feat: add ideogram4 support (#1609 )

perf: ratchet streaming budget so plan stops re-merging every step (#1611 )
fix: Suppress spurious error output for --help (#1607 ) (#1608 )
2026-10-14 05:10:44 +08:00 · 2026-06-06 16:34:16 +08:00 · 2026-06-06 16:32:03 +08:00 · 2026-06-06 16:23:44 +08:00 · 2026-06-06 16:22:18 +08:00 · 2026-06-03 23:32:31 +08:00
126 changed files with 18252 additions and 4341519 deletions
@@ -0,0 +1,15 @@
+## Summary
+
+<!-- Describe what changed and why. Keep the PR focused on one clear change. -->
+
+## Related Issue / Discussion
+
+<!-- Link related issues, discussions, or previous PRs if applicable. -->
+
+## Additional Information
+
+<!-- Add verification notes, screenshots, sample output, or other context when applicable. -->
+
+## Checklist
+
+- [ ] I have read and confirmed this PR follows the [contribution guidelines](https://github.com/leejet/stable-diffusion.cpp/blob/master/CONTRIBUTING.md).
@@ -14,6 +14,8 @@ on:
    paths:
      [
        ".github/workflows/**",
+        ".dockerignore",
+        "Dockerfile*",
        "**/CMakeLists.txt",
        "**/Makefile",
        "**/*.h",
@@ -29,6 +31,8 @@ on:
    paths:
      [
        ".github/workflows/**",
+        ".dockerignore",
+        "Dockerfile*",
        "**/CMakeLists.txt",
        "**/Makefile",
        "**/*.h",
@@ -135,7 +139,7 @@ jobs:
        id: depends
        run: |
          sudo apt-get update
-          sudo apt-get install build-essential libvulkan-dev glslc
+          sudo apt-get install build-essential libvulkan-dev glslc spirv-headers

      - name: Build
        id: cmake_build
@@ -177,7 +181,7 @@ jobs:
  build-and-push-docker-images:
    name: Build and push container images
    if: ${{ github.event_name != 'pull_request' }}
-    runs-on: ubuntu-latest
+    runs-on: ${{ matrix.runner }}

    permissions:
      contents: read
@@ -189,6 +193,20 @@ jobs:
    strategy:
      matrix:
        variant: [musa, sycl, vulkan, cuda]
+        platform: [linux/amd64]
+        runner: [ubuntu-latest]
+        build-args: [""]
+        tag-suffix: [""]
+        include:
+          - variant: cuda
+            platform: linux/arm64
+            runner: ubuntu-24.04-arm
+            tag-suffix: "-spark"
+            build-args: |
+              CUDA_VERSION=13.0.0
+              UBUNTU_VERSION=24.04
+              CUDA_ARCHITECTURES=121
+              GGML_CUDA_FA_ALL_QUANTS=ON

    env:
      REGISTRY: ghcr.io
@@ -243,12 +261,13 @@ jobs:
        uses: docker/build-push-action@v6
        with:
          context: .
-          platforms: linux/amd64
+          platforms: ${{ matrix.platform }}
          push: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
          file: Dockerfile.${{ matrix.variant }}
-          tags: ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}:${{ env.BRANCH_NAME }}-${{ matrix.variant }}
+          tags: ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}:${{ env.BRANCH_NAME }}-${{ matrix.variant }}${{ matrix.tag-suffix }}
          labels: ${{ steps.meta.outputs.labels }}
          annotations: ${{ steps.meta.outputs.annotations }}
+          build-args: ${{ matrix.build-args }}

  macOS-latest-cmake:
    runs-on: macos-latest
@@ -444,12 +463,129 @@ jobs:
          path: |
            sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-x64.zip

+  windows-latest-rocm:
+    runs-on: windows-2022
+
+    env:
+      ROCM_VERSION: "7.13.0"
+      GPU_TARGETS: "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1150;gfx1151;gfx1152;gfx1200;gfx1201"
+
+    steps:
+      - uses: actions/checkout@v3
+        with:
+          submodules: recursive
+
+      - name: Setup Node
+        uses: actions/setup-node@v4
+        with:
+          node-version: 20
+
+      - name: Setup pnpm
+        uses: pnpm/action-setup@v4
+        with:
+          version: 10.15.1
+
+      - name: Cache ROCm Installation
+        id: cache-rocm
+        uses: actions/cache@v4
+        with:
+          path: C:\TheRock\build
+          key: rocm-${{ env.ROCM_VERSION }}-gfx1151-${{ runner.os }}
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.16
+        with:
+          key: windows-latest-rocm-${{ env.ROCM_VERSION }}-x64
+          evict-old-files: 1d
+
+      - name: Install ROCm
+        if: steps.cache-rocm.outputs.cache-hit != 'true'
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "Downloading AMD ROCm ${{ env.ROCM_VERSION }} tarball"
+          Invoke-WebRequest -Uri "https://repo.amd.com/rocm/tarball/therock-dist-windows-gfx1151-${{ env.ROCM_VERSION }}.tar.gz" -OutFile "${env:RUNNER_TEMP}\rocm.tar.gz"
+          write-host "Extracting ROCm tarball"
+          mkdir C:\TheRock\build -Force
+          tar -xzf "${env:RUNNER_TEMP}\rocm.tar.gz" -C C:\TheRock\build --strip-components=1
+          write-host "Completed ROCm extraction"
+
+      - name: Setup ROCm Environment
+        run: |
+          $rocmPath = "C:\TheRock\build"
+          echo "HIP_PATH=$rocmPath" >> $env:GITHUB_ENV
+          echo "HIP_DEVICE_LIB_PATH=$rocmPath\lib\llvm\amdgcn\bitcode" >> $env:GITHUB_ENV
+          echo "HIP_PLATFORM=amd" >> $env:GITHUB_ENV
+          echo "LLVM_PATH=$rocmPath\lib\llvm" >> $env:GITHUB_ENV
+          echo "$rocmPath\bin" >> $env:GITHUB_PATH
+          echo "$rocmPath\lib\llvm\bin" >> $env:GITHUB_PATH
+
+      - name: Build
+        run: |
+          mkdir build
+          cd build
+          cmake .. `
+            -G "Unix Makefiles" `
+            -DCMAKE_PREFIX_PATH="${env:HIP_PATH}" `
+            -DSD_HIPBLAS=ON `
+            -DSD_BUILD_SHARED_LIBS=ON `
+            -DGGML_NATIVE=OFF `
+            -DCMAKE_C_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang.exe" `
+            -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang++.exe" `
+            -DCMAKE_HIP_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang.exe" `
+            -DHIP_PATH="${env:HIP_PATH}" `
+            -DCMAKE_BUILD_TYPE=Release `
+            -DGPU_TARGETS="${{ env.GPU_TARGETS }}"
+          cmake --build . --config Release --parallel ${env:NUMBER_OF_PROCESSORS}
+
+      - name: Get commit hash
+        id: commit
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: pr-mpt/actions-commit-hash@v2
+
+      - name: Pack artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          $ErrorActionPreference = "Stop"
+          $dst = "build\bin"
+          $rocmBin = Join-Path "${env:HIP_PATH}" "bin"
+          $requiredRocmPaths = @(
+            (Join-Path $rocmBin "rocblas.dll"),
+            (Join-Path $rocmBin "rocblas\library")
+          )
+          foreach ($path in $requiredRocmPaths) {
+            if (!(Test-Path $path)) {
+              throw "Missing ROCm runtime dependency: $path"
+            }
+          }
+
+          foreach ($pattern in @("rocblas*.dll", "hipblas*.dll", "libhipblas*.dll")) {
+            Copy-Item -Path (Join-Path $rocmBin $pattern) -Destination $dst -Force -ErrorAction SilentlyContinue
+          }
+
+          foreach ($dir in @("rocblas", "hipblaslt")) {
+            $src = Join-Path $rocmBin $dir
+            if (Test-Path $src) {
+              Copy-Item -Path $src -Destination $dst -Recurse -Force
+            }
+          }
+
+          7z a sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-rocm-${{ env.ROCM_VERSION }}-x64.zip .\build\bin\*
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v4
+        with:
+          name: sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-rocm-${{ env.ROCM_VERSION }}-x64.zip
+          path: |
+            sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-rocm-${{ env.ROCM_VERSION }}-x64.zip
+
  windows-latest-cmake-hip:
    runs-on: windows-2022

    env:
-      HIPSDK_INSTALLER_VERSION: "25.Q3"
-      GPU_TARGETS: "gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
+      HIPSDK_INSTALLER_VERSION: "26.Q1"
+      ROCM_VERSION: "7.1.1"
+      GPU_TARGETS: "gfx1150;gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"

    steps:
      - uses: actions/checkout@v3
@@ -484,7 +620,7 @@ jobs:
        run: |
          $ErrorActionPreference = "Stop"
          write-host "Downloading AMD HIP SDK Installer"
-          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ env.HIPSDK_INSTALLER_VERSION }}-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ env.HIPSDK_INSTALLER_VERSION }}-Win11-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
          write-host "Installing AMD HIP SDK"
          $proc = Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -PassThru
          $completed = $proc.WaitForExit(600000)
@@ -537,47 +673,75 @@ jobs:
        run: |
          md "build\bin\rocblas\library\"
          md "build\bin\hipblaslt\library"
-          cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\"
-          cp "${env:HIP_PATH}\bin\hipblaslt.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\libhipblas.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\libhipblaslt.dll" "build\bin\"
          cp "${env:HIP_PATH}\bin\rocblas.dll" "build\bin\"
          cp "${env:HIP_PATH}\bin\rocblas\library\*" "build\bin\rocblas\library\"
          cp "${env:HIP_PATH}\bin\hipblaslt\library\*" "build\bin\hipblaslt\library\"
-          7z a sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-rocm-x64.zip .\build\bin\*
+          7z a sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-rocm-${{ env.ROCM_VERSION }}-x64.zip .\build\bin\*

      - name: Upload artifacts
        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
        uses: actions/upload-artifact@v4
        with:
-          name: sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-rocm-x64.zip
+          name: sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-rocm-${{ env.ROCM_VERSION }}-x64.zip
          path: |
-            sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-rocm-x64.zip
+            sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-rocm-${{ env.ROCM_VERSION }}-x64.zip

  ubuntu-latest-rocm:
-    runs-on: ubuntu-latest
-    container: rocm/dev-ubuntu-24.04:7.2
+    runs-on: ubuntu-24.04

    env:
-      ROCM_VERSION: "7.2"
      UBUNTU_VERSION: "24.04"
-      GPU_TARGETS: "gfx1151;gfx1150;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201"
+
+    strategy:
+      matrix:
+        include:
+          - ROCM_VERSION: "7.2.1"
+            gpu_targets: "gfx908;gfx90a;gfx942;gfx1030;gfx1031;gfx1032;gfx1100;gfx1101;gfx1102;gfx1151;gfx1150;gfx1200;gfx1201"
+            build: 'x64'
+          - ROCM_VERSION: "7.13.0"
+            gpu_targets: "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1150;gfx1151;gfx1152;gfx1200;gfx1201"
+            build: x64

    steps:
-      - run: apt-get update && apt-get install -y git
      - name: Clone
        id: checkout
        uses: actions/checkout@v6
        with:
          submodules: recursive

-      - name: Setup Node
-        uses: actions/setup-node@v4
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.16
        with:
-          node-version: 20
+          key: ubuntu-rocm-cmake-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
+          evict-old-files: 1d

-      - name: Setup pnpm
-        uses: pnpm/action-setup@v4
-        with:
-          version: 10.15.1
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt install -y build-essential cmake wget zip ninja-build
+
+      - name: Setup Legacy ROCm
+        if: matrix.ROCM_VERSION == '7.2.1'
+        id: legacy_env
+        run: |
+          sudo mkdir --parents --mode=0755 /etc/apt/keyrings
+          wget https://repo.radeon.com/rocm/rocm.gpg.key -O - | \
+            gpg --dearmor | sudo tee /etc/apt/keyrings/rocm.gpg > /dev/null
+
+          sudo tee /etc/apt/sources.list.d/rocm.list << EOF
+          deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/${{ matrix.ROCM_VERSION }} noble main
+          EOF
+
+          sudo tee /etc/apt/preferences.d/rocm-pin-600 << EOF
+          Package: *
+          Pin: release o=repo.radeon.com
+          Pin-Priority: 600
+          EOF
+
+          sudo apt update
+          sudo apt-get install -y libssl-dev rocm-hip-sdk

      - name: Free disk space
        run: |
@@ -592,51 +756,29 @@ jobs:
          sudo rm -rf /var/lib/apt/lists/* || true
          sudo apt clean

-      - name: Dependencies
-        id: depends
+      - name: Setup TheRock
+        if: matrix.ROCM_VERSION != '7.2.1'
+        id: therock_env
        run: |
-          sudo apt-get update
-          sudo apt install -y \
-            cmake \
-            hip-dev \
-            hipblas-dev \
-            ninja-build \
-            rocm-dev \
-            zip
-          # Clean apt caches to recover disk space
-          sudo apt clean
-          sudo rm -rf /var/lib/apt/lists/* || true
+          wget https://repo.amd.com/rocm/tarball/therock-dist-linux-gfx1151-${{ matrix.ROCM_VERSION }}.tar.gz
+          mkdir install
+          tar -xf *.tar.gz -C install
+          export ROCM_PATH=$(pwd)/install
+          echo ROCM_PATH=$ROCM_PATH >> $GITHUB_ENV
+          echo PATH=$PATH:$ROCM_PATH/bin >> $GITHUB_ENV
+          echo LD_LIBRARY_PATH=$ROCM_PATH/lib:$ROCM_PATH/llvm/lib:$ROCM_PATH/lib/rocprofiler-systems >> $GITHUB_ENV

-      - name: Setup ROCm Environment
-        run: |
-          # Add ROCm to PATH for current session
-          echo "/opt/rocm/bin" >> $GITHUB_PATH
+      # setup-node installs into /opt/hostedtoolcache, which is removed above.
+      # Keep Node/pnpm setup after disk cleanup so the server frontend can be embedded.
+      - name: Setup Node
+        uses: actions/setup-node@v4
+        with:
+          node-version: 20

-          # Build regex pattern from ${{ env.GPU_TARGETS }} (match target as substring)
-          TARGET_REGEX="($(printf '%s' "${{ env.GPU_TARGETS }}" | sed 's/;/|/g'))"
-
-          # Remove library files for architectures we're not building for to save disk space
-          echo "Cleaning up unneeded architecture files..."
-          cd /opt/rocm/lib/rocblas/library
-          # Keep only our target architectures
-          for file in *; do
-            if printf '%s' "$file" | grep -q 'gfx'; then
-              if ! printf '%s' "$file" | grep -Eq "$TARGET_REGEX"; then
-                echo "Removing $file" &&
-                sudo rm -f "$file";
-              fi
-            fi
-          done
-
-          cd /opt/rocm/lib/hipblaslt/library
-          for file in *; do
-            if printf '%s' "$file" | grep -q 'gfx'; then
-              if ! printf '%s' "$file" | grep -Eq "$TARGET_REGEX"; then
-                echo "Removing $file" &&
-                sudo rm -f "$file";
-              fi
-            fi
-          done
+      - name: Setup pnpm
+        uses: pnpm/action-setup@v4
+        with:
+          version: 10.15.1

      - name: Build
        id: cmake_build
@@ -644,12 +786,12 @@ jobs:
          mkdir build
          cd build
          cmake .. -G Ninja \
-            -DCMAKE_CXX_COMPILER=amdclang++ \
-            -DCMAKE_C_COMPILER=amdclang \
+            -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" \
+            -DCMAKE_HIP_FLAGS="-mllvm --amdgpu-unroll-threshold-local=600" \
            -DCMAKE_BUILD_TYPE=Release \
            -DSD_HIPBLAS=ON \
-            -DGPU_TARGETS="${{ env.GPU_TARGETS }}" \
-            -DAMDGPU_TARGETS="${{ env.GPU_TARGETS }}" \
+            -DHIP_PLATFORM=amd \
+            -DGPU_TARGETS="${{ matrix.gpu_targets }}" \
            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
            -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
            -DSD_BUILD_SHARED_LIBS=ON
@@ -668,16 +810,6 @@ jobs:
          cp ggml/LICENSE ./build/bin/ggml.txt
          cp LICENSE ./build/bin/stable-diffusion.cpp.txt

-          # Move ROCm runtime libraries (to avoid double space consumption)
-          sudo mv /opt/rocm/lib/librocsparse.so* ./build/bin/
-          sudo mv /opt/rocm/lib/libhsa-runtime64.so* ./build/bin/
-          sudo mv /opt/rocm/lib/libamdhip64.so* ./build/bin/
-          sudo mv /opt/rocm/lib/libhipblas.so* ./build/bin/
-          sudo mv /opt/rocm/lib/libhipblaslt.so* ./build/bin/
-          sudo mv /opt/rocm/lib/librocblas.so* ./build/bin/
-          sudo mv /opt/rocm/lib/rocblas/ ./build/bin/
-          sudo mv /opt/rocm/lib/hipblaslt/ ./build/bin/
-
      - name: Fetch system info
        id: system-info
        run: |
@@ -692,15 +824,15 @@ jobs:
        run: |
          cp ggml/LICENSE ./build/bin/ggml.txt
          cp LICENSE ./build/bin/stable-diffusion.cpp.txt
-          zip -y -r sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-Ubuntu-${{ env.UBUNTU_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}-rocm.zip ./build/bin
+          zip -y -r sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-Ubuntu-${{ env.UBUNTU_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}-rocm-${{ matrix.ROCM_VERSION }}.zip ./build/bin

      - name: Upload artifacts
        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
        uses: actions/upload-artifact@v4
        with:
-          name: sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-Ubuntu-${{ env.UBUNTU_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}-rocm.zip
+          name: sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-Ubuntu-${{ env.UBUNTU_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}-rocm-${{ matrix.ROCM_VERSION }}.zip
          path: |
-            sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-Ubuntu-${{ env.UBUNTU_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}-rocm.zip
+            sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-Ubuntu-${{ env.UBUNTU_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}-rocm-${{ matrix.ROCM_VERSION }}.zip

  release:
    if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
@@ -715,6 +847,7 @@ jobs:
      - macOS-latest-cmake
      - windows-latest-cmake
      - windows-latest-cmake-hip
+      - windows-latest-rocm

    steps:
      - name: Clone
@@ -0,0 +1,55 @@
+name: Close inactive PRs
+
+on:
+  schedule:
+    # Run daily. GitHub cron schedules use UTC.
+    - cron: "30 1 * * *"
+  workflow_dispatch:
+    inputs:
+      debug_only:
+        description: "Dry run: log intended actions without changing PRs"
+        required: false
+        default: false
+        type: boolean
+
+permissions:
+  issues: write
+  pull-requests: write
+
+concurrency:
+  group: ${{ github.workflow }}
+  cancel-in-progress: false
+
+jobs:
+  stale-prs:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Mark and close inactive PRs
+        uses: actions/stale@v10
+        with:
+          days-before-issue-stale: -1
+          days-before-issue-close: -1
+
+          days-before-pr-stale: 365
+          days-before-pr-close: 7
+
+          stale-pr-label: pr:inactive
+          close-pr-label: pr:auto-closed
+          exempt-pr-labels: pr:keep-open
+
+          stale-pr-message: >
+            This PR has been inactive for 365 days. If there is no new activity
+            within 7 days, it will be closed automatically. Comment, push new
+            commits, or remove the pr:inactive label to keep it open. Add
+            pr:keep-open to exempt it from future inactive PR cleanup.
+
+          close-pr-message: >
+            Closing this PR because it has had no activity for 7 days after
+            being marked inactive. If this is still useful or ready to move
+            forward, feel free to reopen it with fresh context or updated
+            details. Sorry for any inconvenience.
+
+          remove-pr-stale-when-updated: true
+          delete-branch: false
+          operations-per-run: 100
+          debug-only: ${{ github.event_name == 'workflow_dispatch' && inputs.debug_only || false }}
@@ -1,6 +1,6 @@
 [submodule "ggml"]
    path = ggml
-	url = https://github.com/ggml-org/ggml.git
+	url = https://github.com/leejet/ggml.git
 [submodule "examples/server/frontend"]
 	path = examples/server/frontend
 	url = https://github.com/leejet/sdcpp-webui.git
@@ -13,13 +13,40 @@ if (MSVC)
    add_compile_definitions(_SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING)
    add_compile_options(
        $<$<COMPILE_LANGUAGE:C>:/MP>
+        $<$<COMPILE_LANGUAGE:C>:/utf-8>
        $<$<COMPILE_LANGUAGE:CXX>:/MP>
+        $<$<COMPILE_LANGUAGE:CXX>:/utf-8>
    )
 endif()

 set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)

+if(APPLE)
+    function(sd_set_macos_rpaths target)
+        get_target_property(target_type ${target} TYPE)
+        if(target_type STREQUAL "EXECUTABLE")
+            set(runtime_paths "@executable_path" "@executable_path/../lib")
+        elseif(target_type STREQUAL "SHARED_LIBRARY" OR target_type STREQUAL "MODULE_LIBRARY")
+            set(runtime_paths "@loader_path" "@loader_path/../lib")
+            set_target_properties(${target} PROPERTIES
+                MACOSX_RPATH ON
+                INSTALL_NAME_DIR "@rpath"
+                BUILD_WITH_INSTALL_NAME_DIR ON
+            )
+        else()
+            return()
+        endif()
+
+        # Release artifacts zip the build output directly, so keep macOS rpaths relocatable.
+        set_target_properties(${target} PROPERTIES
+            BUILD_RPATH "${runtime_paths}"
+            INSTALL_RPATH "${runtime_paths}"
+            BUILD_WITH_INSTALL_RPATH ON
+        )
+    endfunction()
+endif()
+
 if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
    set(SD_STANDALONE ON)
 else()
@@ -69,6 +96,12 @@ option(SD_BUILD_SHARED_GGML_LIB      "sd: build ggml as a separate shared lib" O
 option(SD_USE_SYSTEM_GGML            "sd: use system-installed GGML library" OFF)
 #option(SD_BUILD_SERVER               "sd: build server example"                           ON)

+set(CMAKE_C_STANDARD 11)
+set(CMAKE_C_STANDARD_REQUIRED true)
+
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD_REQUIRED true)
+
 if(SD_CUDA)
    message("-- Use CUDA as backend stable-diffusion")
    set(GGML_CUDA ON)
@@ -92,6 +125,12 @@ endif ()
 if (SD_HIPBLAS)
    message("-- Use HIPBLAS as backend stable-diffusion")
    set(GGML_HIP ON)
+    # ggml-hip's device-stub objects must be position-independent, or the
+    # default-PIE sd-cli link fails with `relocation R_X86_64_32 ... cannot be
+    # used when making a PIE object` on distros that default to PIE
+    # (Ubuntu 24.04, Fedora 40+, Debian 12+). The shared-library branch below
+    # already sets this; the static build (the HIP default) did not.
+    set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 endif ()

 if(SD_MUSA)
@@ -106,19 +145,28 @@ if(SD_WEBP)
          "Or link against system library:\n  cmake (...) -DSD_USE_SYSTEM_WEBP=ON")
    endif()
    if(SD_USE_SYSTEM_WEBP)
-        find_package(WebP REQUIRED)
-        add_library(webp ALIAS WebP::webp)
-        # libwebp CMake target naming is not consistent across versions/distros.
-        # Some export WebP::libwebpmux, others export WebP::webpmux.
-        if(TARGET WebP::libwebpmux)
-            add_library(libwebpmux ALIAS WebP::libwebpmux)
-        elseif(TARGET WebP::webpmux)
-            add_library(libwebpmux ALIAS WebP::webpmux)
+        find_package(WebP)
+        if(WebP_FOUND)
+            add_library(webp ALIAS WebP::webp)
+            # libwebp CMake target naming is not consistent across versions/distros.
+            # Some export WebP::libwebpmux, others export WebP::webpmux.
+            if(TARGET WebP::libwebpmux)
+                add_library(libwebpmux ALIAS WebP::libwebpmux)
+            elseif(TARGET WebP::webpmux)
+                add_library(libwebpmux ALIAS WebP::webpmux)
+            else()
+                message(FATAL_ERROR
+                    "Could not find a compatible webpmux target in system WebP package. "
+                    "Expected WebP::libwebpmux or WebP::webpmux."
+                )
+            endif()
        else()
-            message(FATAL_ERROR
-                "Could not find a compatible webpmux target in system WebP package. "
-                "Expected WebP::libwebpmux or WebP::webpmux."
-            )
+            find_package(PkgConfig REQUIRED)
+            pkg_check_modules(WebP REQUIRED IMPORTED_TARGET GLOBAL libwebp)
+            pkg_check_modules(WebPMux REQUIRED IMPORTED_TARGET GLOBAL libwebpmux)
+            link_libraries(PkgConfig::WebP)
+            link_libraries(PkgConfig::WebPMux)
+            add_library(libwebpmux ALIAS PkgConfig::WebPMux)
        endif()
    endif()
 endif()
@@ -133,18 +181,26 @@ if(SD_WEBM)
          "Or link against system library:\n  cmake (...) -DSD_USE_SYSTEM_WEBM=ON")
    endif()
    if(SD_USE_SYSTEM_WEBM)
-        find_path(WEBM_INCLUDE_DIR
-            NAMES mkvmuxer/mkvmuxer.h mkvparser/mkvparser.h common/webmids.h
-            PATH_SUFFIXES webm
-            REQUIRED)
-        find_library(WEBM_LIBRARY
-            NAMES webm libwebm
-            REQUIRED)
+        find_package(PkgConfig)
+        if(PkgConfig_FOUND)
+            pkg_check_modules(WebM REQUIRED IMPORTED_TARGET GLOBAL libwebm)
+        endif()
+        if(PkgConfig_FOUND AND WebM_FOUND)
+            link_libraries(PkgConfig::WebM)
+        else()
+            find_path(WEBM_INCLUDE_DIR
+                NAMES mkvmuxer/mkvmuxer.h mkvparser/mkvparser.h common/webmids.h
+                PATH_SUFFIXES webm
+                REQUIRED)
+            find_library(WEBM_LIBRARY
+                NAMES webm libwebm
+                REQUIRED)

-        add_library(webm UNKNOWN IMPORTED)
-        set_target_properties(webm PROPERTIES
-            IMPORTED_LOCATION "${WEBM_LIBRARY}"
-            INTERFACE_INCLUDE_DIRECTORIES "${WEBM_INCLUDE_DIR}")
+            add_library(webm UNKNOWN IMPORTED)
+            set_target_properties(webm PROPERTIES
+                IMPORTED_LOCATION "${WEBM_LIBRARY}"
+                INTERFACE_INCLUDE_DIRECTORIES "${WEBM_INCLUDE_DIR}")
+        endif()
    endif()
 endif()

@@ -212,6 +268,10 @@ else()
    add_library(${SD_LIB} STATIC ${SD_LIB_SOURCES})
 endif()

+if(APPLE)
+    sd_set_macos_rpaths(${SD_LIB})
+endif()
+
 if(SD_SYCL)
    message("-- Use SYCL as backend stable-diffusion")
    set(GGML_SYCL ON)
@@ -0,0 +1,65 @@
+# Contributing
+
+This document collects general contribution conventions for this repository.
+
+## Before You Start
+
+Before opening a PR, please search existing PRs to avoid duplicating ongoing work.
+
+For large-scale refactors or changes with broad impact, please open an issue first to discuss the approach before submitting a PR.
+
+If you want to update a third-party dependency, please open an issue first instead of submitting a direct PR. See [Dependency Updates](#dependency-updates) for details.
+
+## Pull Requests
+
+Keep each PR focused on one clear change. Large or overly complex PRs are harder to review and may not be merged.
+
+Follow Conventional Commit-style subjects seen in history: `feat:`, `fix:`, `refactor:`, `ci:`, `docs:`, `chore:`. Keep subjects imperative and scoped.
+
+PRs should include:
+
+- What changed and why (short problem/solution summary).
+- Verification evidence when applicable (commands and key outputs).
+- Linked issue/PR context when applicable.
+- Screenshots or sample outputs for UI/visual behavior changes.
+
+## Code Style
+
+Format code according to the repository style before submitting changes.
+
+Formatting follows `.clang-format` (Chromium base, 4-space indent, no tabs). Run `format-code.sh` before opening a PR. Keep C++ standard at C++17-compatible patterns used in this repo.
+
+Naming conventions:
+
+- Use `PascalCase` for class/struct/type names.
+- In `PascalCase` names, preserve common abbreviations in uppercase, for example `SD`, `API`, `HTTP`, `JSON`, `RGB`, `VAE`, `TAE`, `LoRA`, and `WebP`.
+- Use `snake_case` for functions, methods, variables, and file names unless an existing API requires a different style.
+- Use a trailing underscore for private data member names, for example `hidden_size_` or `tokenizer_`.
+- Use `.h` for C and C++ header files. Do not introduce new `.hpp` headers.
+- Use macro-based header include guards instead of `#pragma once`.
+- Format header include guards as `__SD_{PATH}__`, where `{PATH}` is the header path in uppercase snake case without the file extension. For example, `src/sample.h` should use `__SD_SAMPLE_H__`.
+- Do not introduce anonymous namespaces in new or modified code; prefer `static` file-local functions/variables or an explicit named namespace when scoping is needed.
+- In `class`/`struct` definitions, place data members before member functions unless an existing type already clearly follows a different pattern.
+- Keep `test_*.cpp` / `test_*.py` naming for tests.
+
+Some older code in the project may not fully follow the current conventions. Please do not submit PRs that only rewrite existing code to match style rules.
+
+## AI-Assisted Contributions
+
+AI tools may be used to assist development, but contributors are responsible for the quality and correctness of the submitted code.
+
+If any part of a contribution was generated with AI assistance, the contributor must perform a thorough human review before submitting the PR and understand every changed line.
+
+Do not list AI tools as co-authors. The human contributor is the sole responsible author of the submitted code.
+
+Please do not submit AI-generated code that you do not understand, and do not include meaningless experiments, temporary test code, or unrelated generated output in a PR.
+
+## Dependency Updates
+
+Do not submit PRs that update `ggml`. `ggml` updates are performed only after local validation by the maintainer.
+
+Other third-party dependencies are not updated unless necessary. If you want to update a dependency, please open an issue first instead of submitting a direct PR.
+
+## Security & Configuration
+
+Do not commit model weights, secrets, or local absolute paths. Keep large binaries out of git unless intentionally tracked release assets.
@@ -2,7 +2,18 @@ ARG UBUNTU_VERSION=24.04

 FROM ubuntu:$UBUNTU_VERSION AS build

-RUN apt-get update && apt-get install -y --no-install-recommends build-essential git cmake
+# sd-server embeds the web UI at build time, so the build image needs Node/pnpm.
+RUN apt-get update && apt-get install -y --no-install-recommends build-essential git cmake ca-certificates curl gnupg && \
+    mkdir -p /etc/apt/keyrings && \
+    curl -fsSL https://deb.nodesource.com/gpgkey/nodesource-repo.gpg.key -o /tmp/nodesource-repo.gpg.key && \
+    gpg --dearmor -o /etc/apt/keyrings/nodesource.gpg /tmp/nodesource-repo.gpg.key && \
+    rm /tmp/nodesource-repo.gpg.key && \
+    echo "deb [signed-by=/etc/apt/keyrings/nodesource.gpg] https://deb.nodesource.com/node_20.x nodistro main" > /etc/apt/sources.list.d/nodesource.list && \
+    apt-get update && \
+    apt-get install -y --no-install-recommends nodejs && \
+    npm install -g pnpm@10.15.1 && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*

 WORKDIR /sd.cpp

@@ -20,4 +31,4 @@ RUN apt-get update && \
 COPY --from=build /sd.cpp/build/bin/sd-cli /sd-cli
 COPY --from=build /sd.cpp/build/bin/sd-server /sd-server

-ENTRYPOINT [ "/sd-cli" ]
+ENTRYPOINT [ "/sd-cli" ]
@@ -3,14 +3,31 @@ ARG UBUNTU_VERSION=24.04

 FROM nvidia/cuda:${CUDA_VERSION}-cudnn-devel-ubuntu${UBUNTU_VERSION} AS build

-RUN apt-get update && apt-get install -y --no-install-recommends build-essential git ccache cmake
+# sd-server embeds the web UI at build time, so the build image needs Node/pnpm.
+RUN apt-get update && apt-get install -y --no-install-recommends build-essential git ccache cmake ca-certificates curl gnupg && \
+    mkdir -p /etc/apt/keyrings && \
+    curl -fsSL https://deb.nodesource.com/gpgkey/nodesource-repo.gpg.key -o /tmp/nodesource-repo.gpg.key && \
+    gpg --dearmor -o /etc/apt/keyrings/nodesource.gpg /tmp/nodesource-repo.gpg.key && \
+    rm /tmp/nodesource-repo.gpg.key && \
+    echo "deb [signed-by=/etc/apt/keyrings/nodesource.gpg] https://deb.nodesource.com/node_20.x nodistro main" > /etc/apt/sources.list.d/nodesource.list && \
+    apt-get update && \
+    apt-get install -y --no-install-recommends nodejs && \
+    npm install -g pnpm@10.15.1 && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*

 WORKDIR /sd.cpp

 COPY . .

 ARG CUDACXX=/usr/local/cuda/bin/nvcc
-RUN cmake . -B ./build -DSD_CUDA=ON
+ARG CUDA_ARCHITECTURES=""
+ARG GGML_CUDA_FA_ALL_QUANTS=""
+
+RUN cmake . -B ./build \
+    -DSD_CUDA=ON \
+    ${CUDA_ARCHITECTURES:+-DCMAKE_CUDA_ARCHITECTURES="${CUDA_ARCHITECTURES}"} \
+    ${GGML_CUDA_FA_ALL_QUANTS:+-DGGML_CUDA_FA_ALL_QUANTS=${GGML_CUDA_FA_ALL_QUANTS}}
 RUN cmake --build ./build --config Release -j$(nproc)

 FROM nvidia/cuda:${CUDA_VERSION}-cudnn-runtime-ubuntu${UBUNTU_VERSION} AS runtime
@@ -3,7 +3,18 @@ ARG UBUNTU_VERSION=22.04

 FROM mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}-amd64 as build

-RUN apt-get update && apt-get install -y ccache cmake git
+# sd-server embeds the web UI at build time, so the build image needs Node/pnpm.
+RUN apt-get update && apt-get install -y --no-install-recommends ccache cmake git ca-certificates curl gnupg && \
+    mkdir -p /etc/apt/keyrings && \
+    curl -fsSL https://deb.nodesource.com/gpgkey/nodesource-repo.gpg.key -o /tmp/nodesource-repo.gpg.key && \
+    gpg --dearmor -o /etc/apt/keyrings/nodesource.gpg /tmp/nodesource-repo.gpg.key && \
+    rm /tmp/nodesource-repo.gpg.key && \
+    echo "deb [signed-by=/etc/apt/keyrings/nodesource.gpg] https://deb.nodesource.com/node_20.x nodistro main" > /etc/apt/sources.list.d/nodesource.list && \
+    apt-get update && \
+    apt-get install -y --no-install-recommends nodejs && \
+    npm install -g pnpm@10.15.1 && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*

 WORKDIR /sd.cpp

@@ -21,4 +32,4 @@ FROM mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}-amd64 as runt
 COPY --from=build /sd.cpp/build/bin/sd-cli /sd-cli
 COPY --from=build /sd.cpp/build/bin/sd-server /sd-server

-ENTRYPOINT [ "/sd-cli" ]
+ENTRYPOINT [ "/sd-cli" ]
@@ -1,8 +1,20 @@
-ARG SYCL_VERSION=2025.1.0-0
+# ggml SYCL hardware detection uses BMG G31/WCL architecture enums added in oneAPI 2025.3.
+ARG SYCL_VERSION=2025.3.2-0

 FROM intel/oneapi-basekit:${SYCL_VERSION}-devel-ubuntu24.04 AS build

-RUN apt-get update && apt-get install -y cmake
+# sd-server embeds the web UI at build time, so the build image needs Node/pnpm.
+RUN apt-get update && apt-get install -y --no-install-recommends cmake ca-certificates curl gnupg && \
+    mkdir -p /etc/apt/keyrings && \
+    curl -fsSL https://deb.nodesource.com/gpgkey/nodesource-repo.gpg.key -o /tmp/nodesource-repo.gpg.key && \
+    gpg --dearmor -o /etc/apt/keyrings/nodesource.gpg /tmp/nodesource-repo.gpg.key && \
+    rm /tmp/nodesource-repo.gpg.key && \
+    echo "deb [signed-by=/etc/apt/keyrings/nodesource.gpg] https://deb.nodesource.com/node_20.x nodistro main" > /etc/apt/sources.list.d/nodesource.list && \
+    apt-get update && \
+    apt-get install -y --no-install-recommends nodejs && \
+    npm install -g pnpm@10.15.1 && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*

 WORKDIR /sd.cpp

@@ -2,7 +2,18 @@ ARG UBUNTU_VERSION=24.04

 FROM ubuntu:$UBUNTU_VERSION AS build

-RUN apt-get update && apt-get install -y --no-install-recommends build-essential git cmake libvulkan-dev glslc
+# sd-server embeds the web UI at build time, so the build image needs Node/pnpm.
+RUN apt-get update && apt-get install -y --no-install-recommends build-essential git cmake libvulkan-dev glslc spirv-headers ca-certificates curl gnupg && \
+    mkdir -p /etc/apt/keyrings && \
+    curl -fsSL https://deb.nodesource.com/gpgkey/nodesource-repo.gpg.key -o /tmp/nodesource-repo.gpg.key && \
+    gpg --dearmor -o /etc/apt/keyrings/nodesource.gpg /tmp/nodesource-repo.gpg.key && \
+    rm /tmp/nodesource-repo.gpg.key && \
+    echo "deb [signed-by=/etc/apt/keyrings/nodesource.gpg] https://deb.nodesource.com/node_20.x nodistro main" > /etc/apt/sources.list.d/nodesource.list && \
+    apt-get update && \
+    apt-get install -y --no-install-recommends nodejs && \
+    npm install -g pnpm@10.15.1 && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*

 WORKDIR /sd.cpp

@@ -15,29 +15,18 @@ API and command-line option may change frequently.***

 ## 🔥Important News

+* **2026/06/04** 🚀 stable-diffusion.cpp now supports **Ideogram4**
+* **2026/05/31** 🚀 stable-diffusion.cpp now supports **PiD**
+* **2026/05/27** 🚀 stable-diffusion.cpp now supports **Lens**
+* **2026/05/17** 🚀 stable-diffusion.cpp now supports **LTX-2.3**
 * **2026/04/11** 🚀 stable-diffusion.cpp now uses a brand-new embedded web UI.  
-  👉 Details: [PR #1408](https://github.com/leejet/stable-diffusion.cpp/pull/1408)
-
 * **2026/01/18** 🚀 stable-diffusion.cpp now supports **FLUX.2-klein**  
-  👉 Details: [PR #1193](https://github.com/leejet/stable-diffusion.cpp/pull/1193)
-
 * **2025/12/01** 🚀 stable-diffusion.cpp now supports **Z-Image**  
-  👉 Details: [PR #1020](https://github.com/leejet/stable-diffusion.cpp/pull/1020)
-
 * **2025/11/30** 🚀 stable-diffusion.cpp now supports **FLUX.2-dev**  
-  👉 Details: [PR #1016](https://github.com/leejet/stable-diffusion.cpp/pull/1016)
-
 * **2025/10/13** 🚀 stable-diffusion.cpp now supports **Qwen-Image-Edit / Qwen-Image-Edit 2509**  
-  👉 Details: [PR #877](https://github.com/leejet/stable-diffusion.cpp/pull/877)
-
 * **2025/10/12** 🚀 stable-diffusion.cpp now supports **Qwen-Image**  
-  👉 Details: [PR #851](https://github.com/leejet/stable-diffusion.cpp/pull/851)
-
 * **2025/09/14** 🚀 stable-diffusion.cpp now supports **Wan2.1 Vace**  
-  👉 Details: [PR #819](https://github.com/leejet/stable-diffusion.cpp/pull/819)
-
 * **2025/09/06** 🚀 stable-diffusion.cpp now supports **Wan2.1 / Wan2.2**  
-  👉 Details: [PR #778](https://github.com/leejet/stable-diffusion.cpp/pull/778)

 ## Features

@@ -51,18 +40,25 @@ API and command-line option may change frequently.***
    - [SD3/SD3.5](./docs/sd3.md)
    - [FLUX.1-dev/FLUX.1-schnell](./docs/flux.md)
    - [FLUX.2-dev/FLUX.2-klein](./docs/flux2.md)
+    - [Lens](./docs/lens.md)
    - [Chroma](./docs/chroma.md)
    - [Chroma1-Radiance](./docs/chroma_radiance.md)
    - [Qwen Image](./docs/qwen_image.md)
+    - [PiD](./docs/pid.md)
+    - [LongCat Image](./docs/longcat_image.md)
    - [Z-Image](./docs/z_image.md)
    - [Ovis-Image](./docs/ovis_image.md)
    - [Anima](./docs/anima.md)
    - [ERNIE-Image](./docs/ernie_image.md)
+    - [HiDream-O1-Image](./docs/hidream_o1_image.md)
+    - [Ideogram4](./docs/ideogram4.md)
  - Image Edit Models
    - [FLUX.1-Kontext-dev](./docs/kontext.md)
    - [Qwen Image Edit series](./docs/qwen_image_edit.md)
+    - [LongCat Image Edit](./docs/longcat_image.md)
  - Video Models
    - [Wan2.1/Wan2.2](./docs/wan.md)
+    - [LTX-2.3](./docs/ltx2.md)
  - [PhotoMaker](https://github.com/TencentARC/PhotoMaker) support.
  - Control Net support with SD 1.5
  - LoRA support, same as [stable-diffusion-webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#lora)
@@ -132,9 +128,11 @@ API and command-line option may change frequently.***
 ## Performance

 If you want to improve performance or reduce VRAM/RAM usage, please refer to [performance guide](./docs/performance.md).
+For runtime and parameter backend placement, see the [backend selection guide](./docs/backend.md).

 ## More Guides

+- [Backend selection](./docs/backend.md)
 - [SD1.x/SD2.x/SDXL](./docs/sd.md)
 - [SD3/SD3.5](./docs/sd3.md)
 - [FLUX.1-dev/FLUX.1-schnell](./docs/flux.md)
@@ -144,10 +142,14 @@ If you want to improve performance or reduce VRAM/RAM usage, please refer to [pe
 - [🔥Qwen Image](./docs/qwen_image.md)
 - [🔥Qwen Image Edit series](./docs/qwen_image_edit.md)
 - [🔥Wan2.1/Wan2.2](./docs/wan.md)
+- [🔥LTX-2.3](./docs/ltx2.md)
 - [🔥Z-Image](./docs/z_image.md)
 - [Ovis-Image](./docs/ovis_image.md)
 - [Anima](./docs/anima.md)
 - [ERNIE-Image](./docs/ernie_image.md)
+- [HiDream-O1-Image](./docs/hidream_o1_image.md)
+- [Lens](./docs/lens.md)
+- [LongCat Image / LongCat Image Edit](./docs/longcat_image.md)
 - [LoRA](./docs/lora.md)
 - [LCM/LCM-LoRA](./docs/lcm.md)
 - [Using PhotoMaker to personalize image generation](./docs/photo_maker.md)
@@ -163,6 +165,7 @@ These projects wrap `stable-diffusion.cpp` for easier use in other languages/fra

 * Golang (non-cgo): [seasonjs/stable-diffusion](https://github.com/seasonjs/stable-diffusion)
 * Golang (cgo): [Binozo/GoStableDiffusion](https://github.com/Binozo/GoStableDiffusion)
+* Golang (non-cgo): [l8bloom/gosd](https://github.com/l8bloom/gosd)
 * C#: [DarthAffe/StableDiffusion.NET](https://github.com/DarthAffe/StableDiffusion.NET)
 * Python: [william-murray1204/stable-diffusion-cpp-python](https://github.com/william-murray1204/stable-diffusion-cpp-python)
 * Rust: [newfla/diffusion-rs](https://github.com/newfla/diffusion-rs)
@@ -0,0 +1,122 @@
+# Backend selection
+
+`stable-diffusion.cpp` has two backend assignments:
+
+- `--backend` selects the runtime backend used to execute model graphs.
+- `--params-backend` selects the backend used to allocate model parameters.
+
+If `--params-backend` is not set, parameters use the same backend as their module runtime backend.
+
+## Syntax
+
+A backend assignment can be a single backend name:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend cpu
+```
+
+This applies to every module that does not have a more specific assignment.
+
+Assignments can also target individual modules:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend te=cpu,vae=cuda0,diffusion=vulkan0
+```
+
+The same syntax is used for parameter placement:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend cuda0 --params-backend te=cpu,vae=cpu
+```
+
+Module names are case-insensitive. Hyphens and underscores in module names are ignored, so `clip_vision`, `clip-vision`, and `clipvision` are equivalent.
+
+`all=`, `default=`, and `*=` can be used to set the default backend inside a mixed assignment:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend all=cuda0,te=cpu
+```
+
+## Modules
+
+| Module | Purpose | Accepted names |
+| --- | --- | --- |
+| `diffusion` | UNet, DiT, MMDiT, Flux, Wan, Qwen Image, and other diffusion models | `diffusion`, `model`, `unet`, `dit` |
+| `te` | Text encoders and conditioners | `te`, `clip`, `text`, `textencoder`, `textencoders`, `conditioner`, `cond`, `llm`, `t5`, `t5xxl` |
+| `clip_vision` | CLIP vision encoder | `clip_vision`, `clipvision`, `clip-vision`, `vision` |
+| `vae` | VAE and TAE | `vae`, `firststage`, `autoencoder`, `tae` |
+| `controlnet` | ControlNet | `controlnet`, `control` |
+| `photomaker` | PhotoMaker ID encoder and PhotoMaker LoRA | `photomaker`, `photomakerid`, `pmid`, `photo` |
+| `upscaler` | ESRGAN upscaler | `upscaler`, `esrgan`, `hires` |
+
+`te` is the preferred module name for text encoders. `clip` is kept as an accepted alias because many existing commands and model names use CLIP terminology.
+
+## Backend names
+
+Backend names are resolved against the GGML backend device list. Matching is case-insensitive and accepts exact names or unique prefixes, so common values include names such as:
+
+- `cpu`
+- `cuda0`
+- `vulkan0`
+- `metal`
+
+The special values `auto`, `default`, and an empty backend name select the default backend. The default preference is GPU, then integrated GPU, then CPU.
+
+The special value `gpu` selects the first GPU backend, falling back to the first integrated GPU backend.
+
+## Runtime backend vs. parameter backend
+
+The runtime backend controls where graph execution runs. The parameter backend controls where model weights are allocated.
+
+For example:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend cuda0 --params-backend cpu
+```
+
+This runs all modules on `cuda0`, but stores parameters in CPU RAM. During execution, parameters are moved to the runtime backend as needed.
+
+Per-module assignments can be mixed:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend diffusion=cuda0,te=cpu,vae=cpu --params-backend diffusion=cuda0,te=cpu,vae=cpu
+```
+
+This keeps text encoding and VAE execution on CPU while the diffusion model runs on GPU.
+
+## Backend sharing and lifetime
+
+Backends are managed by `SDBackendManager`.
+
+Within one manager, backend instances are cached by resolved backend device name. If multiple modules request the same backend, they share the same `ggml_backend_t`.
+
+For example:
+
+```shell
+--backend te=cpu,vae=cpu
+```
+
+uses one shared CPU backend for both `te` and `vae` runtime execution.
+
+Runtime and parameter assignments also share the same backend cache. If `--backend diffusion=cuda0` and `--params-backend diffusion=cuda0` resolve to the same device, both use the same backend instance.
+
+`SDBackendManager` owns the backend instances and frees them when the context or upscaler is destroyed. Model runners receive non-owning runtime and parameter backend pointers and do not free them.
+
+## Compatibility flags
+
+The older CPU placement flags are still supported:
+
+- `--clip-on-cpu`
+- `--vae-on-cpu`
+- `--control-net-cpu`
+- `--offload-to-cpu`
+
+`--clip-on-cpu`, `--vae-on-cpu`, and `--control-net-cpu` affect runtime backend assignment only when `--backend` is not set. They map to `te=cpu`, `vae=cpu`, and `controlnet=cpu`.
+
+`--offload-to-cpu` affects parameter backend assignment only when `--params-backend` is not set. It is equivalent to:
+
+```shell
+--params-backend cpu
+```
+
+Explicit `--backend` and `--params-backend` assignments are preferred for new commands.
@@ -102,6 +102,11 @@ cmake --build . --config Release
 ## Build with Vulkan

 Install Vulkan SDK from https://www.lunarg.com/vulkan-sdk/.
+On Ubuntu, install the Vulkan development packages and SPIR-V headers:
+
+```shell
+sudo apt-get install build-essential libvulkan-dev glslc spirv-headers
+```

 ```shell
 mkdir build && cd build
@@ -0,0 +1,20 @@
+# How to Use
+
+## Download weights
+
+- Download HiDream-O1-Image-Dev
+    - safetensors: https://huggingface.co/Comfy-Org/HiDream-O1-Image/tree/main/checkpoints
+- Download HiDream-O1-Image
+    - safetensors: https://huggingface.co/Comfy-Org/HiDream-O1-Image/tree/main/checkpoints
+
+## Examples
+
+### HiDream-O1-Image-Dev
+
+```
+.\bin\Release\sd-cli.exe -m  ..\..\ComfyUI\models\diffusion_models\hidream_o1_image_dev_bf16.safetensors -p "a lovely cat holding a sign says 
+'hidream o1 cpp'" --cfg-scale 1.0  -v -H 1024 -W 1024
+```
+
+<img width="256" alt="HiDream-O1-Image-Dev example" src="../assets/hidream-o1/dev_example.png" />
+
@@ -26,12 +26,12 @@ Fortunately, `AMD` provides complete help documentation, you can use the help do

 Then we must set `ROCM` as environment variables before running cmake.

-Usually if you install according to the official tutorial and do not modify the ROCM path, then there is a high probability that it is here `C:\Program Files\AMD\ROCm\5.5\bin`
+Usually if you install according to the official tutorial and do not modify the ROCM path, then there is a high probability that it is here `C:\Program Files\AMD\ROCm\7.1.1\bin`

 This is what I use to set the clang:
 ```Commandline
-set CC=C:\Program Files\AMD\ROCm\5.5\bin\clang.exe
-set CXX=C:\Program Files\AMD\ROCm\5.5\bin\clang++.exe
+set CC=C:\Program Files\AMD\ROCm\7.1.1\bin\clang.exe
+set CXX=C:\Program Files\AMD\ROCm\7.1.1\bin\clang++.exe
 ```

 ## Ninja
@@ -46,7 +46,7 @@ set ninja=C:\Program Files\ninja\ninja.exe
 ## Building stable-diffusion.cpp

 The thing different from the regular CPU build is `-DSD_HIPBLAS=ON` ,
-`-G "Ninja"`, `-DCMAKE_C_COMPILER=clang`, `-DCMAKE_CXX_COMPILER=clang++`, `-DAMDGPU_TARGETS=gfx1100`
+`-G "Ninja"`, `-DCMAKE_C_COMPILER=clang`, `-DCMAKE_CXX_COMPILER=clang++`, `-DAMDGPU_TARGETS=gfx1150;gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032`

 >**Notice**: check the `clang` and `clang++` information:
 ```Commandline
@@ -59,26 +59,29 @@ If you see like this, we can continue:
 clang version 17.0.0 (git@github.amd.com:Compute-Mirrors/llvm-project e3201662d21c48894f2156d302276eb1cf47c7be)
 Target: x86_64-pc-windows-msvc
 Thread model: posix
-InstalledDir: C:\Program Files\AMD\ROCm\5.5\bin
+InstalledDir: C:\Program Files\AMD\ROCm\7.1.1\bin
 ```

 ```
 clang version 17.0.0 (git@github.amd.com:Compute-Mirrors/llvm-project e3201662d21c48894f2156d302276eb1cf47c7be)
 Target: x86_64-pc-windows-msvc
 Thread model: posix
-InstalledDir: C:\Program Files\AMD\ROCm\5.5\bin
+InstalledDir: C:\Program Files\AMD\ROCm\7.1.1\bin
 ```

->**Notice** that the `gfx1100` is the GPU architecture of my GPU, you can change it to your GPU architecture. Click here to see your architecture [LLVM Target](https://rocm.docs.amd.com/en/latest/release/windows_support.html#windows-supported-gpus)
+>**Notice** that the GPU targets are now compatible with multiple GPU architectures (ROCm 7.1.1 targets). You can change them to match your GPU architecture. Click here to see your architecture [LLVM Target](https://rocm.docs.amd.com/en/latest/release/windows_support.html#windows-supported-gpus)

-My GPU is AMD Radeon™ RX 7900 XTX Graphics, so I set it to `gfx1100`.
+Examples:
+- AMD Radeon™ RX 7900 XTX Graphics: `gfx1100`
+- AMD Radeon™ RX 7900 XT Graphics: `gfx1101`
+- AMD Radeon™ RX 7900 GRE Graphics: `gfx1102`

 option:

 ```commandline
 mkdir build
 cd build
-cmake .. -G "Ninja" -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DSD_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DAMDGPU_TARGETS=gfx1100
+cmake .. -G "Ninja" -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DSD_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DAMDGPU_TARGETS="gfx1150;gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
 cmake --build . --config Release
 ```

@@ -0,0 +1,40 @@
+# How to Use
+
+## Download weights
+
+- Download Ideogram4
+    - safetensors: https://huggingface.co/ideogram-ai/ideogram-4-fp8/tree/main/transformer
+- Download Ideogram4 uncond
+    - safetensors: https://huggingface.co/ideogram-ai/ideogram-4-fp8/tree/main/unconditional_transformer
+- Download vae
+    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-dev/tree/main
+- Download Qwen3-VL-8B-Instruct
+    - gguf: https://huggingface.co/unsloth/Qwen3-VL-8B-Instruct-GGUF/tree/main
+
+## Convert weights
+
+fp8 scale -> bf16
+
+```
+python .\convert_fp8_scale_to_bf16.py --input .\ideogram4_fp8.safetensors --output ideogram4_bf16.safetensors
+python .\convert_fp8_scale_to_bf16.py --input .\ideogram4_uncond_fp8.safetensors --output ideogram4_uncond_bf16.safetensors
+```
+
+bf16 -> q8
+
+```
+.\bin\Release\sd-cli.exe -M convert -m ideogram4_bf16.safetensors -o ideogram4-Q8_0.gguf --tensor-type-rules "^layers.*adaln_modulation.*weight=q8_0,layers.*attention.o.*weight=q8_0,layers.*attention.qkv.*weight=q8_0,layers.*feed_forward.*weight=q8_0" -v
+
+.\bin\Release\sd-cli.exe -M convert -m ideogram4_uncond_bf16.safetensors -o ideogram4_uncond-Q8_0.gguf --tensor-type-rules "^layers.*adaln_modulation.*weight=q8_0,layers.*attention.o.*weight=q8_0,layers.*attention.qkv.*weight=q8_0,layers.*feed_forward.*weight=q8_0" -v
+```
+
+If you want lower VRAM usage, you can change the quantization from q8_0 to a lower-level quantization, such as q4_0.
+
+
+## Examples
+
+```sh
+.\bin\Release\sd-cli.exe --diffusion-model ideogram4-Q8_0.gguf --uncond-diffusion-model ideogram4_uncond-Q8_0.gguf --llm ..\..\llm\Qwen3VL-8B-Instruct-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors -p '{"high_level_description":"A square 1024 x 1024 luxury fashion magazine cover featuring exactly one short chubby fluffy cat as the main model. The cat sits on a soft ivory studio floor, facing the viewer with a stylish calm expression, wearing tiny black sunglasses, a red silk scarf, and a small gold collar charm. In front of the cat on the floor is a wide horizontal luxury nameplate that clearly reads ideogram4.cpp. The whole design feels premium, fashionable, clean, and editorial.","style_description":{"aesthetics":"luxury fashion magazine cover, high-end pet couture campaign, minimalist editorial design, elegant studio photography, soft paper texture, refined typography, fashionable and polished","lighting":"Soft diffused studio lighting, gentle spotlight on the cat, subtle floor shadow, warm ivory highlights, clean separation between subject and background","photo":"high-resolution fashion editorial photography look, front-facing cat portrait, crisp fur details, glossy sunglasses, clear readable nameplate text, shallow depth of field","medium":"mixed media fashion photography and premium editorial graphic design","color_palette":["#F4EFE7","#111111","#D8B56D","#B73A3A","#FFFFFF","#8A7A6A"]},"compositional_deconstruction":{"canvas":"Square 1024 x 1024 canvas with a normal upright orientation. Do not rotate the poster or any text. Use a clean fashion magazine cover layout.","background":"Warm ivory studio backdrop with subtle paper grain, a soft spotlight gradient, faint floor shadow, and a few minimal gold editorial lines. The background is spacious, premium, and uncluttered.","layout":"Top center has a small elegant headline. Center area features one cat as the main fashion model. Lower foreground has a wide horizontal luxury nameplate placed on the floor in front of the cat. Bottom center has a small footer. All text is horizontal, upright, and readable left to right.","elements":[{"type":"text","desc":"Top center headline reading LOOK WHAT I FOUND in a refined high-fashion serif font. The headline is horizontal, centered, elegant, and secondary to the nameplate text."},{"type":"obj","desc":"Exactly one short chubby fluffy cat sitting in the center like a luxury fashion model. The cat has a large round head, compact body, short legs, soft detailed fur, expressive eyes, and a calm confident pose. The cat is cute and rounded, not tall, not stretched, not duplicated."},{"type":"obj","desc":"Tiny glossy black sunglasses worn naturally by the cat, slightly oversized but still showing the cat face clearly. The sunglasses add a chic fashion-editorial attitude."},{"type":"obj","desc":"A red silk scarf tied neatly around the cat neck, with soft folds and a couture feeling. The scarf must not cover the cat face or the nameplate."},{"type":"obj","desc":"A small gold collar charm or fashion accessory under the scarf, subtle and premium, adding a luxury campaign detail."},{"type":"obj","desc":"In the lower foreground, place a wide horizontal luxury nameplate on the floor in front of the cat. The nameplate is low, flat, landscape-oriented, much wider than tall, like a fashion show seat card or premium display plaque. It is centered, front-facing, level, and fully visible. It must not become vertical, tall, standing, rotated, or side-facing."},{"type":"text","desc":"Print the exact text ideogram4.cpp only on the wide horizontal nameplate. Use clean bold black lettering, perfectly spelled, lowercase, with the number 4 and .cpp extension. The text must fit completely inside the nameplate, stay horizontal, and be readable from left to right."},{"type":"obj","desc":"Add sparse premium editorial accents around the edges: thin gold lines, small code brackets, tiny cursor marks, subtle dots, and minimal geometric details. No extra cats, no stickers, no animal faces, no busy decorations."},{"type":"text","desc":"Bottom center footer reading tiny paws, big compile energy in a small refined monospace or editorial font. The footer is horizontal, centered, understated, and much smaller than the nameplate text."}]}}'  --diffusion-fa -v --offload-to-cpu -H 1024 -W 1024
+```
+
+<img alt="ideogram4 image example" src="../assets/ideogram4/example.png" />
@@ -0,0 +1,32 @@
+# How to Use
+
+Lens uses a Lens diffusion transformer, the FLUX.2 VAE, and GPT-OSS-20B as the LLM text encoder.
+
+## Download weights
+
+- Download Lens
+    - safetensors: https://huggingface.co/Comfy-Org/Lens/tree/main/diffusion_models
+- Download Lens Turbo
+    - safetensors: https://huggingface.co/Comfy-Org/Lens/tree/main/diffusion_models
+- Download vae
+    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-dev/tree/main
+- Download GPT-OSS-20B
+    - gguf: https://huggingface.co/unsloth/gpt-oss-20b-GGUF/tree/main
+
+## Examples
+
+### Lens
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model ..\..\ComfyUI\models\diffusion_models\lens_bf16.safetensors --llm "..\..\llm\gpt-oss-20b-UD-Q8_K_XL.gguf" --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors --cfg-scale 5.0  -p "A crystal dragon soaring through an aurora borealis sky, its entire body made of transparent faceted crystal refracting the green and purple aurora light into rainbow spectra, ice particles trailing from its wings, high fantasy digital art" --diffusion-fa -v
+```
+
+<img width="256" alt="Lens example" src="../assets/lens/example.png" />
+
+### Lens Turbo
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model ..\..\ComfyUI\models\diffusion_models\lens_turbo_bf16.safetensors --llm "..\..\llm\gpt-oss-20b-UD-Q8_K_XL.gguf" --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors --cfg-scale 1.0  -p "A crystal dragon soaring through an aurora borealis sky, its entire body made of transparent faceted crystal refracting the green and purple aurora light into rainbow spectra, ice particles trailing from its wings, high fantasy digital art" --diffusion-fa -v --steps 4
+```
+
+<img width="256" alt="Lens Turbo example" src="../assets/lens/turbo_example.png" />
@@ -0,0 +1,30 @@
+# How to Use
+
+LongCat-Image uses a LongCat diffusion transformer, the FLUX VAE, and Qwen2.5-VL as the LLM text encoder.
+
+## Download weights
+
+- Download LongCat Image
+    - safetensors: https://huggingface.co/Comfy-Org/LongCat-Image/tree/main/split_files/diffusion_models
+    - gguf: https://huggingface.co/vantagewithai/LongCat-Image-GGUF/tree/main/comfy
+- Download LongCat Image Edit
+    - LongCat Image Edit Turbo: https://huggingface.co/meituan-longcat/LongCat-Image-Edit-Turbo
+    - gguf: https://huggingface.co/vantagewithai/LongCat-Image-Edit-GGUF/tree/main
+- Download vae
+    - safetensors: https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/ae.safetensors
+- Download qwen_2.5_vl 7b
+    - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main/split_files/text_encoders
+    - gguf: https://huggingface.co/mradermacher/Qwen2.5-VL-7B-Instruct-GGUF/tree/main
+    - For image editing with GGUF text encoders, also download the matching mmproj file and pass it with `--llm_vision`.
+
+## Run
+
+LongCat uses quoted text for character-level text rendering. Put target text inside single quotes, double quotes, or Chinese quotes.
+
+### LongCat Image
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\LongCat-Image-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\ae.sft --llm ..\..\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct-Q8_0.gguf -p "a lovely cat holding a sign says 'longcat.cpp'" --cfg-scale 5.0 --sampling-method euler --flow-shift 3 -v --offload-to-cpu --diffusion-fa
+```
+
+<img alt="longcat example" src="../assets/longcat/example.png" />
@@ -0,0 +1,77 @@
+# How to Use
+
+## Download weights
+
+- Download LTX-2.3
+    - safetensors: https://huggingface.co/Kijai/LTX2.3_comfy/tree/main/diffusion_models
+    - gguf: https://huggingface.co/unsloth/LTX-2.3-GGUF/tree/main
+- Download gemma-3-12b-it
+    - gguf: https://huggingface.co/unsloth/gemma-3-12b-it-GGUF/tree/main
+- Download embeddings connectors
+    - safetensors: https://huggingface.co/unsloth/LTX-2.3-GGUF/tree/main/text_encoders
+- Download vae
+    - safetensors: https://huggingface.co/unsloth/LTX-2.3-GGUF/tree/main/vae
+- Download audio vae
+    - safetensors: https://huggingface.co/unsloth/LTX-2.3-GGUF/tree/main/vae
+- Download LTX spatial latent upscaler
+    - safetensors: https://huggingface.co/Lightricks/LTX-2.3/resolve/main/ltx-2.3-spatial-upscaler-x2-1.1.safetensors
+
+## Examples
+
+### LTX-2.3 dev T2V
+
+```
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\ltx-2.3-22b-dev-UD-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\ltx-2.3-22b-dev_video_vae.safetensors --audio-vae ..\..\ComfyUI\models\vae\ltx-2.3-22b-dev_audio_vae.safetensors --llm ..\..\ComfyUI\models\text_encoders\gemma-3-12b-it-qat-UD-Q4_K_XL.gguf --embeddings-connectors ..\..\ComfyUI\models\text_encoders\ltx-2.3-22b-dev_embeddings_connectors.safetensors  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "worst quality, low quality, blurry, distorted, artifacts" -W 1280 -H 720 --diffusion-fa --offload-to-cpu --video-frames 33 --fps 24 -o t2v.webm
+```
+
+<video
+  src="../assets/ltx2/t2v.webm"
+  controls
+  muted
+  style="max-width: 100%; height: auto;"></video>
+
+### LTX-2.3 dev I2V
+
+```
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\ltx-2.3-22b-dev-UD-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\ltx-2.3-22b-dev_video_vae.safetensors --audio-vae ..\..\ComfyUI\models\vae\ltx-2.3-22b-dev_audio_vae.safetensors --llm ..\..\ComfyUI\models\text_encoders\gemma-3-12b-it-qat-UD-Q4_K_XL.gguf --embeddings-connectors ..\..\ComfyUI\models\text_encoders\ltx-2.3-22b-dev_embeddings_connectors.safetensors  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v  -W 1280 -H 720 --diffusion-fa --offload-to-cpu --video-frames 33 -i ..\assets\ernie_image\turbo_example.png -o i2v.webm
+```
+
+<video
+  src="../assets/ltx2/i2v.webm"
+  controls
+  muted
+  style="max-width: 100%; height: auto;"></video>
+
+### LTX-2.3 dev FLF2V
+
+```
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\ltx-2.3-22b-dev-UD-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\ltx-2.3-22b-dev_video_vae.safetensors --audio-vae ..\..\ComfyUI\models\vae\ltx-2.3-22b-dev_audio_vae.safetensors --llm ..\..\ComfyUI\models\text_encoders\gemma-3-12b-it-qat-UD-Q4_K_XL.gguf --embeddings-connectors ..\..\ComfyUI\models\text_encoders\ltx-2.3-22b-dev_embeddings_connectors.safetensors  -p "glass flower blossom" --cfg-scale 6.0 --sampling-method euler -v  -W 1280 -H 720 --diffusion-fa --offload-to-cpu --video-frames 33 --init-img ..\..\ComfyUI\input\start_image.png --end-img ..\..\ComfyUI\input\end_image.png -o flf2v.webm
+```
+
+<video
+  src="../assets/ltx2/flf2v.webm"
+  controls
+  muted
+  style="max-width: 100%; height: auto;"></video>
+
+### LTX-2.3 spatial latent upscale
+
+LTX spatial latent upscale runs a model-backed x2 latent upsampler between the low-resolution video pass and the high-resolution refine pass. `-W` and `-H` are the pre-upscale generation size; the spatial upsampler produces x2 latent dimensions.
+
+Put `ltx-2.3-spatial-upscaler-x2-1.1.safetensors` under the directory passed to `--hires-upscalers-dir`, then use the model name without path or extension in `--hires-upscaler`.
+
+```
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\ltx-2.3-22b-dev-UD-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\ltx-2.3-22b-dev_video_vae.safetensors --audio-vae ..\..\ComfyUI\models\vae\ltx-2.3-22b-dev_audio_vae.safetensors --llm ..\..\ComfyUI\models\text_encoders\gemma-3-12b-it-qat-UD-Q4_K_XL.gguf --embeddings-connectors ..\..\ComfyUI\models\text_encoders\ltx-2.3-22b-dev_embeddings_connectors.safetensors --hires-upscalers-dir ..\..\ComfyUI\models\latent_upscale_models --hires-upscaler ltx-2.3-spatial-upscaler-x2-1.1 --hires --hires-steps 4 -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v  -W 640 -H 360 --diffusion-fa --offload-to-cpu --video-frames 33 -i ..\assets\ernie_image\turbo_example.png -o hires_i2v.webm
+```
+
+By default, the hires refine pass uses the main sampler and scheduler, then trims the second-pass sigma schedule by `--hires-denoising-strength` (`0.7` by default). To reproduce a ComfyUI-style explicit refine schedule, pass custom hires sigmas:
+
+```
+--hires-sigmas "0.85,0.725,0.421875,0.0"
+```
+
+<video
+  src="../assets/ltx2/hires_i2v.webm"
+  controls
+  muted
+  style="max-width: 100%; height: auto;"></video>
@@ -0,0 +1,39 @@
+# How to Use
+
+PiD is NVIDIA's Pixel Diffusion Decoder. It replaces the usual VAE decode or decode-then-upscale path with a pixel-space diffusion decoder conditioned on a
+source latent and text prompt.
+
+In stable-diffusion.cpp, PiD currently runs as an image edit pipeline: provide a reference image with `-r`/`--ref-image`, encode that image with a matching VAE, then let the PiD diffusion model decode/upscale directly to RGB.
+
+## Download weights
+
+- Download PiD
+    - safetensors: https://huggingface.co/Comfy-Org/PixelDiT/tree/main/diffusion_models
+- Download Gemma 2 2B
+    - safetensors: https://huggingface.co/Comfy-Org/PixelDiT/tree/main/text_encoders
+- Download the VAE that matches the PiD checkpoint backbone
+    - safetensors: https://huggingface.co/nvidia/PiD/tree/main/checkpoints
+    - Flux / Z-Image PiD: use the Flux VAE and pass `--vae-format flux`
+    - SD3 PiD: use the SD3 VAE and pass `--vae-format sd3`
+    - Flux.2 PiD: use the Flux.2 VAE and pass `--vae-format flux2`
+
+The official PiD model card should be checked before use. At the time of the initial PiD release, the official weights are under the NSCLv1 non-commercial license.
+
+## Examples
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model ..\..\ComfyUI\models\diffusion_models\pid_flux1_512_to_2048_4step_bf16.safetensors --llm "..\..\ComfyUI\models\text_encoders\gemma_2_2b_it_elm_bf16.safetensors" --vae ..\..\ComfyUI\models\vae\ae.sft --vae-format flux --cfg-scale 1.0  -p "a lovely cat" -r ..\assets\ernie_image\turbo_example.png --diffusion-fa -v --steps 4 -H 2048 -W 2048 --rng cpu
+```
+
+Before:
+
+<img width="256" alt="ERNIE-Image Turbo example" src="../assets/ernie_image/turbo_example.png" />
+
+After:
+<img width="1024" alt="PiD example" src="../assets/pid/example.png" />
+
+## Notes
+
+- `-r`/`--ref-image` is required. PiD uses the first reference image as the source latent condition.
+- `--vae-format` should match the VAE latent layout used by the PiD checkpoint. This is important when using standalone VAE files because the PiD diffusion
+  checkpoint alone does not identify the VAE format.
@@ -21,7 +21,7 @@ You can run Z-Image with stable-diffusion.cpp on GPUs with 4GB of VRAM — or ev
 ### Z-Image-Turbo

 ```
-.\bin\Release\sd-cli.exe --diffusion-model  z_image_turbo-Q3_K.gguf --vae ..\..\ComfyUI\models\vae\ae.sft  --llm ..\..\ComfyUI\models\text_encoders\Qwen3-4B-Instruct-2507-Q4_K_M.gguf -p "A cinematic, melancholic photograph of a solitary hooded figure walking through a sprawling, rain-slicked metropolis at night. The city lights are a chaotic blur of neon orange and cool blue, reflecting on the wet asphalt. The scene evokes a sense of being a single component in a vast machine. Superimposed over the image in a sleek, modern, slightly glitched font is the philosophical quote: 'THE CITY IS A CIRCUIT BOARD, AND I AM A BROKEN TRANSISTOR.' -- moody, atmospheric, profound, dark academic" --cfg-scale 1.0 -v --offload-to-cpu --diffusion-fa -H 1024 -W 512
+.\bin\Release\sd-cli.exe --diffusion-model  z_image_turbo-Q3_K.gguf --vae ..\..\ComfyUI\models\vae\ae.sft  --llm ..\..\ComfyUI\models\text_encoders\Qwen3-4B-Instruct-2507-Q4_K_M.gguf -p "A cinematic, melancholic photograph of a solitary hooded figure walking through a sprawling, rain-slicked metropolis at night. The city lights are a chaotic blur of neon orange and cool blue, reflecting on the wet asphalt. The scene evokes a sense of being a single component in a vast machine. Superimposed over the image in a sleek, modern, slightly glitched font is the philosophical quote: 'THE CITY IS A CIRCUIT BOARD, AND I AM A BROKEN TRANSISTOR.' -- moody, atmospheric, profound, dark academic" --cfg-scale 1.0 -v --offload-to-cpu --diffusion-fa -H 1024 -W 512 --steps 8
 ```

 <img width="256" alt="z-image example" src="../assets/z_image/q3_K.png" />
@@ -7,6 +7,13 @@ add_executable(${TARGET}
    image_metadata.cpp
    main.cpp
 )
+if(APPLE)
+    sd_set_macos_rpaths(${TARGET})
+endif()
+target_include_directories(${TARGET} PRIVATE
+    "${CMAKE_CURRENT_SOURCE_DIR}/.."
+    "${PROJECT_SOURCE_DIR}/src"
+)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE stable-diffusion zip ${CMAKE_THREAD_LIBS_INIT})
 if(SD_WEBP)
@@ -41,6 +41,8 @@ Context Options:
  --qwen2vl_vision <string>                alias of --llm_vision. Deprecated.
  --diffusion-model <string>               path to the standalone diffusion model
  --high-noise-diffusion-model <string>    path to the standalone high noise diffusion model
+  --uncond-diffusion-model <string>        path to the standalone unconditional diffusion model, currently used by
+                                           Ideogram4 CFG
  --vae <string>                           path to standalone vae model
  --taesd <string>                         path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)
  --tae <string>                           alias of --taesd
@@ -55,7 +57,8 @@ Context Options:
                                           then threads will be set to the number of CPU physical cores
  --chroma-t5-mask-pad <int>               t5 mask pad size of chroma
  --max-vram <float>                       maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables
-                                           graph splitting
+                                           graph splitting; a negative value auto-detects free VRAM, sparing the
+                                           specified value (e.g. -0.5 will keep at least 0.5 GiB free)
  --force-sdxl-vae-conv-scale              force use of conv scale on sdxl vae
  --offload-to-cpu                         place the weights in RAM to save VRAM, and automatically load them into VRAM
                                           when needed
@@ -103,6 +106,12 @@ Generation Options:
  --hires-upscaler <string>                highres fix upscaler, Lanczos, Nearest, Latent, Latent (nearest), Latent
                                           (nearest-exact), Latent (antialiased), Latent (bicubic), Latent (bicubic
                                           antialiased), or a model name under --hires-upscalers-dir (default: Latent)
+  --extra-sample-args <string>             extra sampler/scheduler/guidance args, key=value list. APG supports apg_eta,
+                                           apg_momentum, apg_norm_threshold, apg_norm_threshold_smoothing; SLG supports
+                                           slg_uncond; lcm supports noise_clip_std, noise_scale_start, noise_scale_end;
+                                           ltx2 supports max_shift, base_shift, stretch, terminal; euler_ge supports gamma
+  --extra-tiling-args <string>             extra VAE tiling args, key=value list. LTX video VAE supports
+                                           temporal_tile_frames (default: 4), temporal_tile_overlap (default: 1)
  -H, --height <int>                       image height, in pixel space (default: 512)
  -W, --width <int>                        image width, in pixel space (default: 512)
  --steps <int>                            number of sample steps (default: 20)
@@ -122,8 +131,8 @@ Generation Options:
  --hires-upscale-tile-size <int>          highres fix upscaler tile size, reserved for model-backed upscalers (default:
                                           128)
  --cfg-scale <float>                      unconditional guidance scale: (default: 7.0)
-  --img-cfg-scale <float>                  image guidance scale for inpaint or instruct-pix2pix models: (default: same
-                                           as --cfg-scale)
+  --img-cfg-scale <float>                  image guidance scale for inpaint or image edit models: (default: same as
+                                           --cfg-scale)
  --guidance <float>                       distilled guidance scale for models with guidance input (default: 3.5)
  --slg-scale <float>                      skip layer guidance (SLG) scale, only for DiT models: (default: 0). 0 means
                                           disabled, a value of 2.5 is nice for sd3.5 medium
@@ -133,8 +142,8 @@ Generation Options:
                                           res_2s; 1 for euler_a, er_sde and dpm++2s_a)
  --flow-shift <float>                     shift value for Flow models like SD3.x or WAN (default: auto)
  --high-noise-cfg-scale <float>           (high noise) unconditional guidance scale: (default: 7.0)
-  --high-noise-img-cfg-scale <float>       (high noise) image guidance scale for inpaint or instruct-pix2pix models
-                                           (default: same as --cfg-scale)
+  --high-noise-img-cfg-scale <float>       (high noise) image guidance scale for inpaint or image edit models (default:
+                                           same as --cfg-scale)
  --high-noise-guidance <float>            (high noise) distilled guidance scale for models with guidance input
                                           (default: 3.5)
  --high-noise-slg-scale <float>           (high noise) skip layer guidance (SLG) scale, only for DiT models: (default:
@@ -144,7 +153,7 @@ Generation Options:
  --high-noise-eta <float>                 (high noise) noise multiplier (default: 0 for ddim_trailing, tcd,
                                           res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)
  --strength <float>                       strength for noising/unnoising (default: 0.75)
-  --pm-style-strength <float>              
+  --pm-style-strength <float>
  --control-strength <float>               strength to apply Control Net (default: 0.9). 1.0 corresponds to full
                                           destruction of information in init image
  --moe-boundary <float>                   timestep boundary for Wan2.2 MoE model. (default: 0.875). Only enabled if
@@ -158,19 +167,22 @@ Generation Options:
  --disable-auto-resize-ref-image          disable auto resize of ref images
  --disable-image-metadata                 do not embed generation metadata on image files
  --vae-tiling                             process vae in tiles to reduce memory usage
+  --temporal-tiling                        enable temporal tiling for LTX video VAE decode
  --hires                                  enable highres fix
  -s, --seed                               RNG seed (default: 42, use random seed for < 0)
  --sampling-method                        sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m,
                                           dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s,
-                                           er_sde] (default: euler for Flux/SD3/Wan, euler_a otherwise)
+                                           er_sde, euler_cfg_pp, euler_a_cfg_pp] (default: euler for Flux/SD3/Wan, euler_a otherwise)
  --high-noise-sampling-method             (high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a,
                                           dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep,
-                                           res_2s, er_sde] default: euler for Flux/SD3/Wan, euler_a otherwise
+                                           res_2s, er_sde, euler_cfg_pp, euler_a_cfg_pp] default: euler for Flux/SD3/Wan, euler_a otherwise
  --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits,
-                                           smoothstep, sgm_uniform, simple, kl_optimal, lcm, bong_tangent], default:
-                                           discrete
+                                           smoothstep, sgm_uniform, simple, kl_optimal, lcm, bong_tangent, ltx2], default:
+                                           model-specific
  --sigmas                                 custom sigma values for the sampler, comma-separated (e.g.,
                                           "14.61,7.8,3.5,0.0").
+  --hires-sigmas                           custom sigma values for the highres fix second pass, comma-separated (e.g.,
+                                           "0.85,0.725,0.421875,0.0").
  --skip-layers                            layers to skip for SLG steps (default: [7,8,9])
  --high-noise-skip-layers                 (high noise) layers to skip for SLG steps (default: [7,8,9])
  -r, --ref-image                          reference image for Flux Kontext models (can be used multiple times)
@@ -169,8 +169,9 @@ struct SDCliParams {
            return 1;
        };

-        auto on_help_arg = [&](int argc, const char** argv, int index) {
+        auto on_help_arg = [&](int argc, const char** argv, int index, bool& valid) {
            normal_exit = true;
+            valid       = true;
            return -1;
        };

@@ -385,11 +386,32 @@ std::string format_frame_idx(std::string pattern, int frame_idx) {
    return result;
 }

+static fs::path get_video_audio_sidecar_path(const SDCliParams& cli_params) {
+    fs::path out_path     = cli_params.output_path;
+    fs::path base_path    = out_path;
+    fs::path ext          = out_path.has_extension() ? out_path.extension() : fs::path{};
+    std::string ext_lower = ext.string();
+    std::transform(ext_lower.begin(), ext_lower.end(), ext_lower.begin(), ::tolower);
+    const EncodedImageFormat output_format = encoded_image_format_from_path(out_path.string());
+    if (!ext.empty()) {
+        if (output_format == EncodedImageFormat::JPEG ||
+            output_format == EncodedImageFormat::PNG ||
+            output_format == EncodedImageFormat::WEBP ||
+            ext_lower == ".avi" ||
+            ext_lower == ".webm") {
+            base_path.replace_extension();
+        }
+    }
+    base_path += ".wav";
+    return base_path;
+}
+
 bool save_results(const SDCliParams& cli_params,
                  const SDContextParams& ctx_params,
                  const SDGenerationParams& gen_params,
                  sd_image_t* results,
-                  int num_results) {
+                  int num_results,
+                  const sd_audio_t* generated_audio = nullptr) {
    if (results == nullptr || num_results <= 0) {
        return false;
    }
@@ -442,6 +464,21 @@ bool save_results(const SDCliParams& cli_params,
        return ok;
    };

+    auto write_audio_sidecar = [&](const fs::path& wav_path) {
+        if (generated_audio == nullptr) {
+            return;
+        }
+        if (write_wav_to_file(wav_path.string(),
+                              generated_audio->data,
+                              generated_audio->sample_count,
+                              generated_audio->channels,
+                              generated_audio->sample_rate)) {
+            LOG_INFO("save result audio to '%s'", wav_path.string().c_str());
+        } else {
+            LOG_WARN("failed to save result audio to '%s'", wav_path.string().c_str());
+        }
+    };
+
    int sucessful_reults = 0;

    if (std::regex_search(cli_params.output_path, format_specifier_regex)) {
@@ -465,8 +502,16 @@ bool save_results(const SDCliParams& cli_params,
            ext = ".avi";
        fs::path video_path = base_path;
        video_path += ext;
-        if (create_video_from_sd_images(video_path.string().c_str(), results, num_results, gen_params.fps) == 0) {
+        std::string final_ext_lower = ext.string();
+        std::transform(final_ext_lower.begin(), final_ext_lower.end(), final_ext_lower.begin(), ::tolower);
+        const bool mux_audio = generated_audio != nullptr && (final_ext_lower == ".avi" || final_ext_lower == ".webm");
+        if (create_video_from_sd_images(video_path.string().c_str(), results, num_results, gen_params.fps, 90, mux_audio ? generated_audio : nullptr) == 0) {
            LOG_INFO("save result video to '%s'", video_path.string().c_str());
+            if (generated_audio != nullptr && !mux_audio) {
+                fs::path wav_path = video_path;
+                wav_path.replace_extension(".wav");
+                write_audio_sidecar(wav_path);
+            }
            return true;
        } else {
            LOG_ERROR("Failed to save result video to '%s'", video_path.string().c_str());
@@ -488,6 +533,9 @@ bool save_results(const SDCliParams& cli_params,
        }
    }
    LOG_INFO("%d/%d images saved", sucessful_reults, num_results);
+    if (generated_audio != nullptr) {
+        write_audio_sidecar(get_video_audio_sidecar_path(cli_params));
+    }
    return sucessful_reults != 0;
 }

@@ -701,7 +749,8 @@ int main(int argc, const char* argv[]) {
    sd_ctx_params_t sd_ctx_params = ctx_params.to_sd_ctx_params_t(vae_decode_only, true, cli_params.taesd_preview);

    SDImageVec results;
-    int num_results = 0;
+    int num_results             = 0;
+    sd_audio_t* generated_audio = nullptr;

    if (cli_params.mode == UPSCALE) {
        num_results = 1;
@@ -733,7 +782,10 @@ int main(int argc, const char* argv[]) {
            results.adopt(generate_image(sd_ctx.get(), &img_gen_params), num_results);
        } else if (cli_params.mode == VID_GEN) {
            sd_vid_gen_params_t vid_gen_params = gen_params.to_sd_vid_gen_params_t();
-            sd_image_t* generated_video        = generate_video(sd_ctx.get(), &vid_gen_params, &num_results);
+            sd_image_t* generated_video        = nullptr;
+            if (!generate_video(sd_ctx.get(), &vid_gen_params, &generated_video, &num_results, &generated_audio)) {
+                generated_video = nullptr;
+            }
            results.adopt(generated_video, num_results);
        }

@@ -749,7 +801,9 @@ int main(int argc, const char* argv[]) {
                                                     ctx_params.offload_params_to_cpu,
                                                     ctx_params.diffusion_conv_direct,
                                                     ctx_params.n_threads,
-                                                     gen_params.upscale_tile_size));
+                                                     gen_params.upscale_tile_size,
+                                                     ctx_params.backend.c_str(),
+                                                     ctx_params.params_backend.c_str()));

        if (upscaler_ctx == nullptr) {
            LOG_ERROR("new_upscaler_ctx failed");
@@ -773,9 +827,12 @@ int main(int argc, const char* argv[]) {
        }
    }

-    if (!save_results(cli_params, ctx_params, gen_params, results.data(), num_results)) {
+    if (!save_results(cli_params, ctx_params, gen_params, results.data(), num_results, generated_audio)) {
+        free_sd_audio(generated_audio);
        return 1;
    }

+    free_sd_audio(generated_audio);
+
    return 0;
 }
@@ -35,6 +35,22 @@ const char* const modes_str[] = {
    "metadata",
 };

+static sd_vae_format_t str_to_vae_format(const std::string& value) {
+    if (value == "auto") {
+        return SD_VAE_FORMAT_AUTO;
+    }
+    if (value == "flux") {
+        return SD_VAE_FORMAT_FLUX;
+    }
+    if (value == "sd3") {
+        return SD_VAE_FORMAT_SD3;
+    }
+    if (value == "flux2") {
+        return SD_VAE_FORMAT_FLUX2;
+    }
+    return SD_VAE_FORMAT_COUNT;
+}
+
 #if defined(_WIN32)
 static std::string utf16_to_utf8(const std::wstring& wstr) {
    if (wstr.empty())
@@ -229,6 +245,7 @@ bool parse_options(int argc, const char** argv, const std::vector<ArgOptions>& o
        return false;
    };

+    bool valid = false;
    for (int i = 1; i < argc; i++) {
        arg            = argv[i];
        bool found_arg = false;
@@ -271,7 +288,7 @@ bool parse_options(int argc, const char** argv, const std::vector<ArgOptions>& o
                break;

            if (match_and_apply(options.manual_options, [&](auto& option) {
-                    int ret = option.cb(argc, argv, i);
+                    int ret = option.cb(argc, argv, i, valid);
                    if (ret < 0) {
                        invalid_arg = true;
                        return;
@@ -283,7 +300,9 @@ bool parse_options(int argc, const char** argv, const std::vector<ArgOptions>& o
        }

        if (invalid_arg) {
-            LOG_ERROR("error: invalid parameter for argument: %s", arg.c_str());
+            if (!valid) {
+                LOG_ERROR("error: invalid parameter for argument: %s", arg.c_str());
+            }
            return false;
        }
        if (!found_arg) {
@@ -340,10 +359,26 @@ ArgOptions SDContextParams::get_options() {
         "--high-noise-diffusion-model",
         "path to the standalone high noise diffusion model",
         &high_noise_diffusion_model_path},
+        {"",
+         "--uncond-diffusion-model",
+         "path to the standalone unconditional diffusion model, currently used by Ideogram4 CFG",
+         &uncond_diffusion_model_path},
+        {"",
+         "--embeddings-connectors",
+         "path to LTXAV embeddings connectors",
+         &embeddings_connectors_path},
        {"",
         "--vae",
         "path to standalone vae model",
         &vae_path},
+        {"",
+         "--vae-format",
+         "VAE latent format override: auto, flux, sd3, or flux2 (default: auto)",
+         &vae_format},
+        {"",
+         "--audio-vae",
+         "path to standalone LTX audio vae model",
+         &audio_vae_path},
        {"",
         "--taesd",
         "path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)",
@@ -380,6 +415,14 @@ ArgOptions SDContextParams::get_options() {
         "--upscale-model",
         "path to esrgan model.",
         &esrgan_path},
+        {"",
+         "--backend",
+         "runtime backend assignment, e.g. cpu or clip=cpu,vae=cuda0,diffusion=vulkan0",
+         &backend},
+        {"",
+         "--params-backend",
+         "parameter backend assignment, e.g. cpu or diffusion=cpu,clip=cpu",
+         &params_backend},
    };

    options.int_options = {
@@ -397,11 +440,15 @@ ArgOptions SDContextParams::get_options() {
    options.float_options = {
        {"",
         "--max-vram",
-         "maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables graph splitting",
+         "maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables graph splitting; a negative value auto-detects free VRAM, sparing the specified value (e.g. -0.5 will keep at least 0.5 GiB free)",
         &max_vram},
    };

    options.bool_options = {
+        {"",
+         "--stream-layers",
+         "enable residency+prefetch streaming on top of --max-vram (no effect without --max-vram; defaults to false)",
+         true, &stream_layers},
        {"",
         "--force-sdxl-vae-conv-scale",
         "force use of conv scale on sdxl vae",
@@ -623,6 +670,11 @@ bool SDContextParams::validate(SDMode mode) {
        }
    }

+    if (str_to_vae_format(vae_format) == SD_VAE_FORMAT_COUNT) {
+        LOG_ERROR("error: vae_format must be 'auto', 'flux', 'sd3', or 'flux2'");
+        return false;
+    }
+
    return true;
 }

@@ -661,7 +713,11 @@ std::string SDContextParams::to_string() const {
        << "  llm_vision_path: \"" << llm_vision_path << "\",\n"
        << "  diffusion_model_path: \"" << diffusion_model_path << "\",\n"
        << "  high_noise_diffusion_model_path: \"" << high_noise_diffusion_model_path << "\",\n"
+        << "  uncond_diffusion_model_path: \"" << uncond_diffusion_model_path << "\",\n"
+        << "  embeddings_connectors_path: \"" << embeddings_connectors_path << "\",\n"
        << "  vae_path: \"" << vae_path << "\",\n"
+        << "  vae_format: \"" << vae_format << "\",\n"
+        << "  audio_vae_path: \"" << audio_vae_path << "\",\n"
        << "  taesd_path: \"" << taesd_path << "\",\n"
        << "  esrgan_path: \"" << esrgan_path << "\",\n"
        << "  control_net_path: \"" << control_net_path << "\",\n"
@@ -676,6 +732,9 @@ std::string SDContextParams::to_string() const {
        << "  sampler_rng_type: " << sd_rng_type_name(sampler_rng_type) << ",\n"
        << "  offload_params_to_cpu: " << (offload_params_to_cpu ? "true" : "false") << ",\n"
        << "  max_vram: " << max_vram << ",\n"
+        << "  stream_layers: " << (stream_layers ? "true" : "false") << ",\n"
+        << "  backend: \"" << backend << "\",\n"
+        << "  params_backend: \"" << params_backend << "\",\n"
        << "  enable_mmap: " << (enable_mmap ? "true" : "false") << ",\n"
        << "  control_net_cpu: " << (control_net_cpu ? "true" : "false") << ",\n"
        << "  clip_on_cpu: " << (clip_on_cpu ? "true" : "false") << ",\n"
@@ -718,7 +777,10 @@ sd_ctx_params_t SDContextParams::to_sd_ctx_params_t(bool vae_decode_only, bool f
        llm_vision_path.c_str(),
        diffusion_model_path.c_str(),
        high_noise_diffusion_model_path.c_str(),
+        uncond_diffusion_model_path.c_str(),
+        embeddings_connectors_path.c_str(),
        vae_path.c_str(),
+        audio_vae_path.c_str(),
        taesd_path.c_str(),
        control_net_path.c_str(),
        embedding_vec.data(),
@@ -750,7 +812,11 @@ sd_ctx_params_t SDContextParams::to_sd_ctx_params_t(bool vae_decode_only, bool f
        chroma_use_t5_mask,
        chroma_t5_mask_pad,
        qwen_image_zero_cond_t,
+        str_to_vae_format(vae_format),
        max_vram,
+        stream_layers,
+        backend.c_str(),
+        params_backend.c_str(),
    };
    return sd_ctx_params;
 }
@@ -807,6 +873,14 @@ ArgOptions SDGenerationParams::get_options() {
         "Latent (antialiased), Latent (bicubic), Latent (bicubic antialiased), or a model name "
         "under --hires-upscalers-dir (default: Latent)",
         &hires_upscaler},
+        {"",
+         "--extra-sample-args",
+         "extra sampler/scheduler/guidance args, key=value list. APG supports apg_eta, apg_momentum, apg_norm_threshold, apg_norm_threshold_smoothing; SLG supports slg_uncond; lcm supports noise_clip_std, noise_scale_start, noise_scale_end; ltx2 supports max_shift, base_shift, stretch, terminal; euler_ge supports gamma",
+         &extra_sample_args},
+        {"",
+         "--extra-tiling-args",
+         "extra VAE tiling args, key=value list. LTX video VAE supports temporal_tile_frames (default: 4), temporal_tile_overlap (default: 1)",
+         &extra_tiling_args},
    };

    options.int_options = {
@@ -881,7 +955,7 @@ ArgOptions SDGenerationParams::get_options() {
         &sample_params.guidance.txt_cfg},
        {"",
         "--img-cfg-scale",
-         "image guidance scale for inpaint or instruct-pix2pix models: (default: same as --cfg-scale)",
+         "image guidance scale for inpaint or image edit models: (default: same as --cfg-scale)",
         &sample_params.guidance.img_cfg},
        {"",
         "--guidance",
@@ -913,7 +987,7 @@ ArgOptions SDGenerationParams::get_options() {
         &high_noise_sample_params.guidance.txt_cfg},
        {"",
         "--high-noise-img-cfg-scale",
-         "(high noise) image guidance scale for inpaint or instruct-pix2pix models (default: same as --cfg-scale)",
+         "(high noise) image guidance scale for inpaint or image edit models (default: same as --cfg-scale)",
         &high_noise_sample_params.guidance.img_cfg},
        {"",
         "--high-noise-guidance",
@@ -990,6 +1064,11 @@ ArgOptions SDGenerationParams::get_options() {
         "process vae in tiles to reduce memory usage",
         true,
         &vae_tiling_params.enabled},
+        {"",
+         "--temporal-tiling",
+         "enable temporal tiling for LTX video VAE decode",
+         true,
+         &vae_tiling_params.temporal_tiling},
        {"",
         "--hires",
         "enable highres fix",
@@ -1101,11 +1180,11 @@ ArgOptions SDGenerationParams::get_options() {
        return 1;
    };

-    auto on_sigmas_arg = [&](int argc, const char** argv, int index) {
-        if (++index >= argc) {
+    auto parse_sigmas_arg = [&](const char* value, std::vector<float>* target, const char* option_name) {
+        if (target == nullptr || value == nullptr) {
            return -1;
        }
-        std::string sigmas_str = argv[index];
+        std::string sigmas_str = value;
        if (!sigmas_str.empty() && sigmas_str.front() == '[') {
            sigmas_str.erase(0, 1);
        }
@@ -1113,6 +1192,7 @@ ArgOptions SDGenerationParams::get_options() {
            sigmas_str.pop_back();
        }

+        size_t before = target->size();
        std::stringstream ss(sigmas_str);
        std::string item;
        while (std::getline(ss, item, ',')) {
@@ -1120,24 +1200,38 @@ ArgOptions SDGenerationParams::get_options() {
            item.erase(item.find_last_not_of(" \t\n\r\f\v") + 1);
            if (!item.empty()) {
                try {
-                    custom_sigmas.push_back(std::stof(item));
+                    target->push_back(std::stof(item));
                } catch (const std::invalid_argument&) {
-                    LOG_ERROR("error: invalid float value '%s' in --sigmas", item.c_str());
+                    LOG_ERROR("error: invalid float value '%s' in %s", item.c_str(), option_name);
                    return -1;
                } catch (const std::out_of_range&) {
-                    LOG_ERROR("error: float value '%s' out of range in --sigmas", item.c_str());
+                    LOG_ERROR("error: float value '%s' out of range in %s", item.c_str(), option_name);
                    return -1;
                }
            }
        }

-        if (custom_sigmas.empty() && !sigmas_str.empty()) {
-            LOG_ERROR("error: could not parse any sigma values from '%s'", argv[index]);
+        if (target->size() == before && !sigmas_str.empty()) {
+            LOG_ERROR("error: could not parse any sigma values from '%s'", value);
            return -1;
        }
        return 1;
    };

+    auto on_sigmas_arg = [&](int argc, const char** argv, int index) {
+        if (++index >= argc) {
+            return -1;
+        }
+        return parse_sigmas_arg(argv[index], &custom_sigmas, "--sigmas");
+    };
+
+    auto on_hires_sigmas_arg = [&](int argc, const char** argv, int index) {
+        if (++index >= argc) {
+            return -1;
+        }
+        return parse_sigmas_arg(argv[index], &hires_custom_sigmas, "--hires-sigmas");
+    };
+
    auto on_ref_image_arg = [&](int argc, const char** argv, int index) {
        if (++index >= argc) {
            return -1;
@@ -1244,22 +1338,26 @@ ArgOptions SDGenerationParams::get_options() {
         on_seed_arg},
        {"",
         "--sampling-method",
-         "sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s, er_sde] "
+         "sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s, er_sde, euler_cfg_pp, euler_a_cfg_pp]"
         "(default: euler for Flux/SD3/Wan, euler_a otherwise)",
         on_sample_method_arg},
        {"",
         "--high-noise-sampling-method",
-         "(high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s, er_sde]"
+         "(high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s, er_sde, euler_cfg_pp, euler_a_cfg_pp]"
         " default: euler for Flux/SD3/Wan, euler_a otherwise",
         on_high_noise_sample_method_arg},
        {"",
         "--scheduler",
-         "denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple, kl_optimal, lcm, bong_tangent], default: discrete",
+         "denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple, kl_optimal, lcm, bong_tangent, ltx2], default: model-specific",
         on_scheduler_arg},
        {"",
         "--sigmas",
         "custom sigma values for the sampler, comma-separated (e.g., \"14.61,7.8,3.5,0.0\").",
         on_sigmas_arg},
+        {"",
+         "--hires-sigmas",
+         "custom sigma values for the highres fix second pass, comma-separated (e.g., \"0.85,0.725,0.421875,0.0\").",
+         on_hires_sigmas_arg},
        {"",
         "--skip-layers",
         "layers to skip for SLG steps (default: [7,8,9])",
@@ -1492,11 +1590,31 @@ static bool resolve_model_file_from_dir(const std::string& model_name,
        LOG_ERROR("%s directory is empty", label);
        return false;
    }
+    auto ends_with_valid_ext = [&]() {
+        for (const auto& ext : valid_ext) {
+            if (model_name.size() < ext.size()) {
+                continue;
+            }
+            auto suffix = model_name.substr(model_name.size() - ext.size());
+            std::transform(suffix.begin(), suffix.end(), suffix.begin(), [](unsigned char c) {
+                return static_cast<char>(std::tolower(c));
+            });
+            std::string lower_ext = ext;
+            std::transform(lower_ext.begin(), lower_ext.end(), lower_ext.begin(), [](unsigned char c) {
+                return static_cast<char>(std::tolower(c));
+            });
+            if (suffix == lower_ext) {
+                return true;
+            }
+        }
+        return false;
+    };
+
    if (model_name.empty() ||
        model_name.find('/') != std::string::npos ||
        model_name.find('\\') != std::string::npos ||
        fs::path(model_name).has_root_path() ||
-        fs::path(model_name).has_extension()) {
+        ends_with_valid_ext()) {
        LOG_ERROR("%s must be a model name without path or extension: %s", label, model_name.c_str());
        return false;
    }
@@ -1600,6 +1718,9 @@ bool SDGenerationParams::from_json_str(
        if (hires_json.contains("denoising_strength") && hires_json["denoising_strength"].is_number()) {
            hires_denoising_strength = hires_json["denoising_strength"];
        }
+        if (hires_json.contains("custom_sigmas") && hires_json["custom_sigmas"].is_array()) {
+            hires_custom_sigmas = hires_json["custom_sigmas"].get<std::vector<float>>();
+        }
        if (hires_json.contains("upscale_tile_size") && hires_json["upscale_tile_size"].is_number_integer()) {
            hires_upscale_tile_size = hires_json["upscale_tile_size"];
        }
@@ -1607,6 +1728,7 @@ bool SDGenerationParams::from_json_str(

    auto parse_sample_params_json = [&](const json& sample_json,
                                        sd_sample_params_t& target_params,
+                                        std::string& target_extra_sample_args,
                                        std::vector<int>& target_skip_layers,
                                        std::vector<float>* target_custom_sigmas) {
        if (sample_json.contains("sample_steps") && sample_json["sample_steps"].is_number_integer()) {
@@ -1621,6 +1743,9 @@ bool SDGenerationParams::from_json_str(
        if (sample_json.contains("flow_shift") && sample_json["flow_shift"].is_number()) {
            target_params.flow_shift = sample_json["flow_shift"];
        }
+        if (sample_json.contains("extra_sample_args") && sample_json["extra_sample_args"].is_string()) {
+            target_extra_sample_args = sample_json["extra_sample_args"].get<std::string>();
+        }
        if (target_custom_sigmas != nullptr &&
            sample_json.contains("custom_sigmas") &&
            sample_json["custom_sigmas"].is_array()) {
@@ -1668,11 +1793,12 @@ bool SDGenerationParams::from_json_str(
    };

    if (j.contains("sample_params") && j["sample_params"].is_object()) {
-        parse_sample_params_json(j["sample_params"], sample_params, skip_layers, &custom_sigmas);
+        parse_sample_params_json(j["sample_params"], sample_params, extra_sample_args, skip_layers, &custom_sigmas);
    }
    if (j.contains("high_noise_sample_params") && j["high_noise_sample_params"].is_object()) {
        parse_sample_params_json(j["high_noise_sample_params"],
                                 high_noise_sample_params,
+                                 high_noise_extra_sample_args,
                                 high_noise_skip_layers,
                                 nullptr);
    }
@@ -1682,6 +1808,9 @@ bool SDGenerationParams::from_json_str(
        if (tiling_json.contains("enabled") && tiling_json["enabled"].is_boolean()) {
            vae_tiling_params.enabled = tiling_json["enabled"];
        }
+        if (tiling_json.contains("temporal_tiling") && tiling_json["temporal_tiling"].is_boolean()) {
+            vae_tiling_params.temporal_tiling = tiling_json["temporal_tiling"];
+        }
        if (tiling_json.contains("tile_size_x") && tiling_json["tile_size_x"].is_number_integer()) {
            vae_tiling_params.tile_size_x = tiling_json["tile_size_x"];
        }
@@ -1697,6 +1826,9 @@ bool SDGenerationParams::from_json_str(
        if (tiling_json.contains("rel_size_y") && tiling_json["rel_size_y"].is_number()) {
            vae_tiling_params.rel_size_y = tiling_json["rel_size_y"];
        }
+        if (tiling_json.contains("extra_tiling_args") && tiling_json["extra_tiling_args"].is_string()) {
+            extra_tiling_args = tiling_json["extra_tiling_args"].get<std::string>();
+        }
    }

    if (!parse_lora_json_field(j, lora_path_resolver, lora_map, high_noise_lora_map)) {
@@ -1919,6 +2051,8 @@ bool SDGenerationParams::initialize_cache_params() {
 }

 bool SDGenerationParams::resolve(const std::string& lora_model_dir, const std::string& hires_upscalers_dir, bool strict) {
+    vae_tiling_params.extra_tiling_args = extra_tiling_args.empty() ? nullptr : extra_tiling_args.c_str();
+
    if (high_noise_sample_params.sample_steps <= 0) {
        high_noise_sample_params.sample_steps = -1;
    }
@@ -2039,6 +2173,10 @@ bool SDGenerationParams::validate(SDMode mode) {
            LOG_ERROR("error: hires denoising strength must be in (0.0, 1.0]");
            return false;
        }
+        if (!hires_custom_sigmas.empty() && hires_custom_sigmas.size() < 2) {
+            LOG_ERROR("error: hires custom sigmas must contain at least two values");
+            return false;
+        }
        if (hires_upscale_tile_size < 1) {
            LOG_ERROR("error: hires upscale tile size must be positive");
            return false;
@@ -2099,6 +2237,9 @@ sd_img_gen_params_t SDGenerationParams::to_sd_img_gen_params_t() {
    high_noise_sample_params.guidance.slg.layer_count = high_noise_skip_layers.size();
    sample_params.custom_sigmas                       = custom_sigmas.empty() ? nullptr : custom_sigmas.data();
    sample_params.custom_sigmas_count                 = static_cast<int>(custom_sigmas.size());
+    sample_params.extra_sample_args                   = extra_sample_args.empty() ? nullptr : extra_sample_args.c_str();
+    high_noise_sample_params.extra_sample_args        = high_noise_extra_sample_args.empty() ? nullptr : high_noise_extra_sample_args.c_str();
+    vae_tiling_params.extra_tiling_args               = extra_tiling_args.empty() ? nullptr : extra_tiling_args.c_str();
    cache_params.scm_mask                             = scm_mask.empty() ? nullptr : scm_mask.c_str();

    sd_pm_params_t pm_params = {
@@ -2131,15 +2272,17 @@ sd_img_gen_params_t SDGenerationParams::to_sd_img_gen_params_t() {
    params.vae_tiling_params     = vae_tiling_params;
    params.cache                 = cache_params;

-    params.hires.enabled            = hires_enabled;
-    params.hires.upscaler           = resolved_hires_upscaler;
-    params.hires.model_path         = hires_upscaler_model_path.empty() ? nullptr : hires_upscaler_model_path.c_str();
-    params.hires.scale              = hires_scale;
-    params.hires.target_width       = hires_width;
-    params.hires.target_height      = hires_height;
-    params.hires.steps              = hires_steps;
-    params.hires.denoising_strength = hires_denoising_strength;
-    params.hires.upscale_tile_size  = hires_upscale_tile_size;
+    params.hires.enabled             = hires_enabled;
+    params.hires.upscaler            = resolved_hires_upscaler;
+    params.hires.model_path          = hires_upscaler_model_path.empty() ? nullptr : hires_upscaler_model_path.c_str();
+    params.hires.scale               = hires_scale;
+    params.hires.target_width        = hires_width;
+    params.hires.target_height       = hires_height;
+    params.hires.steps               = hires_steps;
+    params.hires.denoising_strength  = hires_denoising_strength;
+    params.hires.upscale_tile_size   = hires_upscale_tile_size;
+    params.hires.custom_sigmas       = hires_custom_sigmas.empty() ? nullptr : hires_custom_sigmas.data();
+    params.hires.custom_sigmas_count = static_cast<int>(hires_custom_sigmas.size());
    return params;
 }

@@ -2168,28 +2311,43 @@ sd_vid_gen_params_t SDGenerationParams::to_sd_vid_gen_params_t() {
    high_noise_sample_params.guidance.slg.layer_count = high_noise_skip_layers.size();
    sample_params.custom_sigmas                       = custom_sigmas.empty() ? nullptr : custom_sigmas.data();
    sample_params.custom_sigmas_count                 = static_cast<int>(custom_sigmas.size());
+    sample_params.extra_sample_args                   = extra_sample_args.empty() ? nullptr : extra_sample_args.c_str();
+    high_noise_sample_params.extra_sample_args        = high_noise_extra_sample_args.empty() ? nullptr : high_noise_extra_sample_args.c_str();
+    vae_tiling_params.extra_tiling_args               = extra_tiling_args.empty() ? nullptr : extra_tiling_args.c_str();
    cache_params.scm_mask                             = scm_mask.empty() ? nullptr : scm_mask.c_str();

-    params.loras                    = lora_vec.empty() ? nullptr : lora_vec.data();
-    params.lora_count               = static_cast<uint32_t>(lora_vec.size());
-    params.prompt                   = prompt.c_str();
-    params.negative_prompt          = negative_prompt.c_str();
-    params.clip_skip                = clip_skip;
-    params.init_image               = init_image.get();
-    params.end_image                = end_image.get();
-    params.control_frames           = control_frame_views.empty() ? nullptr : control_frame_views.data();
-    params.control_frames_size      = static_cast<int>(control_frame_views.size());
-    params.width                    = get_resolved_width();
-    params.height                   = get_resolved_height();
-    params.sample_params            = sample_params;
-    params.high_noise_sample_params = high_noise_sample_params;
-    params.moe_boundary             = moe_boundary;
-    params.strength                 = strength;
-    params.seed                     = seed;
-    params.video_frames             = video_frames;
-    params.vace_strength            = vace_strength;
-    params.vae_tiling_params        = vae_tiling_params;
-    params.cache                    = cache_params;
+    params.loras                     = lora_vec.empty() ? nullptr : lora_vec.data();
+    params.lora_count                = static_cast<uint32_t>(lora_vec.size());
+    params.prompt                    = prompt.c_str();
+    params.negative_prompt           = negative_prompt.c_str();
+    params.clip_skip                 = clip_skip;
+    params.init_image                = init_image.get();
+    params.end_image                 = end_image.get();
+    params.control_frames            = control_frame_views.empty() ? nullptr : control_frame_views.data();
+    params.control_frames_size       = static_cast<int>(control_frame_views.size());
+    params.width                     = get_resolved_width();
+    params.height                    = get_resolved_height();
+    params.sample_params             = sample_params;
+    params.high_noise_sample_params  = high_noise_sample_params;
+    params.moe_boundary              = moe_boundary;
+    params.strength                  = strength;
+    params.seed                      = seed;
+    params.video_frames              = video_frames;
+    params.fps                       = fps;
+    params.vace_strength             = vace_strength;
+    params.vae_tiling_params         = vae_tiling_params;
+    params.cache                     = cache_params;
+    params.hires.enabled             = hires_enabled;
+    params.hires.upscaler            = resolved_hires_upscaler;
+    params.hires.model_path          = hires_upscaler_model_path.empty() ? nullptr : hires_upscaler_model_path.c_str();
+    params.hires.scale               = hires_scale;
+    params.hires.target_width        = hires_width;
+    params.hires.target_height       = hires_height;
+    params.hires.steps               = hires_steps;
+    params.hires.denoising_strength  = hires_denoising_strength;
+    params.hires.upscale_tile_size   = hires_upscale_tile_size;
+    params.hires.custom_sigmas       = hires_custom_sigmas.empty() ? nullptr : hires_custom_sigmas.data();
+    params.hires.custom_sigmas_count = static_cast<int>(hires_custom_sigmas.size());
    return params;
 }

@@ -2272,14 +2430,17 @@ std::string SDGenerationParams::to_string() const {
        << ", target_height: " << hires_height
        << ", steps: " << hires_steps
        << ", denoising_strength: " << hires_denoising_strength
+        << ", custom_sigmas: " << vec_to_string(hires_custom_sigmas)
        << ", upscale_tile_size: " << hires_upscale_tile_size << " },\n"
        << "  vae_tiling_params: { "
        << vae_tiling_params.enabled << ", "
+        << vae_tiling_params.temporal_tiling << ", "
        << vae_tiling_params.tile_size_x << ", "
        << vae_tiling_params.tile_size_y << ", "
        << vae_tiling_params.target_overlap << ", "
        << vae_tiling_params.rel_size_x << ", "
-        << vae_tiling_params.rel_size_y << " },\n"
+        << vae_tiling_params.rel_size_y << ", "
+        << "\"" << extra_tiling_args << "\" },\n"
        << "}";
    return oss.str();
 }
@@ -2306,6 +2467,7 @@ static json build_sampling_metadata_json(const sd_sample_params_t& sample_params
        {"eta", sample_params.eta},
        {"shifted_timestep", sample_params.shifted_timestep},
        {"flow_shift", sample_params.flow_shift},
+        {"extra_sample_args", safe_json_string(sample_params.extra_sample_args)},
        {"guidance",
         {
             {"txt_cfg", sample_params.guidance.txt_cfg},
@@ -2366,6 +2528,7 @@ std::string build_sdcpp_image_metadata_json(const SDContextParams& ctx_params,
    set_json_basename_if_not_empty(models, "llm_vision", ctx_params.llm_vision_path);
    set_json_basename_if_not_empty(models, "diffusion_model", ctx_params.diffusion_model_path);
    set_json_basename_if_not_empty(models, "high_noise_diffusion_model", ctx_params.high_noise_diffusion_model_path);
+    set_json_basename_if_not_empty(models, "uncond_diffusion_model", ctx_params.uncond_diffusion_model_path);
    set_json_basename_if_not_empty(models, "vae", ctx_params.vae_path);
    set_json_basename_if_not_empty(models, "taesd", ctx_params.taesd_path);
    set_json_basename_if_not_empty(models, "control_net", ctx_params.control_net_path);
@@ -2421,6 +2584,7 @@ std::string build_sdcpp_image_metadata_json(const SDContextParams& ctx_params,
            {"target_height", gen_params.hires_height},
            {"steps", gen_params.hires_steps},
            {"denoising_strength", gen_params.hires_denoising_strength},
+            {"custom_sigmas", gen_params.hires_custom_sigmas},
            {"upscale_tile_size", gen_params.hires_upscale_tile_size},
        };
    }
@@ -2456,14 +2620,18 @@ std::string build_sdcpp_image_metadata_json(const SDContextParams& ctx_params,
        };
    }

-    if (gen_params.vae_tiling_params.enabled) {
+    if (gen_params.vae_tiling_params.enabled ||
+        gen_params.vae_tiling_params.temporal_tiling ||
+        !gen_params.extra_tiling_args.empty()) {
        root["vae_tiling"] = {
            {"enabled", gen_params.vae_tiling_params.enabled},
+            {"temporal_tiling", gen_params.vae_tiling_params.temporal_tiling},
            {"tile_size_x", gen_params.vae_tiling_params.tile_size_x},
            {"tile_size_y", gen_params.vae_tiling_params.tile_size_y},
            {"target_overlap", gen_params.vae_tiling_params.target_overlap},
            {"rel_size_x", gen_params.vae_tiling_params.rel_size_x},
            {"rel_size_y", gen_params.vae_tiling_params.rel_size_y},
+            {"extra_tiling_args", gen_params.extra_tiling_args},
        };
    }

@@ -2497,6 +2665,9 @@ std::string get_image_params(const SDContextParams& ctx_params,
    }
    parameter_string += "Guidance: " + std::to_string(gen_params.sample_params.guidance.distilled_guidance) + ", ";
    parameter_string += "Eta: " + std::to_string(gen_params.sample_params.eta) + ", ";
+    if (!gen_params.extra_sample_args.empty()) {
+        parameter_string += "Extra sample args: " + gen_params.extra_sample_args + ", ";
+    }
    parameter_string += "Seed: " + std::to_string(seed) + ", ";
    parameter_string += "Size: " + std::to_string(gen_params.get_resolved_width()) + "x" + std::to_string(gen_params.get_resolved_height()) + ", ";
    parameter_string += "Model: " + sd_basename(ctx_params.model_path) + ", ";
@@ -2525,6 +2696,9 @@ std::string get_image_params(const SDContextParams& ctx_params,
    if (!ctx_params.diffusion_model_path.empty()) {
        parameter_string += "Unet: " + sd_basename(ctx_params.diffusion_model_path) + ", ";
    }
+    if (!ctx_params.uncond_diffusion_model_path.empty()) {
+        parameter_string += "Uncond Unet: " + sd_basename(ctx_params.uncond_diffusion_model_path) + ", ";
+    }
    if (!ctx_params.vae_path.empty()) {
        parameter_string += "VAE: " + sd_basename(ctx_params.vae_path) + ", ";
    }
@@ -2537,6 +2711,9 @@ std::string get_image_params(const SDContextParams& ctx_params,
        parameter_string += "Hires resize: " + std::to_string(gen_params.hires_width) + "x" + std::to_string(gen_params.hires_height) + ", ";
        parameter_string += "Hires steps: " + std::to_string(gen_params.hires_steps) + ", ";
        parameter_string += "Denoising strength: " + std::to_string(gen_params.hires_denoising_strength) + ", ";
+        if (!gen_params.hires_custom_sigmas.empty()) {
+            parameter_string += "Hires custom sigmas: " + vec_to_string(gen_params.hires_custom_sigmas) + ", ";
+        }
    }
    parameter_string += "Version: stable-diffusion.cpp";
    parameter_string += ", SDCPP: " + build_sdcpp_image_metadata_json(ctx_params, gen_params, seed, mode);
@@ -56,11 +56,42 @@ struct BoolOption {
    bool* target;
 };

+struct ManualFunction {
+    std::function<int(int, const char**, int, bool&)> _func;
+
+    ManualFunction() = default;
+
+    ManualFunction(std::function<int(int argc, const char** argv, int index, bool& valid)> func)
+        : _func(std::move(func)) {
+    }
+
+    template <typename F>
+    ManualFunction(F func)
+        : _func(make_function(func)) {
+    }
+
+    int operator()(int argc, const char** argv, int index, bool& valid) const {
+        return _func(argc, argv, index, valid);
+    }
+
+private:
+    template <typename F>
+    static std::function<int(int, const char**, int, bool&)> make_function(F func) {
+        if constexpr (std::is_invocable_v<F, int, const char**, int, bool&>) {
+            return func;
+        } else {
+            return [func](int argc, const char** argv, int index, bool&) {
+                return func(argc, argv, index);
+            };
+        }
+    }
+};
+
 struct ManualOption {
    std::string short_name;
    std::string long_name;
    std::string desc;
-    std::function<int(int argc, const char** argv, int index)> cb;
+    ManualFunction cb;
 };

 struct ArgOptions {
@@ -92,7 +123,11 @@ struct SDContextParams {
    std::string llm_vision_path;
    std::string diffusion_model_path;
    std::string high_noise_diffusion_model_path;
+    std::string uncond_diffusion_model_path;
+    std::string embeddings_connectors_path;
    std::string vae_path;
+    std::string vae_format = "auto";
+    std::string audio_vae_path;
    std::string taesd_path;
    std::string esrgan_path;
    std::string control_net_path;
@@ -110,14 +145,17 @@ struct SDContextParams {
    rng_type_t sampler_rng_type = RNG_TYPE_COUNT;
    bool offload_params_to_cpu  = false;
    float max_vram              = 0.f;
-    bool enable_mmap            = false;
-    bool control_net_cpu        = false;
-    bool clip_on_cpu            = false;
-    bool vae_on_cpu             = false;
-    bool flash_attn             = false;
-    bool diffusion_flash_attn   = false;
-    bool diffusion_conv_direct  = false;
-    bool vae_conv_direct        = false;
+    bool stream_layers          = false;
+    std::string backend;
+    std::string params_backend;
+    bool enable_mmap           = false;
+    bool control_net_cpu       = false;
+    bool clip_on_cpu           = false;
+    bool vae_on_cpu            = false;
+    bool flash_attn            = false;
+    bool diffusion_flash_attn  = false;
+    bool diffusion_conv_direct = false;
+    bool vae_conv_direct       = false;

    bool circular   = false;
    bool circular_x = false;
@@ -168,6 +206,8 @@ struct SDGenerationParams {

    sd_sample_params_t sample_params;
    sd_sample_params_t high_noise_sample_params;
+    std::string extra_sample_args;
+    std::string high_noise_extra_sample_args;
    std::vector<int> skip_layers            = {7, 8, 9};
    std::vector<int> high_noise_skip_layers = {7, 8, 9};

@@ -183,7 +223,8 @@ struct SDGenerationParams {
    int video_frames                     = 1;
    int fps                              = 16;
    float vace_strength                  = 1.f;
-    sd_tiling_params_t vae_tiling_params = {false, 0, 0, 0.5f, 0.0f, 0.0f};
+    sd_tiling_params_t vae_tiling_params = {false, false, 0, 0, 0.5f, 0.0f, 0.0f, nullptr};
+    std::string extra_tiling_args;

    std::string pm_id_images_dir;
    std::string pm_id_embed_path;
@@ -201,6 +242,7 @@ struct SDGenerationParams {
    int hires_steps                = 0;
    float hires_denoising_strength = 0.7f;
    int hires_upscale_tile_size    = 128;
+    std::vector<float> hires_custom_sigmas;

    std::map<std::string, float> lora_map;
    std::map<std::string, float> high_noise_lora_map;
@@ -613,6 +613,13 @@ typedef struct {
    uint32_t size;
 } avi_index_entry;

+typedef struct {
+    char fourcc[4];
+    uint32_t flags;
+    uint32_t offset;
+    uint32_t size;
+} avi_chunk_index_entry;
+
 void write_u32_le(FILE* f, uint32_t val) {
    fwrite(&val, 4, 1, f);
 }
@@ -647,6 +654,33 @@ void write_fourcc(std::vector<uint8_t>& data, const char* fourcc) {
    data.insert(data.end(), fourcc, fourcc + 4);
 }

+static std::vector<uint8_t> audio_to_pcm16_bytes(const sd_audio_t* audio) {
+    if (audio == nullptr || audio->data == nullptr || audio->sample_count == 0 || audio->channels == 0 || audio->sample_rate == 0) {
+        return {};
+    }
+
+    const size_t pcm_samples = static_cast<size_t>(audio->sample_count) * static_cast<size_t>(audio->channels);
+    std::vector<uint8_t> bytes(pcm_samples * sizeof(int16_t));
+    auto* pcm = reinterpret_cast<int16_t*>(bytes.data());
+    for (size_t i = 0; i < pcm_samples; ++i) {
+        const float sample = std::clamp(audio->data[i], -1.0f, 1.0f);
+        pcm[i]             = static_cast<int16_t>(std::lrint(sample * 32767.0f));
+    }
+    return bytes;
+}
+
+static std::pair<uint64_t, uint64_t> audio_sample_range_for_video_frame(const sd_audio_t* audio, int frame_idx, int num_frames, int fps) {
+    if (audio == nullptr || fps <= 0 || num_frames <= 0) {
+        return {0, 0};
+    }
+    const uint64_t total = audio->sample_count;
+    const uint64_t start = static_cast<uint64_t>((static_cast<long double>(frame_idx) * total) / num_frames);
+    const uint64_t end   = frame_idx + 1 == num_frames
+                               ? total
+                               : static_cast<uint64_t>((static_cast<long double>(frame_idx + 1) * total) / num_frames);
+    return {start, std::max(start, end)};
+}
+
 EncodedImageFormat encoded_image_format_from_path(const std::string& path) {
    std::string ext = fs::path(path).extension().string();
    std::transform(ext.begin(), ext.end(), ext.begin(), ::tolower);
@@ -776,7 +810,7 @@ uint8_t* load_image_from_memory(const char* image_bytes,
    return load_image_common(true, image_bytes, len, width, height, expected_width, expected_height, expected_channel);
 }

-std::vector<uint8_t> create_mjpg_avi_from_sd_images_to_vector(sd_image_t* images, int num_images, int fps, int quality) {
+std::vector<uint8_t> create_mjpg_avi_from_sd_images_to_vector(sd_image_t* images, int num_images, int fps, int quality, const sd_audio_t* audio) {
    if (num_images == 0) {
        fprintf(stderr, "Error: Image array is empty.\n");
        return {};
@@ -793,7 +827,13 @@ std::vector<uint8_t> create_mjpg_avi_from_sd_images_to_vector(sd_image_t* images
    // stb_image_write changes JPEG sampling behavior above quality 90.
    // MJPG AVI playback is more compatible when we keep the encoder on the
    // <= 90 path.
-    const int mjpg_quality = std::clamp(quality, 1, 90);
+    const int mjpg_quality               = std::clamp(quality, 1, 90);
+    const bool has_audio                 = audio != nullptr && audio->data != nullptr && audio->sample_count > 0 && audio->channels > 0 && audio->sample_rate > 0;
+    const std::vector<uint8_t> audio_pcm = audio_to_pcm16_bytes(audio);
+    const uint16_t audio_bits_per_sample = 16;
+    const uint16_t audio_block_align     = has_audio ? static_cast<uint16_t>(audio->channels * (audio_bits_per_sample / 8)) : 0;
+    const uint32_t audio_byte_rate       = has_audio ? static_cast<uint32_t>(audio->sample_rate * audio_block_align) : 0;
+    const uint32_t audio_data_size       = has_audio ? static_cast<uint32_t>(audio_pcm.size()) : 0;

    std::vector<uint8_t> avi_data;
    avi_data.reserve(static_cast<size_t>(num_images) * 1024);
@@ -804,7 +844,11 @@ std::vector<uint8_t> create_mjpg_avi_from_sd_images_to_vector(sd_image_t* images
    write_fourcc(avi_data, "AVI ");

    write_fourcc(avi_data, "LIST");
-    write_u32_le(avi_data, 4 + 8 + 56 + 8 + 4 + 8 + 56 + 8 + 40);
+    uint32_t hdrl_size = 4 + 8 + 56 + 8 + 4 + 8 + 56 + 8 + 40;
+    if (has_audio) {
+        hdrl_size += 8 + (4 + 8 + 56 + 8 + 16);
+    }
+    write_u32_le(avi_data, hdrl_size);
    write_fourcc(avi_data, "hdrl");

    write_fourcc(avi_data, "avih");
@@ -815,7 +859,7 @@ std::vector<uint8_t> create_mjpg_avi_from_sd_images_to_vector(sd_image_t* images
    write_u32_le(avi_data, 0x110);
    write_u32_le(avi_data, num_images);
    write_u32_le(avi_data, 0);
-    write_u32_le(avi_data, 1);
+    write_u32_le(avi_data, has_audio ? 2 : 1);
    write_u32_le(avi_data, width * height * 3);
    write_u32_le(avi_data, width);
    write_u32_le(avi_data, height);
@@ -862,12 +906,48 @@ std::vector<uint8_t> create_mjpg_avi_from_sd_images_to_vector(sd_image_t* images
    write_u32_le(avi_data, 0);
    write_u32_le(avi_data, 0);

+    if (has_audio) {
+        write_fourcc(avi_data, "LIST");
+        write_u32_le(avi_data, 4 + 8 + 56 + 8 + 16);
+        write_fourcc(avi_data, "strl");
+
+        write_fourcc(avi_data, "strh");
+        write_u32_le(avi_data, 56);
+        write_fourcc(avi_data, "auds");
+        write_u32_le(avi_data, 0);
+        write_u32_le(avi_data, 0);
+        write_u16_le(avi_data, 0);
+        write_u16_le(avi_data, 0);
+        write_u32_le(avi_data, 0);
+        write_u32_le(avi_data, audio_block_align);
+        write_u32_le(avi_data, audio_byte_rate);
+        write_u32_le(avi_data, 0);
+        write_u32_le(avi_data, static_cast<uint32_t>(audio->sample_count));
+        write_u32_le(avi_data, audio_data_size);
+        write_u32_le(avi_data, static_cast<uint32_t>(-1));
+        write_u32_le(avi_data, audio_block_align);
+        write_u16_le(avi_data, 0);
+        write_u16_le(avi_data, 0);
+        write_u16_le(avi_data, 0);
+        write_u16_le(avi_data, 0);
+
+        write_fourcc(avi_data, "strf");
+        write_u32_le(avi_data, 16);
+        write_u16_le(avi_data, 1);
+        write_u16_le(avi_data, static_cast<uint16_t>(audio->channels));
+        write_u32_le(avi_data, audio->sample_rate);
+        write_u32_le(avi_data, audio_byte_rate);
+        write_u16_le(avi_data, audio_block_align);
+        write_u16_le(avi_data, audio_bits_per_sample);
+    }
+
    write_fourcc(avi_data, "LIST");
    const size_t movi_size_pos = avi_data.size();
    write_u32_le(avi_data, 0);
    write_fourcc(avi_data, "movi");

-    std::vector<avi_index_entry> index(static_cast<size_t>(num_images));
+    std::vector<avi_chunk_index_entry> index;
+    index.reserve(static_cast<size_t>(num_images) + (has_audio ? 1 : 0));
    std::vector<uint8_t> jpeg_data;

    for (int i = 0; i < num_images; i++) {
@@ -884,27 +964,46 @@ std::vector<uint8_t> create_mjpg_avi_from_sd_images_to_vector(sd_image_t* images
            return {};
        }

-        index[i].offset = static_cast<uint32_t>(avi_data.size());
+        avi_chunk_index_entry video_entry = {};
+        memcpy(video_entry.fourcc, "00dc", 4);
+        video_entry.flags  = 0x10;
+        video_entry.offset = static_cast<uint32_t>(avi_data.size());
        write_fourcc(avi_data, "00dc");
        write_u32_le(avi_data, static_cast<uint32_t>(jpeg_data.size()));
-        index[i].size = (uint32_t)jpeg_data.size();
+        video_entry.size = static_cast<uint32_t>(jpeg_data.size());
        avi_data.insert(avi_data.end(), jpeg_data.begin(), jpeg_data.end());
+        index.push_back(video_entry);

        if (jpeg_data.size() % 2) {
            avi_data.push_back(0);
        }
    }

+    if (has_audio && !audio_pcm.empty()) {
+        avi_chunk_index_entry audio_entry = {};
+        memcpy(audio_entry.fourcc, "01wb", 4);
+        audio_entry.flags  = 0;
+        audio_entry.offset = static_cast<uint32_t>(avi_data.size());
+        audio_entry.size   = static_cast<uint32_t>(audio_pcm.size());
+        write_fourcc(avi_data, "01wb");
+        write_u32_le(avi_data, static_cast<uint32_t>(audio_pcm.size()));
+        avi_data.insert(avi_data.end(), audio_pcm.begin(), audio_pcm.end());
+        index.push_back(audio_entry);
+        if (audio_pcm.size() % 2 != 0) {
+            avi_data.push_back(0);
+        }
+    }
+
    const size_t movi_size = avi_data.size() - movi_size_pos - 4;
    patch_u32_le(avi_data, movi_size_pos, static_cast<uint32_t>(movi_size));

    write_fourcc(avi_data, "idx1");
-    write_u32_le(avi_data, num_images * 16);
-    for (int i = 0; i < num_images; i++) {
-        write_fourcc(avi_data, "00dc");
-        write_u32_le(avi_data, 0x10);
-        write_u32_le(avi_data, index[i].offset);
-        write_u32_le(avi_data, index[i].size);
+    write_u32_le(avi_data, static_cast<uint32_t>(index.size() * 16));
+    for (const auto& entry : index) {
+        write_fourcc(avi_data, entry.fourcc);
+        write_u32_le(avi_data, entry.flags);
+        write_u32_le(avi_data, entry.offset);
+        write_u32_le(avi_data, entry.size);
    }

    const size_t file_size = avi_data.size() - riff_size_pos - 4;
@@ -913,8 +1012,8 @@ std::vector<uint8_t> create_mjpg_avi_from_sd_images_to_vector(sd_image_t* images
    return avi_data;
 }

-int create_mjpg_avi_from_sd_images(const char* filename, sd_image_t* images, int num_images, int fps, int quality) {
-    std::vector<uint8_t> avi_data = create_mjpg_avi_from_sd_images_to_vector(images, num_images, fps, quality);
+int create_mjpg_avi_from_sd_images(const char* filename, sd_image_t* images, int num_images, int fps, int quality, const sd_audio_t* audio) {
+    std::vector<uint8_t> avi_data = create_mjpg_avi_from_sd_images_to_vector(images, num_images, fps, quality, audio);
    if (avi_data.empty()) {
        return -1;
    }
@@ -1044,7 +1143,7 @@ int create_animated_webp_from_sd_images(const char* filename, sd_image_t* images
 #endif

 #ifdef SD_USE_WEBM
-std::vector<uint8_t> create_webm_from_sd_images_to_vector(sd_image_t* images, int num_images, int fps, int quality) {
+std::vector<uint8_t> create_webm_from_sd_images_to_vector(sd_image_t* images, int num_images, int fps, int quality, const sd_audio_t* audio) {
    if (num_images == 0) {
        fprintf(stderr, "Error: Image array is empty.\n");
        return {};
@@ -1089,6 +1188,25 @@ std::vector<uint8_t> create_webm_from_sd_images_to_vector(sd_image_t* images, in
            video_track->set_display_height(static_cast<uint64_t>(height));
            video_track->set_frame_rate(static_cast<double>(fps));
        }
+
+        uint64_t audio_track_number    = 0;
+        std::vector<uint8_t> audio_pcm = audio_to_pcm16_bytes(audio);
+        if (audio != nullptr && !audio_pcm.empty()) {
+            audio_track_number = segment.AddAudioTrack(static_cast<int32_t>(audio->sample_rate), static_cast<int32_t>(audio->channels), 0);
+            if (audio_track_number == 0) {
+                fprintf(stderr, "Error: Failed to add audio track.\n");
+                return -1;
+            }
+            auto* audio_track = static_cast<mkvmuxer::AudioTrack*>(segment.GetTrackByNumber(audio_track_number));
+            if (audio_track == nullptr) {
+                fprintf(stderr, "Error: Failed to get audio track.\n");
+                return -1;
+            }
+            audio_track->set_codec_id("A_PCM/INT/LIT");
+            audio_track->set_bit_depth(16);
+            audio_track->set_sample_rate(static_cast<double>(audio->sample_rate));
+            audio_track->set_channels(audio->channels);
+        }
        segment.GetSegmentInfo()->set_writing_app("stable-diffusion.cpp");
        segment.GetSegmentInfo()->set_muxing_app("stable-diffusion.cpp");

@@ -1118,6 +1236,23 @@ std::vector<uint8_t> create_webm_from_sd_images_to_vector(sd_image_t* images, in
                return -1;
            }

+            if (audio_track_number != 0) {
+                auto [audio_begin, audio_end] = audio_sample_range_for_video_frame(audio, i, num_images, fps);
+                const uint64_t frame_samples  = audio_end - audio_begin;
+                if (frame_samples > 0) {
+                    const uint64_t frame_bytes = frame_samples * audio->channels * sizeof(int16_t);
+                    const uint8_t* frame_ptr   = audio_pcm.data() + audio_begin * audio->channels * sizeof(int16_t);
+                    if (!segment.AddFrame(frame_ptr,
+                                          frame_bytes,
+                                          audio_track_number,
+                                          timestamp_ns,
+                                          true)) {
+                        fprintf(stderr, "Error: Failed to mux audio chunk %d into WebM.\n", i);
+                        return -1;
+                    }
+                }
+            }
+
            timestamp_ns += frame_duration_ns;
        }

@@ -1133,8 +1268,8 @@ std::vector<uint8_t> create_webm_from_sd_images_to_vector(sd_image_t* images, in
    return writer.data();
 }

-int create_webm_from_sd_images(const char* filename, sd_image_t* images, int num_images, int fps, int quality) {
-    std::vector<uint8_t> webm_data = create_webm_from_sd_images_to_vector(images, num_images, fps, quality);
+int create_webm_from_sd_images(const char* filename, sd_image_t* images, int num_images, int fps, int quality, const sd_audio_t* audio) {
+    std::vector<uint8_t> webm_data = create_webm_from_sd_images_to_vector(images, num_images, fps, quality, audio);
    if (webm_data.empty()) {
        return -1;
    }
@@ -1150,7 +1285,8 @@ std::vector<uint8_t> create_video_from_sd_images_to_vector(const std::string& ou
                                                           sd_image_t* images,
                                                           int num_images,
                                                           int fps,
-                                                           int quality) {
+                                                           int quality,
+                                                           const sd_audio_t* audio) {
    std::string format = output_format;
    std::transform(format.begin(), format.end(), format.begin(),
                   [](unsigned char c) { return static_cast<char>(tolower(c)); });
@@ -1160,7 +1296,7 @@ std::vector<uint8_t> create_video_from_sd_images_to_vector(const std::string& ou

 #ifdef SD_USE_WEBM
    if (format == "webm") {
-        return create_webm_from_sd_images_to_vector(images, num_images, fps, quality);
+        return create_webm_from_sd_images_to_vector(images, num_images, fps, quality, audio);
    }
 #endif

@@ -1170,14 +1306,14 @@ std::vector<uint8_t> create_video_from_sd_images_to_vector(const std::string& ou
    }
 #endif

-    return create_mjpg_avi_from_sd_images_to_vector(images, num_images, fps, quality);
+    return create_mjpg_avi_from_sd_images_to_vector(images, num_images, fps, quality, audio);
 }

-int create_video_from_sd_images(const char* filename, sd_image_t* images, int num_images, int fps, int quality) {
+int create_video_from_sd_images(const char* filename, sd_image_t* images, int num_images, int fps, int quality, const sd_audio_t* audio) {
    std::string path                = filename ? filename : "";
    auto pos                        = path.find_last_of('.');
    std::string ext                 = pos == std::string::npos ? "" : path.substr(pos);
-    std::vector<uint8_t> video_data = create_video_from_sd_images_to_vector(ext, images, num_images, fps, quality);
+    std::vector<uint8_t> video_data = create_video_from_sd_images_to_vector(ext, images, num_images, fps, quality, audio);
    if (video_data.empty()) {
        return -1;
    }
@@ -1187,3 +1323,54 @@ int create_video_from_sd_images(const char* filename, sd_image_t* images, int nu
    }
    return 0;
 }
+
+bool write_wav_to_file(const std::string& path,
+                       const float* interleaved_samples,
+                       uint64_t sample_count,
+                       uint32_t channels,
+                       uint32_t sample_rate) {
+    if (interleaved_samples == nullptr || sample_count == 0 || channels == 0 || sample_rate == 0) {
+        return false;
+    }
+
+    std::ofstream file(path, std::ios::binary);
+    if (!file.is_open()) {
+        return false;
+    }
+
+    uint32_t bits_per_sample  = 16;
+    uint32_t bytes_per_sample = bits_per_sample / 8;
+    uint32_t block_align      = channels * bytes_per_sample;
+    uint32_t byte_rate        = sample_rate * block_align;
+    uint32_t data_size        = static_cast<uint32_t>(sample_count * channels * bytes_per_sample);
+    uint32_t riff_size        = 36 + data_size;
+
+    file.write("RIFF", 4);
+    file.write(reinterpret_cast<const char*>(&riff_size), sizeof(riff_size));
+    file.write("WAVE", 4);
+    file.write("fmt ", 4);
+
+    uint32_t fmt_size            = 16;
+    uint16_t audio_format        = 1;
+    uint16_t wav_channels        = static_cast<uint16_t>(channels);
+    uint16_t wav_block_align     = static_cast<uint16_t>(block_align);
+    uint16_t wav_bits_per_sample = static_cast<uint16_t>(bits_per_sample);
+    file.write(reinterpret_cast<const char*>(&fmt_size), sizeof(fmt_size));
+    file.write(reinterpret_cast<const char*>(&audio_format), sizeof(audio_format));
+    file.write(reinterpret_cast<const char*>(&wav_channels), sizeof(wav_channels));
+    file.write(reinterpret_cast<const char*>(&sample_rate), sizeof(sample_rate));
+    file.write(reinterpret_cast<const char*>(&byte_rate), sizeof(byte_rate));
+    file.write(reinterpret_cast<const char*>(&wav_block_align), sizeof(wav_block_align));
+    file.write(reinterpret_cast<const char*>(&wav_bits_per_sample), sizeof(wav_bits_per_sample));
+
+    file.write("data", 4);
+    file.write(reinterpret_cast<const char*>(&data_size), sizeof(data_size));
+
+    std::vector<int16_t> pcm(sample_count * channels);
+    for (size_t i = 0; i < pcm.size(); ++i) {
+        float sample = std::max(-1.0f, std::min(1.0f, interleaved_samples[i]));
+        pcm[i]       = static_cast<int16_t>(std::lrint(sample * 32767.0f));
+    }
+    file.write(reinterpret_cast<const char*>(pcm.data()), static_cast<std::streamsize>(pcm.size() * sizeof(int16_t)));
+    return file.good();
+}
@@ -57,11 +57,13 @@ int create_mjpg_avi_from_sd_images(const char* filename,
                                   sd_image_t* images,
                                   int num_images,
                                   int fps,
-                                   int quality = 90);
+                                   int quality             = 90,
+                                   const sd_audio_t* audio = nullptr);
 std::vector<uint8_t> create_mjpg_avi_from_sd_images_to_vector(sd_image_t* images,
                                                              int num_images,
                                                              int fps,
-                                                              int quality = 90);
+                                                              int quality             = 90,
+                                                              const sd_audio_t* audio = nullptr);

 #ifdef SD_USE_WEBP
 int create_animated_webp_from_sd_images(const char* filename,
@@ -80,22 +82,32 @@ int create_webm_from_sd_images(const char* filename,
                               sd_image_t* images,
                               int num_images,
                               int fps,
-                               int quality = 90);
+                               int quality             = 90,
+                               const sd_audio_t* audio = nullptr);
 std::vector<uint8_t> create_webm_from_sd_images_to_vector(sd_image_t* images,
                                                          int num_images,
                                                          int fps,
-                                                          int quality = 90);
+                                                          int quality             = 90,
+                                                          const sd_audio_t* audio = nullptr);
 #endif

 int create_video_from_sd_images(const char* filename,
                                sd_image_t* images,
                                int num_images,
                                int fps,
-                                int quality = 90);
+                                int quality             = 90,
+                                const sd_audio_t* audio = nullptr);
 std::vector<uint8_t> create_video_from_sd_images_to_vector(const std::string& output_format,
                                                           sd_image_t* images,
                                                           int num_images,
                                                           int fps,
-                                                           int quality = 90);
+                                                           int quality             = 90,
+                                                           const sd_audio_t* audio = nullptr);
+
+bool write_wav_to_file(const std::string& path,
+                       const float* interleaved_samples,
+                       uint64_t sample_count,
+                       uint32_t channels,
+                       uint32_t sample_rate);

 #endif  // __MEDIA_IO_H__
@@ -74,6 +74,9 @@ add_executable(${TARGET}
    routes_sdapi.cpp
    routes_sdcpp.cpp
 )
+if(APPLE)
+    sd_set_macos_rpaths(${TARGET})
+endif()

 if(HAVE_FRONTEND_BUILD)
    add_dependencies(${TARGET} ${TARGET}_frontend)
@@ -143,6 +143,8 @@ Context Options:
  --qwen2vl_vision <string>                alias of --llm_vision. Deprecated.
  --diffusion-model <string>               path to the standalone diffusion model
  --high-noise-diffusion-model <string>    path to the standalone high noise diffusion model
+  --uncond-diffusion-model <string>        path to the standalone unconditional diffusion model, currently used by
+                                           Ideogram4 CFG
  --vae <string>                           path to standalone vae model
  --taesd <string>                         path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)
  --tae <string>                           alias of --taesd
@@ -157,7 +159,8 @@ Context Options:
                                           then threads will be set to the number of CPU physical cores
  --chroma-t5-mask-pad <int>               t5 mask pad size of chroma
  --max-vram <float>                       maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables
-                                           graph splitting
+                                           graph splitting; a negative value auto-detects free VRAM, sparing the
+                                           specified value (e.g. -0.5 will keep at least 0.5 GiB free)
  --force-sdxl-vae-conv-scale              force use of conv scale on sdxl vae
  --offload-to-cpu                         place the weights in RAM to save VRAM, and automatically load them into VRAM
                                           when needed
@@ -205,6 +208,12 @@ Default Generation Options:
  --hires-upscaler <string>                highres fix upscaler, Lanczos, Nearest, Latent, Latent (nearest), Latent
                                           (nearest-exact), Latent (antialiased), Latent (bicubic), Latent (bicubic
                                           antialiased), or a model name under --hires-upscalers-dir (default: Latent)
+  --extra-sample-args <string>             extra sampler/scheduler/guidance args, key=value list. APG supports apg_eta,
+                                           apg_momentum, apg_norm_threshold, apg_norm_threshold_smoothing; SLG supports
+                                           slg_uncond; lcm supports noise_clip_std, noise_scale_start, noise_scale_end;
+                                           ltx2 supports max_shift, base_shift, stretch, terminal; euler_ge supports gamma
+  --extra-tiling-args <string>             extra VAE tiling args, key=value list. LTX video VAE supports
+                                           temporal_tile_frames (default: 4), temporal_tile_overlap (default: 1)
  -H, --height <int>                       image height, in pixel space (default: 512)
  -W, --width <int>                        image width, in pixel space (default: 512)
  --steps <int>                            number of sample steps (default: 20)
@@ -224,8 +233,8 @@ Default Generation Options:
  --hires-upscale-tile-size <int>          highres fix upscaler tile size, reserved for model-backed upscalers (default:
                                           128)
  --cfg-scale <float>                      unconditional guidance scale: (default: 7.0)
-  --img-cfg-scale <float>                  image guidance scale for inpaint or instruct-pix2pix models: (default: same
-                                           as --cfg-scale)
+  --img-cfg-scale <float>                  image guidance scale for inpaint or image edit models: (default: same as
+                                           --cfg-scale)
  --guidance <float>                       distilled guidance scale for models with guidance input (default: 3.5)
  --slg-scale <float>                      skip layer guidance (SLG) scale, only for DiT models: (default: 0). 0 means
                                           disabled, a value of 2.5 is nice for sd3.5 medium
@@ -235,8 +244,8 @@ Default Generation Options:
                                           res_2s; 1 for euler_a, er_sde and dpm++2s_a)
  --flow-shift <float>                     shift value for Flow models like SD3.x or WAN (default: auto)
  --high-noise-cfg-scale <float>           (high noise) unconditional guidance scale: (default: 7.0)
-  --high-noise-img-cfg-scale <float>       (high noise) image guidance scale for inpaint or instruct-pix2pix models
-                                           (default: same as --cfg-scale)
+  --high-noise-img-cfg-scale <float>       (high noise) image guidance scale for inpaint or image edit models (default:
+                                           same as --cfg-scale)
  --high-noise-guidance <float>            (high noise) distilled guidance scale for models with guidance input
                                           (default: 3.5)
  --high-noise-slg-scale <float>           (high noise) skip layer guidance (SLG) scale, only for DiT models: (default:
@@ -246,7 +255,7 @@ Default Generation Options:
  --high-noise-eta <float>                 (high noise) noise multiplier (default: 0 for ddim_trailing, tcd,
                                           res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)
  --strength <float>                       strength for noising/unnoising (default: 0.75)
-  --pm-style-strength <float>              
+  --pm-style-strength <float>
  --control-strength <float>               strength to apply Control Net (default: 0.9). 1.0 corresponds to full
                                           destruction of information in init image
  --moe-boundary <float>                   timestep boundary for Wan2.2 MoE model. (default: 0.875). Only enabled if
@@ -260,19 +269,22 @@ Default Generation Options:
  --disable-auto-resize-ref-image          disable auto resize of ref images
  --disable-image-metadata                 do not embed generation metadata on image files
  --vae-tiling                             process vae in tiles to reduce memory usage
+  --temporal-tiling                        enable temporal tiling for LTX video VAE decode
  --hires                                  enable highres fix
  -s, --seed                               RNG seed (default: 42, use random seed for < 0)
  --sampling-method                        sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m,
                                           dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s,
-                                           er_sde] (default: euler for Flux/SD3/Wan, euler_a otherwise)
+                                           er_sde, euler_cfg_pp, euler_a_cfg_pp] (default: euler for Flux/SD3/Wan, euler_a otherwise)
  --high-noise-sampling-method             (high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a,
                                           dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep,
-                                           res_2s, er_sde] default: euler for Flux/SD3/Wan, euler_a otherwise
+                                           res_2s, er_sde, euler_cfg_pp, euler_a_cfg_pp] default: euler for Flux/SD3/Wan, euler_a otherwise
  --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits,
-                                           smoothstep, sgm_uniform, simple, kl_optimal, lcm, bong_tangent], default:
-                                           discrete
+                                           smoothstep, sgm_uniform, simple, kl_optimal, lcm, bong_tangent, ltx2], default:
+                                           model-specific
  --sigmas                                 custom sigma values for the sampler, comma-separated (e.g.,
                                           "14.61,7.8,3.5,0.0").
+  --hires-sigmas                           custom sigma values for the highres fix second pass, comma-separated (e.g.,
+                                           "0.85,0.725,0.421875,0.0").
  --skip-layers                            layers to skip for SLG steps (default: [7,8,9])
  --high-noise-skip-layers                 (high noise) layers to skip for SLG steps (default: [7,8,9])
  -r, --ref-image                          reference image for Flux Kontext models (can be used multiple times)
@@ -504,11 +504,13 @@ Shared default fields used by both `img_gen` and `vid_gen`:
 | `sample_params.guidance.slg.scale` | `number` |
 | `vae_tiling_params` | `object` |
 | `vae_tiling_params.enabled` | `boolean` |
+| `vae_tiling_params.temporal_tiling` | `boolean` |
 | `vae_tiling_params.tile_size_x` | `integer` |
 | `vae_tiling_params.tile_size_y` | `integer` |
 | `vae_tiling_params.target_overlap` | `number` |
 | `vae_tiling_params.rel_size_x` | `number` |
 | `vae_tiling_params.rel_size_y` | `number` |
+| `vae_tiling_params.extra_tiling_args` | `string` |
 | `cache_mode` | `string` |
 | `cache_option` | `string` |
 | `scm_mask` | `string` |
@@ -516,6 +518,8 @@ Shared default fields used by both `img_gen` and `vid_gen`:
 | `output_format` | `string` |
 | `output_compression` | `integer` |

+`vae_tiling_params.extra_tiling_args` accepts a key=value list. For LTX video VAE temporal tiling, `temporal_tile_frames` defaults to `4` and `temporal_tile_overlap` defaults to `1`.
+
 `img_gen`-specific default fields:

 | Field | Type |
@@ -532,6 +536,7 @@ Shared default fields used by both `img_gen` and `vid_gen`:
 | `hires.target_height` | `integer` |
 | `hires.steps` | `integer` |
 | `hires.denoising_strength` | `number` |
+| `hires.custom_sigmas` | `array<number>` |
 | `hires.upscale_tile_size` | `integer` |

 `vid_gen`-specific default fields:
@@ -685,16 +690,19 @@ Example:
    "target_height": 0,
    "steps": 0,
    "denoising_strength": 0.7,
+    "custom_sigmas": [],
    "upscale_tile_size": 128
  },

  "vae_tiling_params": {
    "enabled": false,
+    "temporal_tiling": false,
    "tile_size_x": 0,
    "tile_size_y": 0,
    "target_overlap": 0.5,
    "rel_size_x": 0.0,
-    "rel_size_y": 0.0
+    "rel_size_y": 0.0,
+    "extra_tiling_args": ""
  },

  "cache_mode": "disabled",
@@ -799,14 +807,23 @@ Other native fields:
 | `hires.target_height` | `integer` |
 | `hires.steps` | `integer` |
 | `hires.denoising_strength` | `number` |
+| `hires.custom_sigmas` | `array<number>` |
 | `hires.upscale_tile_size` | `integer` |
 | `vae_tiling_params` | `object` |
+| `vae_tiling_params.enabled` | `boolean` |
+| `vae_tiling_params.temporal_tiling` | `boolean` |
+| `vae_tiling_params.tile_size_x` | `integer` |
+| `vae_tiling_params.tile_size_y` | `integer` |
+| `vae_tiling_params.target_overlap` | `number` |
+| `vae_tiling_params.rel_size_x` | `number` |
+| `vae_tiling_params.rel_size_y` | `number` |
+| `vae_tiling_params.extra_tiling_args` | `string` |
 | `cache_mode` | `string` |
 | `cache_option` | `string` |
 | `scm_mask` | `string` |
 | `scm_policy_dynamic` | `boolean` |

-For `hires.upscaler`, use `Lanczos`, `Nearest`, `Latent`, `Latent (nearest)`, `Latent (nearest-exact)`, `Latent (antialiased)`, `Latent (bicubic)`, `Latent (bicubic antialiased)`, or an `upscalers[].name` value from `GET /sdcpp/v1/capabilities`. Model-backed upscalers are resolved as `--hires-upscalers-dir / (name + ext)` and must live directly in that directory.
+For `hires.upscaler`, use `Lanczos`, `Nearest`, `Latent`, `Latent (nearest)`, `Latent (nearest-exact)`, `Latent (antialiased)`, `Latent (bicubic)`, `Latent (bicubic antialiased)`, or an `upscalers[].name` value from `GET /sdcpp/v1/capabilities`. Model-backed upscalers are resolved as `--hires-upscalers-dir / (name + ext)` and must live directly in that directory. `hires.custom_sigmas`, when present, overrides the generated second-pass hires sigma schedule; otherwise the hires schedule is trimmed by `hires.denoising_strength`.

 HTTP-only output fields:

@@ -1009,11 +1026,13 @@ Example:

  "vae_tiling_params": {
    "enabled": false,
+    "temporal_tiling": false,
    "tile_size_x": 0,
    "tile_size_y": 0,
    "target_overlap": 0.5,
    "rel_size_x": 0.0,
-    "rel_size_y": 0.0
+    "rel_size_y": 0.0,
+    "extra_tiling_args": ""
  },

  "cache_mode": "disabled",
@@ -1131,6 +1150,14 @@ Other native fields:
 | Field | Type |
 | --- | --- |
 | `vae_tiling_params` | `object` |
+| `vae_tiling_params.enabled` | `boolean` |
+| `vae_tiling_params.temporal_tiling` | `boolean` |
+| `vae_tiling_params.tile_size_x` | `integer` |
+| `vae_tiling_params.tile_size_y` | `integer` |
+| `vae_tiling_params.target_overlap` | `number` |
+| `vae_tiling_params.rel_size_x` | `number` |
+| `vae_tiling_params.rel_size_y` | `number` |
+| `vae_tiling_params.extra_tiling_args` | `string` |
 | `cache_mode` | `string` |
 | `cache_option` | `string` |
 | `scm_mask` | `string` |
@@ -231,16 +231,21 @@ bool execute_vid_gen_job(ServerRuntime& runtime,
    sd_vid_gen_params_t params = job.vid_gen.to_sd_vid_gen_params_t();

    SDImageVec results;
-    int num_results = 0;
+    int num_results             = 0;
+    sd_audio_t* generated_audio = nullptr;

    {
        std::lock_guard<std::mutex> lock(*runtime.sd_ctx_mutex);
-        sd_image_t* raw_results = generate_video(runtime.sd_ctx, &params, &num_results);
+        sd_image_t* raw_results = nullptr;
+        if (!generate_video(runtime.sd_ctx, &params, &raw_results, &num_results, &generated_audio)) {
+            raw_results = nullptr;
+        }
        results.adopt(raw_results, num_results);
    }

    num_results = results.count();
    if (num_results <= 0) {
+        free_sd_audio(generated_audio);
        error_message = "generate_video returned no results";
        return false;
    }
@@ -249,7 +254,9 @@ bool execute_vid_gen_job(ServerRuntime& runtime,
                                                                             results.data(),
                                                                             num_results,
                                                                             job.vid_gen.gen_params.fps,
-                                                                             job.vid_gen.output_compression);
+                                                                             job.vid_gen.output_compression,
+                                                                             generated_audio);
+    free_sd_audio(generated_audio);
    if (video_bytes.empty()) {
        error_message = "failed to encode generated video container";
        return false;
@@ -145,7 +145,7 @@ int main(int argc, const char** argv) {
    register_sdapi_endpoints(svr, runtime);
    register_sdcpp_api_endpoints(svr, runtime);

-    LOG_INFO("listening on: %s:%d\n", svr_params.listen_ip.c_str(), svr_params.listen_port);
+    LOG_INFO("listening on: http://%s:%d\n", svr_params.listen_ip.c_str(), svr_params.listen_port);
    svr.listen(svr_params.listen_ip, svr_params.listen_port);

    {
@@ -67,6 +67,10 @@ static enum sample_method_t get_sdapi_sample_method(std::string name) {
        {"k_res_multistep", RES_MULTISTEP_SAMPLE_METHOD},
        {"res 2s", RES_2S_SAMPLE_METHOD},
        {"k_res_2s", RES_2S_SAMPLE_METHOD},
+        {"euler_cfg_pp", EULER_CFG_PP_SAMPLE_METHOD},
+        {"k_euler_cfg_pp", EULER_CFG_PP_SAMPLE_METHOD},
+        {"euler_a_cfg_pp", EULER_CFG_PP_SAMPLE_METHOD},
+        {"k_euler_a_cfg_pp", EULER_CFG_PP_SAMPLE_METHOD},
    };
    auto it = hardcoded.find(name);
    return it != hardcoded.end() ? it->second : SAMPLE_METHOD_COUNT;
@@ -56,11 +56,13 @@ static const char* capability_sample_method_name(enum sample_method_t sample_met
 static json make_vae_tiling_json(const sd_tiling_params_t& params) {
    return {
        {"enabled", params.enabled},
+        {"temporal_tiling", params.temporal_tiling},
        {"tile_size_x", params.tile_size_x},
        {"tile_size_y", params.tile_size_y},
        {"target_overlap", params.target_overlap},
        {"rel_size_x", params.rel_size_x},
        {"rel_size_y", params.rel_size_y},
+        {"extra_tiling_args", params.extra_tiling_args ? params.extra_tiling_args : ""},
    };
 }

@@ -100,6 +102,20 @@ static json make_sample_params_json(const sd_sample_params_t& sample_params, con
    };
 }

+static json make_hires_json(const SDGenerationParams& defaults) {
+    return {
+        {"enabled", defaults.hires_enabled},
+        {"upscaler", defaults.hires_upscaler},
+        {"scale", defaults.hires_scale},
+        {"target_width", defaults.hires_width},
+        {"target_height", defaults.hires_height},
+        {"steps", defaults.hires_steps},
+        {"denoising_strength", defaults.hires_denoising_strength},
+        {"custom_sigmas", defaults.hires_custom_sigmas},
+        {"upscale_tile_size", defaults.hires_upscale_tile_size},
+    };
+}
+
 static json make_img_gen_defaults_json(const SDGenerationParams& defaults, const std::string& output_format) {
    return {
        {"prompt", defaults.prompt},
@@ -114,17 +130,7 @@ static json make_img_gen_defaults_json(const SDGenerationParams& defaults, const
        {"increase_ref_index", defaults.increase_ref_index},
        {"control_strength", defaults.control_strength},
        {"sample_params", make_sample_params_json(defaults.sample_params, defaults.skip_layers)},
-        {"hires",
-         {
-             {"enabled", defaults.hires_enabled},
-             {"upscaler", defaults.hires_upscaler},
-             {"scale", defaults.hires_scale},
-             {"target_width", defaults.hires_width},
-             {"target_height", defaults.hires_height},
-             {"steps", defaults.hires_steps},
-             {"denoising_strength", defaults.hires_denoising_strength},
-             {"upscale_tile_size", defaults.hires_upscale_tile_size},
-         }},
+        {"hires", make_hires_json(defaults)},
        {"vae_tiling_params", make_vae_tiling_json(defaults.vae_tiling_params)},
        {"cache_mode", defaults.cache_mode},
        {"cache_option", defaults.cache_option},
@@ -150,6 +156,7 @@ static json make_vid_gen_defaults_json(const SDGenerationParams& defaults, const
        {"vace_strength", defaults.vace_strength},
        {"sample_params", make_sample_params_json(defaults.sample_params, defaults.skip_layers)},
        {"high_noise_sample_params", make_sample_params_json(defaults.high_noise_sample_params, defaults.high_noise_skip_layers)},
+        {"hires", make_hires_json(defaults)},
        {"vae_tiling_params", make_vae_tiling_json(defaults.vae_tiling_params)},
        {"cache_mode", defaults.cache_mode},
        {"cache_option", defaults.cache_option},
@@ -203,8 +203,9 @@ ArgOptions SDSvrParams::get_options() {
        {"", "--color", "colors the logging tags according to level", true, &color},
    };

-    auto on_help_arg = [&](int, const char**, int) {
+    auto on_help_arg = [&](int, const char**, int, bool& valid) {
        normal_exit = true;
+        valid       = true;
        return -1;
    };

@@ -254,7 +255,7 @@ void refresh_lora_cache(ServerRuntime& rt) {

    fs::path lora_dir = rt.ctx_params->lora_model_dir;
    if (fs::exists(lora_dir) && fs::is_directory(lora_dir)) {
-        for (auto& entry : fs::recursive_directory_iterator(lora_dir)) {
+        for (auto& entry : fs::recursive_directory_iterator(lora_dir, fs::directory_options::skip_permission_denied)) {
            if (!entry.is_regular_file()) {
                continue;
            }
@@ -51,6 +51,9 @@ enum sample_method_t {
    RES_MULTISTEP_SAMPLE_METHOD,
    RES_2S_SAMPLE_METHOD,
    ER_SDE_SAMPLE_METHOD,
+    EULER_CFG_PP_SAMPLE_METHOD,
+    EULER_A_CFG_PP_SAMPLE_METHOD,
+    EULER_GE_SAMPLE_METHOD,
    SAMPLE_METHOD_COUNT
 };

@@ -66,6 +69,7 @@ enum scheduler_t {
    KL_OPTIMAL_SCHEDULER,
    LCM_SCHEDULER,
    BONG_TANGENT_SCHEDULER,
+    LTX2_SCHEDULER,
    SCHEDULER_COUNT
 };

@@ -122,7 +126,8 @@ enum sd_type_t {
    // SD_TYPE_IQ4_NL_8_8 = 38,
    SD_TYPE_MXFP4 = 39,  // MXFP4 (1 block)
    SD_TYPE_NVFP4 = 40,  // NVFP4 (4 blocks, E4M3 scale)
-    SD_TYPE_COUNT = 41,
+    SD_TYPE_Q1_0  = 41,
+    SD_TYPE_COUNT = 42,
 };

 enum sd_log_level_t {
@@ -149,11 +154,13 @@ enum lora_apply_mode_t {

 typedef struct {
    bool enabled;
+    bool temporal_tiling;
    int tile_size_x;
    int tile_size_y;
    float target_overlap;
    float rel_size_x;
    float rel_size_y;
+    const char* extra_tiling_args;
 } sd_tiling_params_t;

 typedef struct {
@@ -161,6 +168,14 @@ typedef struct {
    const char* path;
 } sd_embedding_t;

+enum sd_vae_format_t {
+    SD_VAE_FORMAT_AUTO = -1,
+    SD_VAE_FORMAT_FLUX,
+    SD_VAE_FORMAT_SD3,
+    SD_VAE_FORMAT_FLUX2,
+    SD_VAE_FORMAT_COUNT,
+};
+
 typedef struct {
    const char* model_path;
    const char* clip_l_path;
@@ -171,7 +186,10 @@ typedef struct {
    const char* llm_vision_path;
    const char* diffusion_model_path;
    const char* high_noise_diffusion_model_path;
+    const char* uncond_diffusion_model_path;
+    const char* embeddings_connectors_path;
    const char* vae_path;
+    const char* audio_vae_path;
    const char* taesd_path;
    const char* control_net_path;
    const sd_embedding_t* embeddings;
@@ -203,9 +221,20 @@ typedef struct {
    bool chroma_use_t5_mask;
    int chroma_t5_mask_pad;
    bool qwen_image_zero_cond_t;
-    float max_vram;
+    enum sd_vae_format_t vae_format;
+    float max_vram;  // GiB budget for graph-cut segmented param offload (0 = disabled, -1 = auto free VRAM minus 1 GiB)
+    bool stream_layers;  // Enable residency+prefetch streaming on top of --max-vram (no effect without --max-vram)
+    const char* backend;
+    const char* params_backend;
 } sd_ctx_params_t;

+typedef struct {
+    uint32_t sample_rate;
+    uint32_t channels;
+    uint64_t sample_count;
+    float* data;
+} sd_audio_t;
+
 typedef struct {
    uint32_t width;
    uint32_t height;
@@ -238,6 +267,7 @@ typedef struct {
    float* custom_sigmas;
    int custom_sigmas_count;
    float flow_shift;
+    const char* extra_sample_args;
 } sd_sample_params_t;

 typedef struct {
@@ -314,6 +344,8 @@ typedef struct {
    int steps;
    float denoising_strength;
    int upscale_tile_size;
+    float* custom_sigmas;
+    int custom_sigmas_count;
 } sd_hires_params_t;

 typedef struct {
@@ -360,9 +392,11 @@ typedef struct {
    float strength;
    int64_t seed;
    int video_frames;
+    int fps;
    float vace_strength;
    sd_tiling_params_t vae_tiling_params;
    sd_cache_params_t cache;
+    sd_hires_params_t hires;
 } sd_vid_gen_params_t;

 typedef struct sd_ctx_t sd_ctx_t;
@@ -404,6 +438,7 @@ SD_API char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params);

 SD_API sd_ctx_t* new_sd_ctx(const sd_ctx_params_t* sd_ctx_params);
 SD_API void free_sd_ctx(sd_ctx_t* sd_ctx);
+SD_API void free_sd_audio(sd_audio_t* audio);

 SD_API void sd_sample_params_init(sd_sample_params_t* sample_params);
 SD_API char* sd_sample_params_to_str(const sd_sample_params_t* sample_params);
@@ -416,7 +451,11 @@ SD_API char* sd_img_gen_params_to_str(const sd_img_gen_params_t* sd_img_gen_para
 SD_API sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* sd_img_gen_params);

 SD_API void sd_vid_gen_params_init(sd_vid_gen_params_t* sd_vid_gen_params);
-SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* sd_vid_gen_params, int* num_frames_out);
+SD_API bool generate_video(sd_ctx_t* sd_ctx,
+                           const sd_vid_gen_params_t* sd_vid_gen_params,
+                           sd_image_t** frames_out,
+                           int* num_frames_out,
+                           sd_audio_t** audio_out);

 typedef struct upscaler_ctx_t upscaler_ctx_t;

@@ -424,7 +463,9 @@ SD_API upscaler_ctx_t* new_upscaler_ctx(const char* esrgan_path,
                                        bool offload_params_to_cpu,
                                        bool direct,
                                        int n_threads,
-                                        int tile_size);
+                                        int tile_size,
+                                        const char* backend,
+                                        const char* params_backend);
 SD_API void free_upscaler_ctx(upscaler_ctx_t* upscaler_ctx);

 SD_API sd_image_t upscale(upscaler_ctx_t* upscaler_ctx,
@@ -0,0 +1,283 @@
+#!/usr/bin/env python
+import argparse
+import json
+import math
+import os
+import struct
+from collections import Counter
+from pathlib import Path
+
+import torch
+from safetensors import safe_open
+
+
+FLOAT_DTYPES = {
+    "BF16",
+    "F16",
+    "F32",
+    "F64",
+    "F8_E4M3",
+    "F8_E4M3FN",
+    "F8_E5M2",
+}
+
+FP8_DTYPES = {
+    "F8_E4M3",
+    "F8_E4M3FN",
+    "F8_E5M2",
+}
+
+DTYPE_SIZES = {
+    "BOOL": 1,
+    "U8": 1,
+    "I8": 1,
+    "F8_E4M3": 1,
+    "F8_E4M3FN": 1,
+    "F8_E5M2": 1,
+    "U16": 2,
+    "I16": 2,
+    "F16": 2,
+    "BF16": 2,
+    "U32": 4,
+    "I32": 4,
+    "F32": 4,
+    "U64": 8,
+    "I64": 8,
+    "F64": 8,
+}
+
+
+def read_safetensors_header(path: Path):
+    with path.open("rb") as f:
+        header_len = struct.unpack("<Q", f.read(8))[0]
+        header = f.read(header_len).decode("utf-8").rstrip()
+    return json.loads(header)
+
+
+def numel(shape):
+    return math.prod(shape) if shape else 1
+
+
+def scale_key_for_weight(name: str):
+    if name.endswith(".weight"):
+        return name[:-len(".weight")] + ".weight_scale"
+    if name.endswith("weight"):
+        return name + "_scale"
+    return None
+
+
+def tensor_nbytes(dtype: str, shape):
+    return numel(shape) * DTYPE_SIZES[dtype]
+
+
+def build_output_plan(header):
+    entries = {k: v for k, v in header.items() if k != "__metadata__"}
+    paired_scale_keys = set()
+    plan = []
+
+    for name, info in entries.items():
+        scale_key = scale_key_for_weight(name)
+        if info["dtype"] in FP8_DTYPES and scale_key in entries:
+            paired_scale_keys.add(scale_key)
+
+    for name, info in entries.items():
+        if name in paired_scale_keys:
+            continue
+
+        dtype = info["dtype"]
+        shape = info["shape"]
+        scale_key = scale_key_for_weight(name)
+
+        if dtype in FP8_DTYPES and scale_key in entries:
+            scale_info = entries[scale_key]
+            plan.append(
+                {
+                    "name": name,
+                    "source_dtype": dtype,
+                    "output_dtype": "BF16",
+                    "shape": shape,
+                    "mode": "fp8_scaled_weight",
+                    "scale_key": scale_key,
+                }
+            )
+            continue
+
+        if dtype in FLOAT_DTYPES:
+            plan.append(
+                {
+                    "name": name,
+                    "source_dtype": dtype,
+                    "output_dtype": "BF16",
+                    "shape": shape,
+                    "mode": "float_to_bf16",
+                }
+            )
+        else:
+            plan.append(
+                {
+                    "name": name,
+                    "source_dtype": dtype,
+                    "output_dtype": dtype,
+                    "shape": shape,
+                    "mode": "copy",
+                }
+            )
+
+    metadata = dict(header.get("__metadata__", {}) or {})
+    metadata["format"] = "pt"
+    metadata["conversion"] = "fp8_weight_scale_to_bf16"
+
+    output_header = {"__metadata__": metadata}
+    offset = 0
+    for item in plan:
+        size = tensor_nbytes(item["output_dtype"], item["shape"])
+        output_header[item["name"]] = {
+            "dtype": item["output_dtype"],
+            "shape": item["shape"],
+            "data_offsets": [offset, offset + size],
+        }
+        offset += size
+
+    return plan, output_header, offset
+
+
+def write_tensor_bytes(out, tensor):
+    tensor = tensor.detach().cpu().contiguous()
+    if tensor.numel() == 0:
+        return
+    if tensor.dtype == torch.bfloat16:
+        tensor.view(torch.uint16).numpy().tofile(out)
+    elif tensor.dtype in (getattr(torch, "float8_e4m3fn", None), getattr(torch, "float8_e5m2", None)):
+        tensor.view(torch.uint8).numpy().tofile(out)
+    else:
+        tensor.numpy().tofile(out)
+
+
+def scale_view_for_chunk(scale, chunk, first_dim_start=0, first_dim_end=None):
+    scale = scale.to(torch.float32)
+
+    if scale.numel() == 1:
+        return scale.reshape((1,) * chunk.ndim)
+
+    if chunk.ndim > 0 and scale.ndim == 1:
+        if first_dim_end is not None and scale.shape[0] >= first_dim_end:
+            scale = scale[first_dim_start:first_dim_end]
+        if scale.shape[0] == chunk.shape[0]:
+            return scale.reshape((scale.shape[0],) + (1,) * (chunk.ndim - 1))
+
+    return scale
+
+
+def write_scaled_fp8_weight(out, weight, scale, chunk_rows):
+    if weight.ndim == 0:
+        result = weight.to(torch.float32) * scale_view_for_chunk(scale, weight)
+        write_tensor_bytes(out, result.to(torch.bfloat16))
+        return
+
+    rows = weight.shape[0]
+    for start in range(0, rows, chunk_rows):
+        end = min(start + chunk_rows, rows)
+        chunk = weight[start:end].to(torch.float32)
+        scale_view = scale_view_for_chunk(scale, chunk, start, end)
+        result = chunk * scale_view
+        write_tensor_bytes(out, result.to(torch.bfloat16))
+
+
+def write_float_as_bf16(out, tensor, chunk_rows):
+    if tensor.dtype == torch.bfloat16:
+        write_tensor_bytes(out, tensor)
+        return
+
+    if tensor.ndim == 0:
+        write_tensor_bytes(out, tensor.to(torch.bfloat16))
+        return
+
+    rows = tensor.shape[0]
+    for start in range(0, rows, chunk_rows):
+        end = min(start + chunk_rows, rows)
+        write_tensor_bytes(out, tensor[start:end].to(torch.bfloat16))
+
+
+def convert(input_path: Path, output_path: Path, chunk_rows: int, dry_run: bool):
+    header = read_safetensors_header(input_path)
+    plan, output_header, data_size = build_output_plan(header)
+
+    source_counts = Counter(item["source_dtype"] for item in plan)
+    output_counts = Counter(item["output_dtype"] for item in plan)
+    scaled_count = sum(item["mode"] == "fp8_scaled_weight" for item in plan)
+    dropped_scales = sum(item["mode"] == "fp8_scaled_weight" for item in plan)
+    header_bytes = json.dumps(output_header, separators=(",", ":")).encode("utf-8")
+    expected_size = 8 + len(header_bytes) + data_size
+
+    print(f"input:  {input_path}")
+    print(f"output: {output_path}")
+    print(f"tensors written: {len(plan)}")
+    print(f"scaled fp8 weights dequantized: {scaled_count}")
+    print(f"weight_scale tensors dropped: {dropped_scales}")
+    print(f"source dtypes: {dict(sorted(source_counts.items()))}")
+    print(f"output dtypes: {dict(sorted(output_counts.items()))}")
+    print(f"expected output size: {expected_size / (1024 ** 3):.2f} GiB")
+
+    if dry_run:
+        return
+
+    if output_path.exists():
+        raise FileExistsError(f"{output_path} already exists; pass --overwrite to replace it")
+
+    tmp_path = output_path.with_suffix(output_path.suffix + ".tmp")
+    if tmp_path.exists():
+        raise FileExistsError(f"{tmp_path} already exists; remove it or choose another output")
+
+    with safe_open(str(input_path), framework="pt", device="cpu") as sf, tmp_path.open("wb") as out:
+        out.write(struct.pack("<Q", len(header_bytes)))
+        out.write(header_bytes)
+
+        for index, item in enumerate(plan, 1):
+            name = item["name"]
+            print(f"[{index:04d}/{len(plan):04d}] {name} -> {item['output_dtype']}")
+
+            tensor = sf.get_tensor(name)
+            if item["mode"] == "fp8_scaled_weight":
+                scale = sf.get_tensor(item["scale_key"])
+                write_scaled_fp8_weight(out, tensor, scale, chunk_rows)
+            elif item["mode"] == "float_to_bf16":
+                write_float_as_bf16(out, tensor, chunk_rows)
+            else:
+                write_tensor_bytes(out, tensor)
+
+        actual_size = out.tell()
+
+    if actual_size != expected_size:
+        tmp_path.unlink(missing_ok=True)
+        raise RuntimeError(f"wrote {actual_size} bytes, expected {expected_size} bytes")
+
+    tmp_path.replace(output_path)
+    print("done")
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Convert an fp8 safetensors checkpoint with weight_scale tensors to bf16."
+    )
+    parser.add_argument("--input", default="ideogram4_fp8.safetensors", type=Path)
+    parser.add_argument("--output", default="ideogram4_bf16.safetensors", type=Path)
+    parser.add_argument("--chunk-rows", default=1024, type=int)
+    parser.add_argument("--dry-run", action="store_true")
+    parser.add_argument("--overwrite", action="store_true")
+    args = parser.parse_args()
+
+    input_path = args.input.resolve()
+    output_path = args.output.resolve()
+
+    if args.chunk_rows < 1:
+        raise ValueError("--chunk-rows must be >= 1")
+    if not input_path.exists():
+        raise FileNotFoundError(input_path)
+    if args.overwrite and output_path.exists():
+        output_path.unlink()
+
+    convert(input_path, output_path, args.chunk_rows, args.dry_run)
+
+
+if __name__ == "__main__":
+    main()
@@ -7,6 +7,7 @@
 #include <vector>

 #include "common_block.hpp"
+#include "diffusion_model.hpp"
 #include "flux.hpp"
 #include "rope.hpp"

@@ -518,7 +519,7 @@ namespace Anima {
        }
    };

-    struct AnimaRunner : public GGMLRunner {
+    struct AnimaRunner : public DiffusionModelRunner {
    public:
        std::vector<float> image_pe_vec;
        std::vector<float> adapter_q_pe_vec;
@@ -526,10 +527,10 @@ namespace Anima {
        AnimaNet net;

        AnimaRunner(ggml_backend_t backend,
-                    bool offload_params_to_cpu,
+                    ggml_backend_t params_backend,
                    const String2TensorStorage& tensor_storage_map = {},
                    const std::string prefix                       = "model.diffusion_model")
-            : GGMLRunner(backend, offload_params_to_cpu) {
+            : DiffusionModelRunner(backend, params_backend, prefix) {
            int64_t num_layers    = 0;
            std::string layer_tag = prefix + ".net.blocks.";
            for (const auto& kv : tensor_storage_map) {
@@ -559,7 +560,7 @@ namespace Anima {
            return "anima";
        }

-        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
            net.get_param_tensors(tensors, prefix + ".net");
        }

@@ -598,7 +599,8 @@ namespace Anima {
                                          {},
                                          empty_ref_latents,
                                          false,
-                                          1.0f);
+                                          1.0f,
+                                          false);

            std::vector<float> axis_thetas = {
                static_cast<float>(theta) * calc_ntk_factor(t_extrapolation_ratio, axes_dim[0]),
@@ -683,6 +685,19 @@ namespace Anima {
            };
            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), x.dim());
        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const DiffusionParams& diffusion_params) override {
+            GGML_ASSERT(diffusion_params.x != nullptr);
+            GGML_ASSERT(diffusion_params.timesteps != nullptr);
+            const auto* extra = diffusion_extra_as<AnimaDiffusionExtra>(diffusion_params);
+            return compute(n_threads,
+                           *diffusion_params.x,
+                           *diffusion_params.timesteps,
+                           tensor_or_empty(diffusion_params.context),
+                           tensor_or_empty(extra->t5_ids),
+                           tensor_or_empty(extra->t5_weights));
+        }
    };
 }  // namespace Anima

@@ -664,13 +664,13 @@ struct AutoEncoderKL : public VAE {
    AutoEncoderKLModel ae;

    AutoEncoderKL(ggml_backend_t backend,
-                  bool offload_params_to_cpu,
+                  ggml_backend_t params_backend,
                  const String2TensorStorage& tensor_storage_map,
                  const std::string prefix,
                  bool decode_only       = false,
                  bool use_video_decoder = false,
                  SDVersion version      = VERSION_SD1)
-        : decode_only(decode_only), VAE(version, backend, offload_params_to_cpu) {
+        : decode_only(decode_only), VAE(version, backend, params_backend) {
        if (sd_version_is_sd1(version) || sd_version_is_sd2(version)) {
            scale_factor = 0.18215f;
            shift_factor = 0.f;
@@ -680,7 +680,7 @@ struct AutoEncoderKL : public VAE {
        } else if (sd_version_is_sd3(version)) {
            scale_factor = 1.5305f;
            shift_factor = 0.0609f;
-        } else if (sd_version_is_flux(version) || sd_version_is_z_image(version)) {
+        } else if (sd_version_is_flux(version) || sd_version_is_z_image(version) || sd_version_is_longcat(version)) {
            scale_factor = 0.3611f;
            shift_factor = 0.1159f;
        } else if (sd_version_uses_flux2_vae(version)) {
@@ -469,13 +469,13 @@ struct CLIPTextModelRunner : public GGMLRunner {
    std::vector<float> attention_mask_vec;

    CLIPTextModelRunner(ggml_backend_t backend,
-                        bool offload_params_to_cpu,
+                        ggml_backend_t params_backend,
                        const String2TensorStorage& tensor_storage_map,
                        const std::string prefix,
                        CLIPVersion version = OPENAI_CLIP_VIT_L_14,
                        bool with_final_ln  = true,
                        bool force_clip_f32 = false)
-        : GGMLRunner(backend, offload_params_to_cpu) {
+        : GGMLRunner(backend, params_backend) {
        bool proj_in = false;
        for (const auto& [name, tensor_storage] : tensor_storage_map) {
            if (!starts_with(name, prefix)) {
@@ -103,6 +103,64 @@ namespace DiT {
        x         = ggml_ext_slice(ctx, x, 0, 0, W);               // [N, C, H, W]
        return x;
    }
+
+    inline ggml_tensor* patchify(ggml_context* ctx,
+                                 ggml_tensor* x,
+                                 int pt,
+                                 int ph,
+                                 int pw,
+                                 int64_t N = 1) {
+        // x: [N*C, T, H, W]
+        // return: [N, h*w, C*pt*ph*pw]
+        int64_t C     = x->ne[3] / N;
+        int64_t T     = x->ne[2];
+        int64_t H     = x->ne[1];
+        int64_t W     = x->ne[0];
+        int64_t t_len = T / pt;
+        int64_t h_len = H / ph;
+        int64_t w_len = W / pw;
+
+        GGML_ASSERT(C * N == x->ne[3]);
+        GGML_ASSERT(t_len * pt == T && h_len * ph == H && w_len * pw == W);
+
+        x = ggml_reshape_4d(ctx, x, pw * w_len, ph * h_len, pt, t_len * C * N);      // [N*C*t_len, pt, h_len*ph, w_len*pw]
+        x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 2, 1, 3));          // [N*C*t_len, h_len*ph, pt, w_len*pw]
+        x = ggml_reshape_4d(ctx, x, pw * w_len, pt, ph, h_len * t_len * C * N);      // [N*C*t_len*h_len, ph, pt, w_len*pw]
+        x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 2, 1, 3));          // [N*C*t_len*h_len, pt, ph, w_len*pw]
+        x = ggml_reshape_4d(ctx, x, pw, w_len, ph * pt, h_len * t_len * C * N);      // [N*C*t_len*h_len, pt*ph, w_len, pw]
+        x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 2, 1, 3));          // [N*C*t_len*h_len, w_len, pt*ph, pw]
+        x = ggml_reshape_4d(ctx, x, pw * ph * pt, w_len * h_len * t_len, C, N);      // [N, C, t_len*h_len*w_len, pt*ph*pw]
+        x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 2, 1, 3));          // [N, t_len*h_len*w_len, C, pt*ph*pw]
+        x = ggml_reshape_4d(ctx, x, pw * ph * pt * C, w_len * h_len * t_len, N, 1);  // [N, t_len*h_len*w_len, C*pt*ph*pw]
+        return x;
+    }
+
+    inline ggml_tensor* unpatchify(ggml_context* ctx,
+                                   ggml_tensor* x,
+                                   int64_t t_len,
+                                   int64_t h_len,
+                                   int64_t w_len,
+                                   int pt,
+                                   int ph,
+                                   int pw) {
+        // x: [N, t_len*h_len*w_len, pt*ph*pw*C]
+        // return: [N*C, t_len*pt, h_len*ph, w_len*pw]
+        int64_t N = x->ne[3];
+        int64_t C = x->ne[0] / pt / ph / pw;
+
+        GGML_ASSERT(C * pt * ph * pw == x->ne[0]);
+
+        x = ggml_reshape_4d(ctx, x, C, pw * ph * pt, w_len * h_len * t_len, N);  // [N, t_len*h_len*w_len, pt*ph*pw, C]
+        x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 1, 2, 0, 3));      // [N, C, t_len*h_len*w_len, pt*ph*pw]
+        x = ggml_reshape_4d(ctx, x, pw, ph * pt, w_len, h_len * t_len * C * N);  // [N*C*t_len*h_len, w_len, pt*ph, pw]
+        x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 2, 1, 3));      // [N*C*t_len*h_len, pt*ph, w_len, pw]
+        x = ggml_reshape_4d(ctx, x, pw * w_len, ph, pt, h_len * t_len * C * N);  // [N*C*t_len*h_len, pt, ph, w_len*pw]
+        x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 2, 1, 3));      // [N*C*t_len*h_len, ph, pt, w_len*pw]
+        x = ggml_reshape_4d(ctx, x, pw * w_len, pt, ph * h_len, t_len * C * N);  // [N*C*t_len, h_len*ph, pt, w_len*pw]
+        x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 2, 1, 3));      // [N*C*t_len, pt, h_len*ph, w_len*pw]
+        x = ggml_reshape_4d(ctx, x, pw * w_len, ph * h_len, pt * t_len, C * N);  // [N*C, t_len*pt, h_len*ph, w_len*pw]
+        return x;
+    }
 }  // namespace DiT

 #endif  // __COMMON_DIT_HPP__
@@ -1,6 +1,8 @@
 #ifndef __CONDITIONER_HPP__
 #define __CONDITIONER_HPP__

+#include <cmath>
+#include <limits>
 #include <optional>

 #include "clip.hpp"
@@ -14,6 +16,12 @@ struct SDCondition {
    sd::Tensor<float> c_concat;
    sd::Tensor<int32_t> c_t5_ids;
    sd::Tensor<float> c_t5_weights;
+    sd::Tensor<int32_t> c_input_ids;
+    sd::Tensor<int32_t> c_position_ids;
+    sd::Tensor<int32_t> c_token_types;
+    sd::Tensor<int32_t> c_vinput_mask;
+    std::vector<std::pair<int, sd::Tensor<float>>> c_image_embeds;
+    std::vector<sd::Tensor<float>> c_ref_images;

    std::vector<sd::Tensor<float>> extra_c_crossattns;

@@ -26,10 +34,24 @@ struct SDCondition {

    bool empty() const {
        if (!c_crossattn.empty() || !c_vector.empty() || !c_concat.empty() ||
-            !c_t5_ids.empty() || !c_t5_weights.empty()) {
+            !c_t5_ids.empty() || !c_t5_weights.empty() ||
+            !c_input_ids.empty() || !c_position_ids.empty() ||
+            !c_token_types.empty() || !c_vinput_mask.empty()) {
            return false;
        }

+        for (const auto& image_embed : c_image_embeds) {
+            if (!image_embed.second.empty()) {
+                return false;
+            }
+        }
+
+        for (const auto& tensor : c_ref_images) {
+            if (!tensor.empty()) {
+                return false;
+            }
+        }
+
        for (const auto& tensor : extra_c_crossattns) {
            if (!tensor.empty()) {
                return false;
@@ -46,6 +68,17 @@ static inline sd::Tensor<float> apply_token_weights(sd::Tensor<float> hidden_sta
        return hidden_states;
    }

+    bool all_one = true;
+    for (float weight : weights) {
+        if (weight != 1.0f) {
+            all_one = false;
+            break;
+        }
+    }
+    if (all_one) {
+        return hidden_states;
+    }
+
    if (hidden_states.dim() == 1) {
        hidden_states.unsqueeze_(1);
    }
@@ -57,7 +90,7 @@ static inline sd::Tensor<float> apply_token_weights(sd::Tensor<float> hidden_sta
    chunk_weights.reshape_({1, static_cast<int64_t>(weights.size())});
    hidden_states *= chunk_weights;
    float new_mean = hidden_states.mean();
-    if (new_mean != 0.0f) {
+    if (std::isfinite(original_mean) && std::isfinite(new_mean) && new_mean != 0.0f) {
        hidden_states *= (original_mean / new_mean);
    }

@@ -69,7 +102,6 @@ struct ConditionerParams {
    int clip_skip                                    = -1;
    int width                                        = -1;
    int height                                       = -1;
-    int adm_in_channels                              = -1;
    bool zero_out_masked                             = false;
    int num_input_imgs                               = 0;        // for photomaker
    const std::vector<sd::Tensor<float>>* ref_images = nullptr;  // for qwen image edit
@@ -81,11 +113,12 @@ struct Conditioner {
 public:
    virtual SDCondition get_learned_condition(int n_threads,
                                              const ConditionerParams& conditioner_params) = 0;
-    virtual void alloc_params_buffer()                                                     = 0;
+    virtual bool alloc_params_buffer()                                                     = 0;
    virtual void free_params_buffer()                                                      = 0;
    virtual void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors)           = 0;
    virtual size_t get_params_buffer_size()                                                = 0;
    virtual void set_max_graph_vram_bytes(size_t max_vram_bytes) {}
+    virtual void set_stream_layers_enabled(bool enabled) {}
    virtual void set_flash_attention_enabled(bool enabled) = 0;
    virtual void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) {}
    virtual std::tuple<SDCondition, std::vector<bool>> get_learned_condition_with_trigger(int n_threads,
@@ -114,7 +147,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
    std::map<std::string, std::pair<int, int>> embedding_pos_map;

    FrozenCLIPEmbedderWithCustomWords(ggml_backend_t backend,
-                                      bool offload_params_to_cpu,
+                                      ggml_backend_t params_backend,
                                      const String2TensorStorage& tensor_storage_map,
                                      const std::map<std::string, std::string>& orig_embedding_map,
                                      SDVersion version = VERSION_SD1,
@@ -128,12 +161,12 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
        }
        bool force_clip_f32 = !embedding_map.empty();
        if (sd_version_is_sd1(version)) {
-            text_model = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_storage_map, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14, true, force_clip_f32);
+            text_model = std::make_shared<CLIPTextModelRunner>(backend, params_backend, tensor_storage_map, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14, true, force_clip_f32);
        } else if (sd_version_is_sd2(version)) {
-            text_model = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_storage_map, "cond_stage_model.transformer.text_model", OPEN_CLIP_VIT_H_14, true, force_clip_f32);
+            text_model = std::make_shared<CLIPTextModelRunner>(backend, params_backend, tensor_storage_map, "cond_stage_model.transformer.text_model", OPEN_CLIP_VIT_H_14, true, force_clip_f32);
        } else if (sd_version_is_sdxl(version)) {
-            text_model  = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_storage_map, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14, false, force_clip_f32);
-            text_model2 = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_storage_map, "cond_stage_model.1.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, false, force_clip_f32);
+            text_model  = std::make_shared<CLIPTextModelRunner>(backend, params_backend, tensor_storage_map, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14, false, force_clip_f32);
+            text_model2 = std::make_shared<CLIPTextModelRunner>(backend, params_backend, tensor_storage_map, "cond_stage_model.1.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, false, force_clip_f32);
        }
    }

@@ -144,11 +177,16 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
        }
    }

-    void alloc_params_buffer() override {
-        text_model->alloc_params_buffer();
-        if (sd_version_is_sdxl(version)) {
-            text_model2->alloc_params_buffer();
+    bool alloc_params_buffer() override {
+        if (!text_model->alloc_params_buffer()) {
+            return false;
        }
+        if (sd_version_is_sdxl(version)) {
+            if (!text_model2->alloc_params_buffer()) {
+                return false;
+            }
+        }
+        return true;
    }

    void free_params_buffer() override {
@@ -173,6 +211,13 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
        }
    }

+    void set_stream_layers_enabled(bool enabled) override {
+        text_model->set_stream_layers_enabled(enabled);
+        if (sd_version_is_sdxl(version)) {
+            text_model2->set_stream_layers_enabled(enabled);
+        }
+    }
+
    void set_flash_attention_enabled(bool enabled) override {
        text_model->set_flash_attention_enabled(enabled);
        if (sd_version_is_sdxl(version)) {
@@ -469,7 +514,6 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
                                             int clip_skip,
                                             int width,
                                             int height,
-                                             int adm_in_channels  = -1,
                                             bool zero_out_masked = false) {
        int64_t t0 = ggml_time_ms();
        sd::Tensor<float> hidden_states;  // [n_token, hidden_size] or [n_token, hidden_size + hidden_size2]
@@ -555,7 +599,8 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {

        sd::Tensor<float> vec;
        if (sd_version_is_sdxl(version)) {
-            int out_dim = 256;
+            int out_dim         = 256;
+            int adm_in_channels = 2816;
            GGML_ASSERT(!pooled.empty());
            vec = sd::Tensor<float>({adm_in_channels});
            vec.fill_(0.0f);
@@ -614,7 +659,6 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
                                                 conditioner_params.clip_skip,
                                                 conditioner_params.width,
                                                 conditioner_params.height,
-                                                 conditioner_params.adm_in_channels,
                                                 conditioner_params.zero_out_masked);
        return std::make_tuple(cond, clsm);
    }
@@ -641,7 +685,6 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
                                            conditioner_params.clip_skip,
                                            conditioner_params.width,
                                            conditioner_params.height,
-                                            conditioner_params.adm_in_channels,
                                            conditioner_params.zero_out_masked);
    }
 };
@@ -650,9 +693,9 @@ struct FrozenCLIPVisionEmbedder : public GGMLRunner {
    CLIPVisionModelProjection vision_model;

    FrozenCLIPVisionEmbedder(ggml_backend_t backend,
-                             bool offload_params_to_cpu,
+                             ggml_backend_t params_backend,
                             const String2TensorStorage& tensor_storage_map = {})
-        : GGMLRunner(backend, offload_params_to_cpu) {
+        : GGMLRunner(backend, params_backend) {
        std::string prefix = "cond_stage_model.transformer";
        bool proj_in       = false;
        for (const auto& [name, tensor_storage] : tensor_storage_map) {
@@ -709,7 +752,7 @@ struct SD3CLIPEmbedder : public Conditioner {
    std::shared_ptr<T5Runner> t5;

    SD3CLIPEmbedder(ggml_backend_t backend,
-                    bool offload_params_to_cpu,
+                    ggml_backend_t params_backend,
                    const String2TensorStorage& tensor_storage_map = {})
        : clip_g_tokenizer(0) {
        bool use_clip_l = false;
@@ -729,13 +772,13 @@ struct SD3CLIPEmbedder : public Conditioner {
            return;
        }
        if (use_clip_l) {
-            clip_l = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_storage_map, "text_encoders.clip_l.transformer.text_model", OPENAI_CLIP_VIT_L_14, false);
+            clip_l = std::make_shared<CLIPTextModelRunner>(backend, params_backend, tensor_storage_map, "text_encoders.clip_l.transformer.text_model", OPENAI_CLIP_VIT_L_14, false);
        }
        if (use_clip_g) {
-            clip_g = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_storage_map, "text_encoders.clip_g.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, false);
+            clip_g = std::make_shared<CLIPTextModelRunner>(backend, params_backend, tensor_storage_map, "text_encoders.clip_g.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, false);
        }
        if (use_t5) {
-            t5 = std::make_shared<T5Runner>(backend, offload_params_to_cpu, tensor_storage_map, "text_encoders.t5xxl.transformer");
+            t5 = std::make_shared<T5Runner>(backend, params_backend, tensor_storage_map, "text_encoders.t5xxl.transformer");
        }
    }

@@ -751,16 +794,23 @@ struct SD3CLIPEmbedder : public Conditioner {
        }
    }

-    void alloc_params_buffer() override {
+    bool alloc_params_buffer() override {
        if (clip_l) {
-            clip_l->alloc_params_buffer();
+            if (!clip_l->alloc_params_buffer()) {
+                return false;
+            }
        }
        if (clip_g) {
-            clip_g->alloc_params_buffer();
+            if (!clip_g->alloc_params_buffer()) {
+                return false;
+            }
        }
        if (t5) {
-            t5->alloc_params_buffer();
+            if (!t5->alloc_params_buffer()) {
+                return false;
+            }
        }
+        return true;
    }

    void free_params_buffer() override {
@@ -801,6 +851,18 @@ struct SD3CLIPEmbedder : public Conditioner {
        }
    }

+    void set_stream_layers_enabled(bool enabled) override {
+        if (clip_l) {
+            clip_l->set_stream_layers_enabled(enabled);
+        }
+        if (clip_g) {
+            clip_g->set_stream_layers_enabled(enabled);
+        }
+        if (t5) {
+            t5->set_stream_layers_enabled(enabled);
+        }
+    }
+
    void set_flash_attention_enabled(bool enabled) override {
        if (clip_l) {
            clip_l->set_flash_attention_enabled(enabled);
@@ -1077,7 +1139,7 @@ struct FluxCLIPEmbedder : public Conditioner {
    size_t chunk_len = 256;

    FluxCLIPEmbedder(ggml_backend_t backend,
-                     bool offload_params_to_cpu,
+                     ggml_backend_t params_backend,
                     const String2TensorStorage& tensor_storage_map = {}) {
        bool use_clip_l = false;
        bool use_t5     = false;
@@ -1095,12 +1157,12 @@ struct FluxCLIPEmbedder : public Conditioner {
        }

        if (use_clip_l) {
-            clip_l = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_storage_map, "text_encoders.clip_l.transformer.text_model", OPENAI_CLIP_VIT_L_14, true);
+            clip_l = std::make_shared<CLIPTextModelRunner>(backend, params_backend, tensor_storage_map, "text_encoders.clip_l.transformer.text_model", OPENAI_CLIP_VIT_L_14, true);
        } else {
            LOG_WARN("clip_l text encoder not found! Prompt adherence might be degraded.");
        }
        if (use_t5) {
-            t5 = std::make_shared<T5Runner>(backend, offload_params_to_cpu, tensor_storage_map, "text_encoders.t5xxl.transformer");
+            t5 = std::make_shared<T5Runner>(backend, params_backend, tensor_storage_map, "text_encoders.t5xxl.transformer");
        } else {
            LOG_WARN("t5xxl text encoder not found! Prompt adherence might be degraded.");
        }
@@ -1115,13 +1177,18 @@ struct FluxCLIPEmbedder : public Conditioner {
        }
    }

-    void alloc_params_buffer() override {
+    bool alloc_params_buffer() override {
        if (clip_l) {
-            clip_l->alloc_params_buffer();
+            if (!clip_l->alloc_params_buffer()) {
+                return false;
+            }
        }
        if (t5) {
-            t5->alloc_params_buffer();
+            if (!t5->alloc_params_buffer()) {
+                return false;
+            }
        }
+        return true;
    }

    void free_params_buffer() override {
@@ -1153,6 +1220,15 @@ struct FluxCLIPEmbedder : public Conditioner {
        }
    }

+    void set_stream_layers_enabled(bool enabled) override {
+        if (clip_l) {
+            clip_l->set_stream_layers_enabled(enabled);
+        }
+        if (t5) {
+            t5->set_stream_layers_enabled(enabled);
+        }
+    }
+
    void set_flash_attention_enabled(bool enabled) override {
        if (clip_l) {
            clip_l->set_flash_attention_enabled(enabled);
@@ -1331,7 +1407,7 @@ struct T5CLIPEmbedder : public Conditioner {
    bool is_umt5     = false;

    T5CLIPEmbedder(ggml_backend_t backend,
-                   bool offload_params_to_cpu,
+                   ggml_backend_t params_backend,
                   const String2TensorStorage& tensor_storage_map = {},
                   bool use_mask                                  = false,
                   int mask_pad                                   = 0,
@@ -1348,7 +1424,7 @@ struct T5CLIPEmbedder : public Conditioner {
            LOG_WARN("IMPORTANT NOTICE: No text encoders provided, cannot process prompts!");
            return;
        } else {
-            t5 = std::make_shared<T5Runner>(backend, offload_params_to_cpu, tensor_storage_map, "text_encoders.t5xxl.transformer", is_umt5);
+            t5 = std::make_shared<T5Runner>(backend, params_backend, tensor_storage_map, "text_encoders.t5xxl.transformer", is_umt5);
        }
    }

@@ -1358,10 +1434,13 @@ struct T5CLIPEmbedder : public Conditioner {
        }
    }

-    void alloc_params_buffer() override {
+    bool alloc_params_buffer() override {
        if (t5) {
-            t5->alloc_params_buffer();
+            if (!t5->alloc_params_buffer()) {
+                return false;
+            }
        }
+        return true;
    }

    void free_params_buffer() override {
@@ -1384,6 +1463,12 @@ struct T5CLIPEmbedder : public Conditioner {
        }
    }

+    void set_stream_layers_enabled(bool enabled) override {
+        if (t5) {
+            t5->set_stream_layers_enabled(enabled);
+        }
+    }
+
    void set_flash_attention_enabled(bool enabled) override {
        if (t5) {
            t5->set_flash_attention_enabled(enabled);
@@ -1533,12 +1618,12 @@ struct AnimaConditioner : public Conditioner {
    std::shared_ptr<LLM::LLMRunner> llm;

    AnimaConditioner(ggml_backend_t backend,
-                     bool offload_params_to_cpu,
+                     ggml_backend_t params_backend,
                     const String2TensorStorage& tensor_storage_map = {}) {
        qwen_tokenizer = std::make_shared<Qwen2Tokenizer>();
        llm            = std::make_shared<LLM::LLMRunner>(LLM::LLMArch::QWEN3,
                                               backend,
-                                               offload_params_to_cpu,
+                                               params_backend,
                                               tensor_storage_map,
                                               "text_encoders.llm",
                                               false);
@@ -1548,8 +1633,11 @@ struct AnimaConditioner : public Conditioner {
        llm->get_param_tensors(tensors, "text_encoders.llm");
    }

-    void alloc_params_buffer() override {
-        llm->alloc_params_buffer();
+    bool alloc_params_buffer() override {
+        if (!llm->alloc_params_buffer()) {
+            return false;
+        }
+        return true;
    }

    void free_params_buffer() override {
@@ -1564,6 +1652,10 @@ struct AnimaConditioner : public Conditioner {
        llm->set_max_graph_vram_bytes(max_vram_bytes);
    }

+    void set_stream_layers_enabled(bool enabled) override {
+        llm->set_stream_layers_enabled(enabled);
+    }
+
    void set_flash_attention_enabled(bool enabled) override {
        llm->set_flash_attention_enabled(enabled);
    }
@@ -1605,10 +1697,11 @@ struct AnimaConditioner : public Conditioner {
        for (const auto& item : parsed_attention) {
            const std::string& curr_text = item.first;
            float curr_weight            = item.second;
-            std::vector<int> curr_tokens = t5_tokenizer.tokenize(curr_text, nullptr, true);
+            std::vector<int> curr_tokens = t5_tokenizer.encode(curr_text);
            t5_tokens.insert(t5_tokens.end(), curr_tokens.begin(), curr_tokens.end());
            t5_weights.insert(t5_weights.end(), curr_tokens.size(), curr_weight);
        }
+        t5_tokenizer.pad_tokens(t5_tokens, &t5_weights, nullptr);

        return {qwen_tokens, qwen_weights, t5_tokens, t5_weights};
    }
@@ -1651,7 +1744,7 @@ struct LLMEmbedder : public Conditioner {
    std::shared_ptr<LLM::LLMRunner> llm;

    LLMEmbedder(ggml_backend_t backend,
-                bool offload_params_to_cpu,
+                ggml_backend_t params_backend,
                const String2TensorStorage& tensor_storage_map = {},
                SDVersion version                              = VERSION_QWEN_IMAGE,
                const std::string prefix                       = "",
@@ -1662,17 +1755,27 @@ struct LLMEmbedder : public Conditioner {
            arch = LLM::LLMArch::MISTRAL_SMALL_3_2;
        } else if (sd_version_is_ernie_image(version)) {
            arch = LLM::LLMArch::MINISTRAL_3_3B;
+        } else if (sd_version_is_lens(version)) {
+            arch = LLM::LLMArch::GPT_OSS_20B;
+        } else if (sd_version_is_pid(version)) {
+            arch = LLM::LLMArch::GEMMA2_2B;
+        } else if (sd_version_is_ideogram4(version)) {
+            arch = LLM::LLMArch::QWEN3_VL;
        } else if (sd_version_is_z_image(version) || version == VERSION_OVIS_IMAGE || version == VERSION_FLUX2_KLEIN) {
            arch = LLM::LLMArch::QWEN3;
        }
        if (arch == LLM::LLMArch::MISTRAL_SMALL_3_2 || arch == LLM::LLMArch::MINISTRAL_3_3B) {
            tokenizer = std::make_shared<MistralTokenizer>();
+        } else if (arch == LLM::LLMArch::GPT_OSS_20B) {
+            tokenizer = std::make_shared<GPTOSSTokenizer>();
+        } else if (arch == LLM::LLMArch::GEMMA2_2B) {
+            tokenizer = std::make_shared<Gemma2Tokenizer>();
        } else {
            tokenizer = std::make_shared<Qwen2Tokenizer>();
        }
        llm = std::make_shared<LLM::LLMRunner>(arch,
                                               backend,
-                                               offload_params_to_cpu,
+                                               params_backend,
                                               tensor_storage_map,
                                               "text_encoders.llm",
                                               enable_vision);
@@ -1682,8 +1785,11 @@ struct LLMEmbedder : public Conditioner {
        llm->get_param_tensors(tensors, "text_encoders.llm");
    }

-    void alloc_params_buffer() override {
-        llm->alloc_params_buffer();
+    bool alloc_params_buffer() override {
+        if (!llm->alloc_params_buffer()) {
+            return false;
+        }
+        return true;
    }

    void free_params_buffer() override {
@@ -1700,6 +1806,10 @@ struct LLMEmbedder : public Conditioner {
        llm->set_max_graph_vram_bytes(max_vram_bytes);
    }

+    void set_stream_layers_enabled(bool enabled) override {
+        llm->set_stream_layers_enabled(enabled);
+    }
+
    void set_flash_attention_enabled(bool enabled) override {
        llm->set_flash_attention_enabled(enabled);
    }
@@ -1713,7 +1823,8 @@ struct LLMEmbedder : public Conditioner {
    std::tuple<std::vector<int>, std::vector<float>, std::vector<float>> tokenize(std::string text,
                                                                                  const std::pair<int, int>& attn_range,
                                                                                  size_t min_length = 0,
-                                                                                  size_t max_length = 100000000) {
+                                                                                  size_t max_length = 100000000,
+                                                                                  bool spell_quotes = false) {
        std::vector<std::pair<std::string, float>> parsed_attention;
        if (attn_range.first >= 0 && attn_range.second > 0) {
            if (attn_range.first > 0) {
@@ -1721,6 +1832,9 @@ struct LLMEmbedder : public Conditioner {
            }
            if (attn_range.second - attn_range.first > 0) {
                auto new_parsed_attention = parse_prompt_attention(text.substr(attn_range.first, attn_range.second - attn_range.first));
+                if (spell_quotes) {
+                    new_parsed_attention = split_quotation_attention(new_parsed_attention);
+                }
                parsed_attention.insert(parsed_attention.end(),
                                        new_parsed_attention.begin(),
                                        new_parsed_attention.end());
@@ -1770,8 +1884,10 @@ struct LLMEmbedder : public Conditioner {
                                    int hidden_states_min_length,
                                    const std::vector<std::pair<int, sd::Tensor<float>>>& image_embeds,
                                    const std::set<int>& out_layers,
-                                    int prompt_template_encode_start_idx) {
-        auto tokens_weights_mask = tokenize(prompt, prompt_attn_range, min_length);
+                                    int prompt_template_encode_start_idx,
+                                    bool spell_quotes = false,
+                                    int max_length    = 100000000) {
+        auto tokens_weights_mask = tokenize(prompt, prompt_attn_range, min_length, max_length, spell_quotes);
        auto& tokens             = std::get<0>(tokens_weights_mask);
        auto& weights            = std::get<1>(tokens_weights_mask);
        auto& mask               = std::get<2>(tokens_weights_mask);
@@ -1779,12 +1895,16 @@ struct LLMEmbedder : public Conditioner {
        sd::Tensor<int32_t> input_ids({static_cast<int64_t>(tokens.size())}, tokens);
        sd::Tensor<float> attention_mask;
        if (!mask.empty()) {
-            attention_mask = sd::Tensor<float>({static_cast<int64_t>(mask.size()), static_cast<int64_t>(mask.size())});
+            attention_mask                     = sd::Tensor<float>({static_cast<int64_t>(mask.size()), static_cast<int64_t>(mask.size())});
+            const float masked_attention_value = -std::numeric_limits<float>::max() / 4.0f;
            for (size_t i1 = 0; i1 < mask.size(); ++i1) {
                for (size_t i0 = 0; i0 < mask.size(); ++i0) {
                    float value = 0.0f;
-                    if (mask[i0] == 0.0f || i0 > i1) {
-                        value = -INFINITY;
+                    if (mask[i0] == 0.0f) {
+                        value += masked_attention_value;
+                    }
+                    if (i0 > i1) {
+                        value += masked_attention_value;
                    }
                    attention_mask[static_cast<int64_t>(i0 + mask.size() * i1)] = value;
                }
@@ -1831,7 +1951,9 @@ struct LLMEmbedder : public Conditioner {
        std::vector<std::pair<int, sd::Tensor<float>>> image_embeds;
        int prompt_template_encode_start_idx = 34;
        int min_length                       = 0;  // pad tokens
+        int max_length                       = 100000000;
        int hidden_states_min_length         = 0;  // zero pad hidden_states
+        bool spell_quotes                    = false;
        std::set<int> out_layers;

        int64_t t0 = ggml_time_ms();
@@ -1904,6 +2026,71 @@ struct LLMEmbedder : public Conditioner {

                prompt += "<|im_end|>\n<|im_start|>assistant\n";
            }
+        } else if (sd_version_is_longcat(version)) {
+            spell_quotes = true;
+
+            if (llm->enable_vision && conditioner_params.ref_images != nullptr && !conditioner_params.ref_images->empty()) {
+                LOG_INFO("LongCatEditPipeline");
+                prompt_template_encode_start_idx = 67;
+                min_length                       = 512 + prompt_template_encode_start_idx;
+                int image_embed_idx              = 36 + 6;
+
+                int min_pixels          = 384 * 384;
+                int max_pixels          = 560 * 560;
+                std::string placeholder = "<|image_pad|>";
+                std::string img_prompt;
+
+                for (int i = 0; i < conditioner_params.ref_images->size(); i++) {
+                    const auto& image = (*conditioner_params.ref_images)[i];
+                    double factor     = llm->params.vision.patch_size * llm->params.vision.spatial_merge_size;
+                    int height        = static_cast<int>(image.shape()[1]);
+                    int width         = static_cast<int>(image.shape()[0]);
+                    int h_bar         = static_cast<int>(std::round(height / factor) * factor);
+                    int w_bar         = static_cast<int>(std::round(width / factor) * factor);
+
+                    if (static_cast<double>(h_bar) * w_bar > max_pixels) {
+                        double beta = std::sqrt((height * width) / static_cast<double>(max_pixels));
+                        h_bar       = std::max(static_cast<int>(factor),
+                                               static_cast<int>(std::floor(height / beta / factor)) * static_cast<int>(factor));
+                        w_bar       = std::max(static_cast<int>(factor),
+                                               static_cast<int>(std::floor(width / beta / factor)) * static_cast<int>(factor));
+                    } else if (static_cast<double>(h_bar) * w_bar < min_pixels) {
+                        double beta = std::sqrt(static_cast<double>(min_pixels) / (height * width));
+                        h_bar       = static_cast<int>(std::ceil(height * beta / factor)) * static_cast<int>(factor);
+                        w_bar       = static_cast<int>(std::ceil(width * beta / factor)) * static_cast<int>(factor);
+                    }
+
+                    LOG_DEBUG("resize conditioner ref image %d from %dx%d to %dx%d", i, height, width, h_bar, w_bar);
+
+                    auto resized_image = clip_preprocess(image, w_bar, h_bar);
+                    auto image_embed   = llm->encode_image(n_threads, resized_image);
+                    GGML_ASSERT(!image_embed.empty());
+                    image_embeds.emplace_back(image_embed_idx, image_embed);
+                    image_embed_idx += 1 + static_cast<int>(image_embed.shape()[1]) + 6;
+
+                    img_prompt += "<|vision_start|>";
+                    int64_t num_image_tokens = image_embed.shape()[1];
+                    img_prompt.reserve(num_image_tokens * placeholder.size());
+                    for (int j = 0; j < num_image_tokens; j++) {
+                        img_prompt += placeholder;
+                    }
+                    img_prompt += "<|vision_end|>";
+                }
+
+                prompt = "<|im_start|>system\nAs an image editing expert, first analyze the content and attributes of the input image(s). Then, based on the user's editing instructions, clearly and precisely determine how to modify the given image(s), ensuring that only the specified parts are altered and all other aspects remain consistent with the original(s).<|im_end|>\n<|im_start|>user\n";
+                prompt += img_prompt;
+            } else {
+                prompt_template_encode_start_idx = 36;
+                min_length                       = 512 + prompt_template_encode_start_idx;
+
+                prompt = "<|im_start|>system\nAs an image captioning expert, generate a descriptive text prompt based on an image content, suitable for input to a text-to-image model.<|im_end|>\n<|im_start|>user\n";
+            }
+
+            prompt_attn_range.first = static_cast<int>(prompt.size());
+            prompt += conditioner_params.text;
+            prompt_attn_range.second = static_cast<int>(prompt.size());
+
+            prompt += "<|im_end|>\n<|im_start|>assistant\n";
        } else if (version == VERSION_FLUX2) {
            prompt_template_encode_start_idx = 0;
            hidden_states_min_length         = 512;
@@ -1916,6 +2103,14 @@ struct LLMEmbedder : public Conditioner {
            prompt_attn_range.second = static_cast<int>(prompt.size());

            prompt += "[/INST]";
+        } else if (sd_version_is_ideogram4(version)) {
+            prompt_template_encode_start_idx = 0;
+            out_layers                       = {1, 4, 7, 10, 13, 16, 19, 22, 25, 28, 31, 34, 36};
+
+            prompt = "<|im_start|>user\n";
+            prompt += conditioner_params.text;
+            prompt += "<|im_end|>\n<|im_start|>assistant\n";
+            prompt_attn_range = {0, 0};
        } else if (sd_version_is_ernie_image(version)) {
            prompt_template_encode_start_idx = 0;
            out_layers                       = {25};  // -2
@@ -1923,6 +2118,30 @@ struct LLMEmbedder : public Conditioner {
            prompt_attn_range.first = 0;
            prompt += conditioner_params.text;
            prompt_attn_range.second = static_cast<int>(prompt.size());
+        } else if (sd_version_is_lens(version)) {
+            prompt_template_encode_start_idx = 97;
+            min_length                       = 0;
+            max_length                       = 512;
+            out_layers                       = {6, 12, 18, 24};
+
+            prompt =
+                "<|start|>system<|message|>You are ChatGPT, a large language model trained by OpenAI.\n"
+                "Knowledge cutoff: 2024-06\n"
+                "Current date: 2026-05-26\n"  // fix for current date
+                "\n"
+                "Reasoning: medium\n"
+                "\n"
+                "# Valid channels: analysis, commentary, final. Channel must be included for every message.<|end|><|start|>developer<|message|># Instructions\n"
+                "\n"
+                "Describe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background.\n"
+                "\n"
+                "<|end|><|start|>user<|message|>";
+
+            prompt_attn_range.first = static_cast<int>(prompt.size());
+            prompt += conditioner_params.text;
+            prompt_attn_range.second = static_cast<int>(prompt.size());
+
+            prompt += "<|end|><|start|>assistant<|channel|>analysis<|message|>Need to generate one image according to the description.<|end|><|start|>assistant<|channel|>final<|message|>";
        } else if (sd_version_is_z_image(version)) {
            prompt_template_encode_start_idx = 0;
            out_layers                       = {35};  // -2
@@ -1967,6 +2186,53 @@ struct LLMEmbedder : public Conditioner {
            prompt_attn_range.second = static_cast<int>(prompt.size());

            prompt += "<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n";
+        } else if (sd_version_is_pid(version)) {
+            constexpr int pixeldit_max_length = 300;
+            const std::string chi_prompt =
+                "Given a user prompt, generate an \"Enhanced prompt\" that provides detailed visual descriptions suitable for image generation. Evaluate the level of detail in the user prompt:\n"
+                "- If the prompt is simple, focus on adding specifics about colors, shapes, sizes, textures, and spatial relationships to create vivid and concrete scenes.\n"
+                "- If the prompt is already detailed, refine and enhance the existing details slightly without overcomplicating.\n"
+                "Here are examples of how to transform or refine prompts:\n"
+                "- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by pots of blooming red flowers.\n"
+                "- User Prompt: A busy city street -> Enhanced: A bustling city street scene at dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing, and a double-decker bus passing by towering glass skyscrapers.\n"
+                "Please generate only the enhanced description for the prompt below and avoid including any additional commentary or evaluations:\n"
+                "User Prompt: ";
+            auto chi_tokens       = std::get<0>(tokenize(chi_prompt, {0, 0}));
+            size_t num_chi_tokens = chi_tokens.size();
+            max_length            = (int)num_chi_tokens + pixeldit_max_length - 2;
+            min_length            = max_length;
+
+            prompt_attn_range.first = static_cast<int>(prompt.size());
+            prompt += " " + conditioner_params.text;
+            prompt_attn_range.second = static_cast<int>(prompt.size());
+
+            auto hidden_states = encode_prompt(n_threads,
+                                               prompt,
+                                               prompt_attn_range,
+                                               min_length,
+                                               0,
+                                               image_embeds,
+                                               out_layers,
+                                               0,
+                                               false,
+                                               max_length);
+            GGML_ASSERT(!hidden_states.empty());
+
+            if (hidden_states.shape()[1] > pixeldit_max_length) {
+                auto bos      = sd::ops::slice(hidden_states, 1, 0, 1);
+                auto tail     = sd::ops::slice(hidden_states,
+                                               1,
+                                               hidden_states.shape()[1] - (pixeldit_max_length - 1),
+                                               hidden_states.shape()[1]);
+                hidden_states = sd::ops::concat(bos, tail, 1);
+            }
+
+            int64_t t1 = ggml_time_ms();
+            LOG_DEBUG("computing condition graph completed, taking %" PRId64 " ms", t1 - t0);
+
+            SDCondition result;
+            result.c_crossattn = std::move(hidden_states);
+            return result;
        } else {
            GGML_ABORT("unknown version %d", version);
        }
@@ -1978,7 +2244,9 @@ struct LLMEmbedder : public Conditioner {
                                           hidden_states_min_length,
                                           image_embeds,
                                           out_layers,
-                                           prompt_template_encode_start_idx);
+                                           prompt_template_encode_start_idx,
+                                           spell_quotes,
+                                           max_length);
        std::vector<sd::Tensor<float>> extra_hidden_states_vec;
        for (int i = 0; i < extra_prompts.size(); i++) {
            auto extra_hidden_states = encode_prompt(n_threads,
@@ -1988,7 +2256,9 @@ struct LLMEmbedder : public Conditioner {
                                                     hidden_states_min_length,
                                                     image_embeds,
                                                     out_layers,
-                                                     prompt_template_encode_start_idx);
+                                                     prompt_template_encode_start_idx,
+                                                     spell_quotes,
+                                                     max_length);
            extra_hidden_states_vec.push_back(std::move(extra_hidden_states));
        }

@@ -2001,4 +2271,287 @@ struct LLMEmbedder : public Conditioner {
    }
 };

+struct LTXAVTextProjection : public GGMLBlock {
+    static constexpr int64_t kHiddenSize = 3840;
+    static constexpr int64_t kNumStates  = 49;
+    bool dual_projection                 = false;
+
+    LTXAVTextProjection(bool dual_projection = false)
+        : dual_projection(dual_projection) {
+        if (dual_projection) {
+            blocks["video_aggregate_embed"] = std::make_shared<Linear>(kHiddenSize * kNumStates, 4096, true);
+            blocks["audio_aggregate_embed"] = std::make_shared<Linear>(kHiddenSize * kNumStates, 2048, true);
+        } else {
+            blocks["projection"] = std::make_shared<Linear>(kHiddenSize * kNumStates, kHiddenSize, false);
+        }
+    }
+
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+        if (!dual_projection) {
+            auto projection = std::dynamic_pointer_cast<Linear>(blocks["projection"]);
+            return projection->forward(ctx, x);
+        }
+
+        auto video_projection = std::dynamic_pointer_cast<Linear>(blocks["video_aggregate_embed"]);
+        auto audio_projection = std::dynamic_pointer_cast<Linear>(blocks["audio_aggregate_embed"]);
+        auto video_in         = ggml_ext_scale(ctx->ggml_ctx, x, std::sqrt(4096.f / static_cast<float>(kHiddenSize)));
+        auto audio_in         = ggml_ext_scale(ctx->ggml_ctx, x, std::sqrt(2048.f / static_cast<float>(kHiddenSize)));
+        auto video            = video_projection->forward(ctx, video_in);
+        auto audio            = audio_projection->forward(ctx, audio_in);
+        return ggml_concat(ctx->ggml_ctx, video, audio, 0);
+    }
+};
+
+struct LTXAVTextProjectionRunner : public GGMLRunner {
+    LTXAVTextProjection model;
+
+    LTXAVTextProjectionRunner(ggml_backend_t backend,
+                              ggml_backend_t params_backend,
+                              const String2TensorStorage& tensor_storage_map = {},
+                              const std::string& prefix                      = "")
+        : GGMLRunner(backend, params_backend),
+          model(tensor_storage_map.find(prefix + ".video_aggregate_embed.weight") != tensor_storage_map.end()) {
+        model.init(params_ctx, tensor_storage_map, prefix);
+    }
+
+    std::string get_desc() override {
+        return "ltxav_text_projection";
+    }
+
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) {
+        model.get_param_tensors(tensors, prefix);
+    }
+
+    ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor) {
+        ggml_cgraph* gf = ggml_new_graph(compute_ctx);
+        auto x          = make_input(x_tensor);
+        auto runner_ctx = get_context();
+        auto out        = model.forward(&runner_ctx, x);
+        ggml_build_forward_expand(gf, out);
+        return gf;
+    }
+
+    sd::Tensor<float> compute(int n_threads, const sd::Tensor<float>& x) {
+        auto get_graph = [&]() -> ggml_cgraph* {
+            return build_graph(x);
+        };
+        return take_or_empty(GGMLRunner::compute<float>(get_graph, n_threads, true));
+    }
+};
+
+struct LTXAVEmbedder : public Conditioner {
+    static constexpr int64_t kHiddenSize = 3840;
+    static constexpr int64_t kNumStates  = 49;
+    static constexpr int64_t kMinLength  = 1024;
+
+    std::shared_ptr<GemmaTokenizer> tokenizer;
+    std::shared_ptr<LLM::LLMRunner> llm;
+    std::shared_ptr<LTXAVTextProjectionRunner> projector;
+    bool dual_projection = false;
+
+    LTXAVEmbedder(ggml_backend_t backend,
+                  ggml_backend_t params_backend,
+                  const String2TensorStorage& tensor_storage_map = {},
+                  const std::string& llm_prefix                  = "text_encoders.llm",
+                  const std::string& projector_prefix            = "text_embedding_projection") {
+        tokenizer       = std::make_shared<GemmaTokenizer>();
+        llm             = std::make_shared<LLM::LLMRunner>(LLM::LLMArch::GEMMA3_12B,
+                                               backend,
+                                               params_backend,
+                                               tensor_storage_map,
+                                               llm_prefix,
+                                               false);
+        dual_projection = tensor_storage_map.find(projector_prefix + ".video_aggregate_embed.weight") != tensor_storage_map.end();
+        projector       = std::make_shared<LTXAVTextProjectionRunner>(backend,
+                                                                params_backend,
+                                                                tensor_storage_map,
+                                                                projector_prefix);
+    }
+
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+        llm->get_param_tensors(tensors, "text_encoders.llm");
+        projector->get_param_tensors(tensors, "text_embedding_projection");
+    }
+
+    bool alloc_params_buffer() override {
+        if (!llm->alloc_params_buffer()) {
+            return false;
+        }
+        if (!projector->alloc_params_buffer()) {
+            return false;
+        }
+        return true;
+    }
+
+    void free_params_buffer() override {
+        llm->free_params_buffer();
+        projector->free_params_buffer();
+    }
+
+    size_t get_params_buffer_size() override {
+        return llm->get_params_buffer_size() + projector->get_params_buffer_size();
+    }
+
+    void set_flash_attention_enabled(bool enabled) override {
+        llm->set_flash_attention_enabled(enabled);
+        projector->set_flash_attention_enabled(enabled);
+    }
+
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        llm->set_max_graph_vram_bytes(max_vram_bytes);
+        projector->set_max_graph_vram_bytes(max_vram_bytes);
+    }
+
+    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
+        llm->set_weight_adapter(adapter);
+        projector->set_weight_adapter(adapter);
+    }
+
+    std::tuple<std::vector<int>, std::vector<float>, std::vector<float>> tokenize(std::string text,
+                                                                                  const std::pair<int, int>& attn_range) {
+        std::vector<std::pair<std::string, float>> parsed_attention;
+        if (attn_range.first >= 0 && attn_range.second > 0) {
+            if (attn_range.first > 0) {
+                parsed_attention.emplace_back(text.substr(0, attn_range.first), 1.f);
+            }
+            if (attn_range.second - attn_range.first > 0) {
+                auto new_parsed_attention = parse_prompt_attention(text.substr(attn_range.first, attn_range.second - attn_range.first));
+                parsed_attention.insert(parsed_attention.end(), new_parsed_attention.begin(), new_parsed_attention.end());
+            }
+            if (static_cast<size_t>(attn_range.second) < text.size()) {
+                parsed_attention.emplace_back(text.substr(attn_range.second), 1.f);
+            }
+        } else {
+            parsed_attention.emplace_back(text, 1.f);
+        }
+
+        std::vector<int> tokens;
+        std::vector<float> weights;
+        for (const auto& item : parsed_attention) {
+            auto curr_tokens = tokenizer->encode(item.first, nullptr);
+            tokens.insert(tokens.end(), curr_tokens.begin(), curr_tokens.end());
+            weights.insert(weights.end(), curr_tokens.size(), item.second);
+        }
+
+        std::vector<float> mask;
+        tokenizer->pad_tokens(tokens, &weights, &mask, kMinLength);
+        return {tokens, weights, mask};
+    }
+
+    sd::Tensor<float> encode_prompt(int n_threads,
+                                    const std::string& prompt,
+                                    const std::pair<int, int>& prompt_attn_range) {
+        auto tokens_weights_mask = tokenize(prompt, prompt_attn_range);
+        auto& tokens             = std::get<0>(tokens_weights_mask);
+        auto& weights            = std::get<1>(tokens_weights_mask);
+        auto& mask               = std::get<2>(tokens_weights_mask);
+
+        sd::Tensor<int32_t> input_ids({static_cast<int64_t>(tokens.size())}, std::vector<int32_t>(tokens.begin(), tokens.end()));
+        sd::Tensor<float> attention_mask;
+        if (!mask.empty()) {
+            const float mask_min = std::numeric_limits<float>::lowest() / 4.0f;
+            attention_mask       = sd::Tensor<float>({static_cast<int64_t>(mask.size()), static_cast<int64_t>(mask.size())});
+            for (size_t i1 = 0; i1 < mask.size(); ++i1) {
+                for (size_t i0 = 0; i0 < mask.size(); ++i0) {
+                    float value = 0.0f;
+                    if (mask[i0] == 0.0f) {
+                        value += mask_min;
+                    }
+                    if (i0 > i1) {
+                        value += mask_min;
+                    }
+                    attention_mask[static_cast<int64_t>(i0 + mask.size() * i1)] = value;
+                }
+            }
+        }
+
+        auto hidden_states = llm->compute(n_threads,
+                                          input_ids,
+                                          attention_mask,
+                                          {},
+                                          {},
+                                          true);
+        GGML_ASSERT(!hidden_states.empty());
+        hidden_states = apply_token_weights(std::move(hidden_states), weights);
+
+        int64_t valid_tokens = 0;
+        for (float value : mask) {
+            valid_tokens += static_cast<int64_t>(value > 0.0f);
+        }
+        GGML_ASSERT(valid_tokens > 0);
+
+        hidden_states = sd::ops::slice(hidden_states,
+                                       1,
+                                       hidden_states.shape()[1] - valid_tokens,
+                                       hidden_states.shape()[1]);
+        hidden_states.reshape_({kHiddenSize, kNumStates, valid_tokens});
+        hidden_states = hidden_states.permute({1, 0, 2});
+
+        if (dual_projection) {
+            for (int64_t state_idx = 0; state_idx < kNumStates; ++state_idx) {
+                for (int64_t token_idx = 0; token_idx < valid_tokens; ++token_idx) {
+                    double sq_sum = 0.0;
+                    for (int64_t hidden_idx = 0; hidden_idx < kHiddenSize; ++hidden_idx) {
+                        float value = hidden_states.index(state_idx, hidden_idx, token_idx);
+                        sq_sum += static_cast<double>(value) * static_cast<double>(value);
+                    }
+
+                    float inv_rms = 1.0f / std::sqrt(static_cast<float>(sq_sum / static_cast<double>(kHiddenSize)) + 1e-6f);
+                    for (int64_t hidden_idx = 0; hidden_idx < kHiddenSize; ++hidden_idx) {
+                        hidden_states.index(state_idx, hidden_idx, token_idx) *= inv_rms;
+                    }
+                }
+            }
+        } else {
+            for (int64_t state_idx = 0; state_idx < kNumStates; ++state_idx) {
+                double sum      = 0.0;
+                float min_value = std::numeric_limits<float>::infinity();
+                float max_value = -std::numeric_limits<float>::infinity();
+                for (int64_t token_idx = 0; token_idx < valid_tokens; ++token_idx) {
+                    for (int64_t hidden_idx = 0; hidden_idx < kHiddenSize; ++hidden_idx) {
+                        float value = hidden_states.index(state_idx, hidden_idx, token_idx);
+                        sum += value;
+                        min_value = std::min(min_value, value);
+                        max_value = std::max(max_value, value);
+                    }
+                }
+
+                float mean_value  = static_cast<float>(sum / static_cast<double>(kHiddenSize * valid_tokens));
+                float denom       = max_value - min_value + 1e-6f;
+                float scale_value = 8.0f / denom;
+                for (int64_t token_idx = 0; token_idx < valid_tokens; ++token_idx) {
+                    for (int64_t hidden_idx = 0; hidden_idx < kHiddenSize; ++hidden_idx) {
+                        float value                                           = hidden_states.index(state_idx, hidden_idx, token_idx);
+                        hidden_states.index(state_idx, hidden_idx, token_idx) = (value - mean_value) * scale_value;
+                    }
+                }
+            }
+        }
+
+        hidden_states.reshape_({kNumStates * kHiddenSize, valid_tokens});
+        return projector->compute(n_threads, hidden_states);
+    }
+
+    SDCondition get_learned_condition(int n_threads,
+                                      const ConditionerParams& conditioner_params) override {
+        int64_t t0 = ggml_time_ms();
+
+        std::string prompt;
+        std::pair<int, int> prompt_attn_range;
+        prompt_attn_range.first = static_cast<int>(prompt.size());
+        prompt += conditioner_params.text;
+        prompt_attn_range.second = static_cast<int>(prompt.size());
+
+        auto hidden_states = encode_prompt(n_threads, prompt, prompt_attn_range);
+        GGML_ASSERT(!hidden_states.empty());
+
+        int64_t t1 = ggml_time_ms();
+        LOG_DEBUG("computing LTXAV condition graph completed, taking %" PRId64 " ms", t1 - t0);
+
+        SDCondition result;
+        result.c_crossattn = std::move(hidden_states);
+        return result;
+    }
+};
+
 #endif
@@ -319,10 +319,10 @@ struct ControlNet : public GGMLRunner {
    bool guided_hint_cached = false;

    ControlNet(ggml_backend_t backend,
-               bool offload_params_to_cpu,
+               ggml_backend_t params_backend,
               const String2TensorStorage& tensor_storage_map = {},
               SDVersion version                              = VERSION_SD1)
-        : GGMLRunner(backend, offload_params_to_cpu), control_net(version) {
+        : GGMLRunner(backend, params_backend), control_net(version) {
        control_net.init(params_ctx, tensor_storage_map, "");
    }

@@ -457,7 +457,11 @@ struct ControlNet : public GGMLRunner {

    bool load_from_file(const std::string& file_path, int n_threads) {
        LOG_INFO("loading control net from '%s'", file_path.c_str());
-        alloc_params_buffer();
+        if (!alloc_params_buffer()) {
+            LOG_ERROR("control net model buffer allocation failed");
+            return false;
+        }
+
        std::map<std::string, ggml_tensor*> tensors;
        control_net.get_param_tensors(tensors);
        std::set<std::string> ignore_tensors;
@@ -8,7 +8,7 @@
 #include "model_io/safetensors_io.h"
 #include "util.h"

-#include "ggml-cpu.h"
+#include "ggml_extend_backend.h"

 static ggml_type get_export_tensor_type(ModelLoader& model_loader,
                                        const TensorStorage& tensor_storage,
@@ -103,7 +103,7 @@ bool convert(const char* input_path,
    bool output_is_safetensors = ends_with(output_path, ".safetensors");
    TensorTypeRules type_rules = parse_tensor_type_rules(tensor_type_rules);

-    auto backend    = ggml_backend_cpu_init();
+    auto backend    = sd_backend_cpu_init();
    size_t mem_size = 1 * 1024 * 1024;  // for padding
    mem_size += model_loader.get_tensor_storage_map().size() * ggml_tensor_overhead();
    mem_size += model_loader.get_params_mem_size(backend, type);
@@ -1,11 +1,16 @@
 #ifndef __DENOISER_HPP__
 #define __DENOISER_HPP__

+#include <algorithm>
+#include <cctype>
 #include <cmath>
+#include <functional>
+#include <string>
 #include <utility>

 #include "ggml_extend.hpp"
 #include "gits_noise.inl"
+#include "guidance.h"
 #include "tensor.hpp"

 /*================================================= CompVisDenoiser ==================================================*/
@@ -479,6 +484,81 @@ struct KLOptimalScheduler : SigmaScheduler {
    }
 };

+struct LTX2Scheduler : SigmaScheduler {
+    int token_count  = 4096;
+    float max_shift  = 2.05f;
+    float base_shift = 0.95f;
+    bool stretch     = true;
+    float terminal   = 0.1f;
+
+    explicit LTX2Scheduler(int token_count, const char* extra_sample_args = nullptr)
+        : token_count(token_count > 0 ? token_count : 4096) {
+        parse_extra_sample_args(extra_sample_args);
+    }
+
+    void parse_extra_sample_args(const char* extra_sample_args) {
+        for (const auto& [key, value] : parse_key_value_args(extra_sample_args, "ltx2 scheduler arg")) {
+            if (key == "max_shift") {
+                if (!parse_strict_float(value, max_shift)) {
+                    LOG_WARN("ignoring invalid ltx2 scheduler arg '%s=%s'", key.c_str(), value.c_str());
+                }
+            } else if (key == "base_shift") {
+                if (!parse_strict_float(value, base_shift)) {
+                    LOG_WARN("ignoring invalid ltx2 scheduler arg '%s=%s'", key.c_str(), value.c_str());
+                }
+            } else if (key == "terminal") {
+                if (!parse_strict_float(value, terminal)) {
+                    LOG_WARN("ignoring invalid ltx2 scheduler arg '%s=%s'", key.c_str(), value.c_str());
+                }
+            } else if (key == "stretch") {
+                if (!parse_strict_bool(value, stretch)) {
+                    LOG_WARN("ignoring invalid ltx2 scheduler arg '%s=%s'", key.c_str(), value.c_str());
+                }
+            }
+        }
+    }
+
+    std::vector<float> get_sigmas(uint32_t n, float /*sigma_min*/, float /*sigma_max*/, t_to_sigma_t /*t_to_sigma*/) override {
+        std::vector<float> sigmas;
+        if (n == 0) {
+            sigmas.push_back(0.0f);
+            return sigmas;
+        }
+
+        constexpr float base_shift_anchor = 1024.0f;
+        constexpr float max_shift_anchor  = 4096.0f;
+        float m                           = (max_shift - base_shift) / (max_shift_anchor - base_shift_anchor);
+        float b                           = base_shift - m * base_shift_anchor;
+        float sigma_shift                 = static_cast<float>(token_count) * m + b;
+        float exp_shift                   = std::exp(sigma_shift);
+        float target_terminal             = std::clamp(terminal, 0.0f, 0.99f);
+
+        LOG_DEBUG("LTX2 scheduler: tokens=%d, shift=%.4f, stretch=%d, terminal=%.4f", token_count, sigma_shift, stretch ? 1 : 0, target_terminal);
+
+        sigmas.reserve(n + 1);
+        for (uint32_t i = 0; i <= n; ++i) {
+            float sigma = 1.0f - static_cast<float>(i) / static_cast<float>(n);
+            if (sigma != 0.0f) {
+                sigma = exp_shift / (exp_shift + (1.0f / sigma - 1.0f));
+            }
+            sigmas.push_back(sigma);
+        }
+
+        if (stretch && sigmas.size() > 2) {
+            float one_minus_last = 1.0f - sigmas[n - 1];
+            float scale_factor   = one_minus_last / (1.0f - target_terminal);
+            if (scale_factor > 1e-8f) {
+                for (uint32_t i = 0; i < n; ++i) {
+                    sigmas[i] = 1.0f - (1.0f - sigmas[i]) / scale_factor;
+                }
+            }
+        }
+
+        sigmas[n] = 0.0f;
+        return sigmas;
+    }
+};
+
 struct Denoiser {
    virtual float sigma_min()                                                        = 0;
    virtual float sigma_max()                                                        = 0;
@@ -491,7 +571,7 @@ struct Denoiser {
    virtual sd::Tensor<float> inverse_noise_scaling(float sigma,
                                                    const sd::Tensor<float>& latent) = 0;

-    virtual std::vector<float> get_sigmas(uint32_t n, int /*image_seq_len*/, scheduler_t scheduler_type, SDVersion version) {
+    virtual std::vector<float> get_sigmas(uint32_t n, int image_seq_len, scheduler_t scheduler_type, SDVersion version, const char* extra_sample_args = nullptr) {
        auto bound_t_to_sigma = std::bind(&Denoiser::t_to_sigma, this, std::placeholders::_1);
        std::shared_ptr<SigmaScheduler> scheduler;
        switch (scheduler_type) {
@@ -539,6 +619,10 @@ struct Denoiser {
                LOG_INFO("get_sigmas with LCM scheduler");
                scheduler = std::make_shared<LCMScheduler>();
                break;
+            case LTX2_SCHEDULER:
+                LOG_INFO("get_sigmas with LTX2 scheduler");
+                scheduler = std::make_shared<LTX2Scheduler>(image_seq_len, extra_sample_args);
+                break;
            default:
                LOG_INFO("get_sigmas with discrete scheduler (default)");
                scheduler = std::make_shared<DiscreteScheduler>();
@@ -744,15 +828,15 @@ struct Flux2FlowDenoiser : public FluxFlowDenoiser {
        return mu;
    }

-    std::vector<float> get_sigmas(uint32_t n, int image_seq_len, scheduler_t scheduler_type, SDVersion version) override {
+    std::vector<float> get_sigmas(uint32_t n, int image_seq_len, scheduler_t scheduler_type, SDVersion version, const char* extra_sample_args = nullptr) override {
        float mu = compute_empirical_mu(n, image_seq_len);
        LOG_DEBUG("Flux2FlowDenoiser: set shift to %.3f", mu);
        set_shift(mu);
-        return Denoiser::get_sigmas(n, image_seq_len, scheduler_type, version);
+        return Denoiser::get_sigmas(n, image_seq_len, scheduler_type, version, extra_sample_args);
    }
 };

-typedef std::function<sd::Tensor<float>(const sd::Tensor<float>&, float, int)> denoise_cb_t;
+typedef std::function<sd::guidance::GuiderOutput(const sd::Tensor<float>&, float, int)> denoise_cb_t;

 static std::pair<float, float> get_ancestral_step(float sigma_from,
                                                  float sigma_to,
@@ -823,45 +907,33 @@ static std::tuple<float, float, float> get_ancestral_step(float sigma_from,
 static sd::Tensor<float> sample_euler_ancestral(denoise_cb_t model,
                                                sd::Tensor<float> x,
                                                const std::vector<float>& sigmas,
-                                                std::shared_ptr<RNG> rng,
-                                                float eta) {
+                                                std::shared_ptr<RNG> rng = nullptr,
+                                                bool is_flow_denoiser    = false,
+                                                float eta                = 0.f) {
    int steps = static_cast<int>(sigmas.size()) - 1;
    for (int i = 0; i < steps; i++) {
        float sigma       = sigmas[i];
+        float sigma_to    = sigmas[i + 1];
        auto denoised_opt = model(x, sigma, i + 1);
-        if (denoised_opt.empty()) {
+        if (denoised_opt.pred.empty()) {
            return {};
        }
-        sd::Tensor<float> denoised  = std::move(denoised_opt);
-        sd::Tensor<float> d         = (x - denoised) / sigma;
-        auto [sigma_down, sigma_up] = get_ancestral_step(sigmas[i], sigmas[i + 1], eta);
-        x += d * (sigma_down - sigmas[i]);
-        if (sigmas[i + 1] > 0) {
-            x += sd::Tensor<float>::randn_like(x, rng) * sigma_up;
-        }
-    }
-    return x;
-}
-
-static sd::Tensor<float> sample_euler_flow(denoise_cb_t model,
-                                           sd::Tensor<float> x,
-                                           const std::vector<float>& sigmas,
-                                           std::shared_ptr<RNG> rng,
-                                           float eta) {
-    int steps = static_cast<int>(sigmas.size()) - 1;
-    for (int i = 0; i < steps; i++) {
-        float sigma       = sigmas[i];
-        auto denoised_opt = model(x, sigma, i + 1);
-        if (denoised_opt.empty()) {
-            return {};
-        }
-        sd::Tensor<float> denoised               = std::move(denoised_opt);
-        auto [sigma_down, sigma_up, alpha_scale] = get_ancestral_step_flow(sigma, sigmas[i + 1], eta);
-        float sigma_ratio                        = sigma_down / sigma;
-        x                                        = sigma_ratio * x + (1.0f - sigma_ratio) * denoised;
-
-        if (sigma_up > 0.0f) {
-            x = alpha_scale * x + sd::Tensor<float>::randn_like(x, rng) * sigma_up;
+        sd::Tensor<float> denoised = std::move(denoised_opt.pred);
+        if (sigma_to == 0.f) {
+            x = denoised;
+        } else if (eta == 0.f) {
+            float sigma_ratio = sigma_to / sigma;
+            x                 = sigma_ratio * x + (1.0 - sigma_ratio) * denoised;
+        } else {
+            auto [sigma_down, sigma_up, alpha_scale] = get_ancestral_step(sigma, sigma_to, eta, is_flow_denoiser);
+            float sigma_ratio                        = sigma_down / sigma;
+            x                                        = sigma_ratio * x + (1.0f - sigma_ratio) * denoised;
+            if (sigma_up > 0.f) {
+                if (is_flow_denoiser) {
+                    x *= alpha_scale;
+                }
+                x += sd::Tensor<float>::randn_like(x, rng) * sigma_up;
+            }
        }
    }
    return x;
@@ -874,10 +946,10 @@ static sd::Tensor<float> sample_euler(denoise_cb_t model,
    for (int i = 0; i < steps; i++) {
        float sigma       = sigmas[i];
        auto denoised_opt = model(x, sigma, i + 1);
-        if (denoised_opt.empty()) {
+        if (denoised_opt.pred.empty()) {
            return {};
        }
-        sd::Tensor<float> denoised = std::move(denoised_opt);
+        sd::Tensor<float> denoised = std::move(denoised_opt.pred);
        sd::Tensor<float> d        = (x - denoised) / sigma;
        x += d * (sigmas[i + 1] - sigma);
    }
@@ -890,10 +962,10 @@ static sd::Tensor<float> sample_heun(denoise_cb_t model,
    int steps = static_cast<int>(sigmas.size()) - 1;
    for (int i = 0; i < steps; i++) {
        auto denoised_opt = model(x, sigmas[i], -(i + 1));
-        if (denoised_opt.empty()) {
+        if (denoised_opt.pred.empty()) {
            return {};
        }
-        sd::Tensor<float> denoised = std::move(denoised_opt);
+        sd::Tensor<float> denoised = std::move(denoised_opt.pred);
        sd::Tensor<float> d        = (x - denoised) / sigmas[i];
        float dt                   = sigmas[i + 1] - sigmas[i];
        if (sigmas[i + 1] == 0) {
@@ -901,10 +973,10 @@ static sd::Tensor<float> sample_heun(denoise_cb_t model,
        } else {
            sd::Tensor<float> x2 = x + d * dt;
            auto denoised2_opt   = model(x2, sigmas[i + 1], i + 1);
-            if (denoised2_opt.empty()) {
+            if (denoised2_opt.pred.empty()) {
                return {};
            }
-            sd::Tensor<float> denoised2 = std::move(denoised2_opt);
+            sd::Tensor<float> denoised2 = std::move(denoised2_opt.pred);
            d                           = (d + (x2 - denoised2) / sigmas[i + 1]) / 2.0f;
            x += d * dt;
        }
@@ -918,10 +990,10 @@ static sd::Tensor<float> sample_dpm2(denoise_cb_t model,
    int steps = static_cast<int>(sigmas.size()) - 1;
    for (int i = 0; i < steps; i++) {
        auto denoised_opt = model(x, sigmas[i], -(i + 1));
-        if (denoised_opt.empty()) {
+        if (denoised_opt.pred.empty()) {
            return {};
        }
-        sd::Tensor<float> denoised = std::move(denoised_opt);
+        sd::Tensor<float> denoised = std::move(denoised_opt.pred);
        sd::Tensor<float> d        = (x - denoised) / sigmas[i];
        if (sigmas[i + 1] == 0) {
            x += d * (sigmas[i + 1] - sigmas[i]);
@@ -931,10 +1003,10 @@ static sd::Tensor<float> sample_dpm2(denoise_cb_t model,
            float dt_2           = sigmas[i + 1] - sigmas[i];
            sd::Tensor<float> x2 = x + d * dt_1;
            auto denoised2_opt   = model(x2, sigma_mid, i + 1);
-            if (denoised2_opt.empty()) {
+            if (denoised2_opt.pred.empty()) {
                return {};
            }
-            sd::Tensor<float> denoised2 = std::move(denoised2_opt);
+            sd::Tensor<float> denoised2 = std::move(denoised2_opt.pred);
            x += ((x2 - denoised2) / sigma_mid) * dt_2;
        }
    }
@@ -952,10 +1024,10 @@ static sd::Tensor<float> sample_dpmpp_2s_ancestral(denoise_cb_t model,
    int steps = static_cast<int>(sigmas.size()) - 1;
    for (int i = 0; i < steps; i++) {
        auto denoised_opt = model(x, sigmas[i], -(i + 1));
-        if (denoised_opt.empty()) {
+        if (denoised_opt.pred.empty()) {
            return {};
        }
-        sd::Tensor<float> denoised  = std::move(denoised_opt);
+        sd::Tensor<float> denoised  = std::move(denoised_opt.pred);
        auto [sigma_down, sigma_up] = get_ancestral_step(sigmas[i], sigmas[i + 1], eta);

        if (sigma_down == 0) {
@@ -968,10 +1040,10 @@ static sd::Tensor<float> sample_dpmpp_2s_ancestral(denoise_cb_t model,
            float sigma_s        = sigma_fn(s);
            sd::Tensor<float> x2 = (sigma_s / sigma_fn(t)) * x - (exp(-h * 0.5f) - 1) * denoised;
            auto denoised2_opt   = model(x2, sigma_s, i + 1);
-            if (denoised2_opt.empty()) {
+            if (denoised2_opt.pred.empty()) {
                return {};
            }
-            sd::Tensor<float> denoised2 = std::move(denoised2_opt);
+            sd::Tensor<float> denoised2 = std::move(denoised2_opt.pred);
            x                           = (sigma_fn(t_next) / sigma_fn(t)) * x - (exp(-h) - 1) * denoised2;
        }

@@ -995,10 +1067,10 @@ static sd::Tensor<float> sample_dpmpp_2s_ancestral_flow(denoise_cb_t model,
        bool opt_first_step = (1.0 - sigma < 1e-6);

        auto denoised_opt = model(x, sigma, (opt_first_step ? 1 : -1) * (i + 1));
-        if (denoised_opt.empty()) {
+        if (denoised_opt.pred.empty()) {
            return {};
        }
-        sd::Tensor<float> denoised = std::move(denoised_opt);
+        sd::Tensor<float> denoised = std::move(denoised_opt.pred);

        if (sigma_to == 0.0f) {
            // Euler method (final step, no noise)
@@ -1058,10 +1130,10 @@ static sd::Tensor<float> sample_dpmpp_2s_ancestral_flow(denoise_cb_t model,
                sd::Tensor<float> u   = (x * sigma_s_i_ratio) + (denoised * (1.0f - sigma_s_i_ratio));

                auto denoised2_opt = model(u, sigma_s, i + 1);
-                if (denoised2_opt.empty()) {
+                if (denoised2_opt.pred.empty()) {
                    return {};
                }
-                D_i = std::move(denoised2_opt);
+                D_i = std::move(denoised2_opt.pred);
            }

            float sigma_down_i_ratio = sigma_down / sigma;
@@ -1085,10 +1157,10 @@ static sd::Tensor<float> sample_dpmpp_2m(denoise_cb_t model,
    int steps = static_cast<int>(sigmas.size()) - 1;
    for (int i = 0; i < steps; i++) {
        auto denoised_opt = model(x, sigmas[i], i + 1);
-        if (denoised_opt.empty()) {
+        if (denoised_opt.pred.empty()) {
            return {};
        }
-        sd::Tensor<float> denoised = std::move(denoised_opt);
+        sd::Tensor<float> denoised = std::move(denoised_opt.pred);
        float t                    = t_fn(sigmas[i]);
        float t_next               = t_fn(sigmas[i + 1]);
        float h                    = t_next - t;
@@ -1117,10 +1189,10 @@ static sd::Tensor<float> sample_dpmpp_2m_v2(denoise_cb_t model,
    int steps = static_cast<int>(sigmas.size()) - 1;
    for (int i = 0; i < steps; i++) {
        auto denoised_opt = model(x, sigmas[i], i + 1);
-        if (denoised_opt.empty()) {
+        if (denoised_opt.pred.empty()) {
            return {};
        }
-        sd::Tensor<float> denoised = std::move(denoised_opt);
+        sd::Tensor<float> denoised = std::move(denoised_opt.pred);
        float t                    = t_fn(sigmas[i]);
        float t_next               = t_fn(sigmas[i + 1]);
        float h                    = t_next - t;
@@ -1144,23 +1216,85 @@ static sd::Tensor<float> sample_dpmpp_2m_v2(denoise_cb_t model,
    return x;
 }

+using SamplerExtraArgs = KeyValueArgs;
+
 static sd::Tensor<float> sample_lcm(denoise_cb_t model,
                                    sd::Tensor<float> x,
                                    const std::vector<float>& sigmas,
                                    std::shared_ptr<RNG> rng,
-                                    bool is_flow_denoiser) {
+                                    bool is_flow_denoiser,
+                                    const SamplerExtraArgs& extra_sample_args) {
+    struct LCMSampleArgs {
+        float noise_clip_std    = 0.0f;
+        float noise_scale_start = 1.0f;
+        float noise_scale_end   = 1.0f;
+    };
+
+    LCMSampleArgs args;
+    bool noise_scale_end_was_set   = false;
+    bool noise_scale_start_was_set = false;
+
+    for (const auto& [key, value] : extra_sample_args) {
+        float parsed = 0.0f;
+        if (key == "noise_clip_std") {
+            if (!parse_strict_float(value, parsed)) {
+                LOG_WARN("ignoring invalid lcm extra sample arg '%s=%s'", key.c_str(), value.c_str());
+                continue;
+            }
+            args.noise_clip_std = parsed;
+        } else if (key == "noise_scale_start") {
+            if (!parse_strict_float(value, parsed)) {
+                LOG_WARN("ignoring invalid lcm extra sample arg '%s=%s'", key.c_str(), value.c_str());
+                continue;
+            }
+            args.noise_scale_start    = parsed;
+            noise_scale_start_was_set = true;
+        } else if (key == "noise_scale_end") {
+            if (!parse_strict_float(value, parsed)) {
+                LOG_WARN("ignoring invalid lcm extra sample arg '%s=%s'", key.c_str(), value.c_str());
+                continue;
+            }
+            args.noise_scale_end    = parsed;
+            noise_scale_end_was_set = true;
+        }
+    }
+
+    if (noise_scale_start_was_set && !noise_scale_end_was_set) {
+        args.noise_scale_end = args.noise_scale_start;
+    }
+
    int steps = static_cast<int>(sigmas.size()) - 1;
    for (int i = 0; i < steps; i++) {
        auto denoised_opt = model(x, sigmas[i], i + 1);
-        if (denoised_opt.empty()) {
+        if (denoised_opt.pred.empty()) {
            return {};
        }
-        x = std::move(denoised_opt);
+        x = std::move(denoised_opt.pred);
        if (sigmas[i + 1] > 0) {
            if (is_flow_denoiser) {
                x *= (1 - sigmas[i + 1]);
            }
-            x += sd::Tensor<float>::randn_like(x, rng) * sigmas[i + 1];
+            auto noise = sd::Tensor<float>::randn_like(x, rng);
+            if (args.noise_clip_std > 0.0f && noise.numel() > 0) {
+                double mean = 0.0;
+                for (int64_t j = 0; j < noise.numel(); ++j) {
+                    mean += static_cast<double>(noise[j]);
+                }
+                mean /= static_cast<double>(noise.numel());
+
+                double variance = 0.0;
+                for (int64_t j = 0; j < noise.numel(); ++j) {
+                    double centered = static_cast<double>(noise[j]) - mean;
+                    variance += centered * centered;
+                }
+                variance /= static_cast<double>(noise.numel());
+
+                float clip_val = args.noise_clip_std * static_cast<float>(std::sqrt(variance));
+                noise          = sd::ops::clamp(noise, -clip_val, clip_val);
+            }
+            float t           = steps > 1 ? static_cast<float>(i) / static_cast<float>(steps - 1) : 0.0f;
+            float noise_scale = args.noise_scale_start + (args.noise_scale_end - args.noise_scale_start) * t;
+            x += noise * (sigmas[i + 1] * noise_scale);
        }
    }
    return x;
@@ -1178,10 +1312,10 @@ static sd::Tensor<float> sample_ipndm(denoise_cb_t model,
        float sigma_next = sigmas[i + 1];

        auto denoised_opt = model(x, sigma, i + 1);
-        if (denoised_opt.empty()) {
+        if (denoised_opt.pred.empty()) {
            return {};
        }
-        sd::Tensor<float> denoised = std::move(denoised_opt);
+        sd::Tensor<float> denoised = std::move(denoised_opt.pred);

        sd::Tensor<float> d_cur = (x - denoised) / sigma;
        int order               = std::min(max_order, i + 1);
@@ -1222,10 +1356,10 @@ static sd::Tensor<float> sample_ipndm_v(denoise_cb_t model,
        float t_next = sigmas[i + 1];

        auto denoised_opt = model(x, sigma, i + 1);
-        if (denoised_opt.empty()) {
+        if (denoised_opt.pred.empty()) {
            return {};
        }
-        sd::Tensor<float> denoised = std::move(denoised_opt);
+        sd::Tensor<float> denoised = std::move(denoised_opt.pred);

        sd::Tensor<float> d_cur = (x - denoised) / sigma;
        int order               = std::min(max_order, i + 1);
@@ -1284,13 +1418,13 @@ static sd::Tensor<float> sample_res_multistep(denoise_cb_t model,
    int steps = static_cast<int>(sigmas.size()) - 1;
    for (int i = 0; i < steps; i++) {
        auto denoised_opt = model(x, sigmas[i], i + 1);
-        if (denoised_opt.empty()) {
+        if (denoised_opt.pred.empty()) {
            return {};
        }
-        sd::Tensor<float> denoised = std::move(denoised_opt);
+        sd::Tensor<float> denoised = std::move(denoised_opt.pred);

-        float sigma_from            = sigmas[i];
-        float sigma_to              = sigmas[i + 1];
+        float sigma_from = sigmas[i];
+        float sigma_to   = sigmas[i + 1];

        auto [sigma_down, sigma_up, alpha_scale] = get_ancestral_step(sigma_from, sigma_to, eta, is_flow_denoiser);

@@ -1361,10 +1495,10 @@ static sd::Tensor<float> sample_res_2s(denoise_cb_t model,
        float sigma_to   = sigmas[i + 1];

        auto denoised_opt = model(x, sigma_from, -(i + 1));
-        if (denoised_opt.empty()) {
+        if (denoised_opt.pred.empty()) {
            return {};
        }
-        sd::Tensor<float> denoised = std::move(denoised_opt);
+        sd::Tensor<float> denoised = std::move(denoised_opt.pred);

        auto [sigma_down, sigma_up, alpha_scale] = get_ancestral_step(sigma_from, sigma_to, eta, is_flow_denoiser);

@@ -1387,10 +1521,10 @@ static sd::Tensor<float> sample_res_2s(denoise_cb_t model,
            sd::Tensor<float> x2   = x0 + eps1 * (h * a21);

            auto denoised2_opt = model(x2, sigma_c2, i + 1);
-            if (denoised2_opt.empty()) {
+            if (denoised2_opt.pred.empty()) {
                return {};
            }
-            sd::Tensor<float> denoised2 = std::move(denoised2_opt);
+            sd::Tensor<float> denoised2 = std::move(denoised2_opt.pred);
            sd::Tensor<float> eps2      = denoised2 - x0;
            x                           = x0 + h * (b1 * eps1 + b2 * eps2);
        }
@@ -1463,10 +1597,11 @@ static sd::Tensor<float> sample_er_sde(denoise_cb_t model,

    int steps = static_cast<int>(sigmas.size()) - 1;
    for (int i = 0; i < steps; i++) {
-        sd::Tensor<float> denoised = model(x, sigmas[i], i + 1);
-        if (denoised.empty()) {
+        auto denoised_opt = model(x, sigmas[i], i + 1);
+        if (denoised_opt.pred.empty()) {
            return {};
        }
+        sd::Tensor<float> denoised = std::move(denoised_opt.pred);

        int stage_used = std::min(max_stage, i + 1);

@@ -1539,46 +1674,6 @@ static sd::Tensor<float> sample_er_sde(denoise_cb_t model,
    return x;
 }

-static sd::Tensor<float> sample_ddim_trailing(denoise_cb_t model,
-                                              sd::Tensor<float> x,
-                                              const std::vector<float>& sigmas,
-                                              std::shared_ptr<RNG> rng,
-                                              float eta) {
-    int steps = static_cast<int>(sigmas.size()) - 1;
-    for (int i = 0; i < steps; i++) {
-        float sigma    = sigmas[i];
-        float sigma_to = sigmas[i + 1];
-
-        auto model_output_opt = model(x, sigma, i + 1);
-        if (model_output_opt.empty()) {
-            return {};
-        }
-        sd::Tensor<float> model_output = std::move(model_output_opt);
-        model_output                   = (x - model_output) * (1.0f / sigma);
-
-        float alpha_prod_t      = 1.0f / (sigma * sigma + 1.0f);
-        float alpha_prod_t_prev = 1.0f / (sigma_to * sigma_to + 1.0f);
-        float beta_prod_t       = 1.0f - alpha_prod_t;
-
-        sd::Tensor<float> pred_original_sample = ((x / std::sqrt(sigma * sigma + 1)) -
-                                                  std::sqrt(beta_prod_t) * model_output) *
-                                                 (1.0f / std::sqrt(alpha_prod_t));
-
-        float beta_prod_t_prev = 1.0f - alpha_prod_t_prev;
-        float variance         = (beta_prod_t_prev / beta_prod_t) *
-                         (1.0f - alpha_prod_t / alpha_prod_t_prev);
-        float std_dev_t = eta * std::sqrt(variance);
-
-        x = pred_original_sample +
-            std::sqrt((1.0f - alpha_prod_t_prev - std::pow(std_dev_t, 2)) / alpha_prod_t_prev) * model_output;
-
-        if (eta > 0) {
-            x += std_dev_t / std::sqrt(alpha_prod_t_prev) * sd::Tensor<float>::randn_like(x, rng);
-        }
-    }
-    return x;
-}
-
 static sd::Tensor<float> sample_tcd(denoise_cb_t model,
                                    sd::Tensor<float> x,
                                    const std::vector<float>& sigmas,
@@ -1621,12 +1716,12 @@ static sd::Tensor<float> sample_tcd(denoise_cb_t model,
        int timestep_s    = (int)floor((1 - eta) * prev_timestep);
        float sigma       = sigmas[i];

-        auto model_output_opt = model(x, sigma, i + 1);
-        if (model_output_opt.empty()) {
+        auto denoised_opt = model(x, sigma, i + 1);
+        if (denoised_opt.pred.empty()) {
            return {};
        }
-        sd::Tensor<float> model_output = std::move(model_output_opt);
-        model_output                   = (x - model_output) * (1.0f / sigma);
+        sd::Tensor<float> denoised = std::move(denoised_opt.pred);
+        sd::Tensor<float> d        = (x - denoised) / sigma;

        float alpha_prod_t      = 1.0f / (sigma * sigma + 1.0f);
        float beta_prod_t       = 1.0f - alpha_prod_t;
@@ -1634,12 +1729,8 @@ static sd::Tensor<float> sample_tcd(denoise_cb_t model,
        float alpha_prod_s      = static_cast<float>(alphas_cumprod[timestep_s]);
        float beta_prod_s       = 1.0f - alpha_prod_s;

-        sd::Tensor<float> pred_original_sample = ((x / std::sqrt(sigma * sigma + 1)) -
-                                                  std::sqrt(beta_prod_t) * model_output) *
-                                                 (1.0f / std::sqrt(alpha_prod_t));
-
-        x = std::sqrt(alpha_prod_s / alpha_prod_t_prev) * pred_original_sample +
-            std::sqrt(beta_prod_s / alpha_prod_t_prev) * model_output;
+        x = std::sqrt(alpha_prod_s / alpha_prod_t_prev) * denoised +
+            std::sqrt(beta_prod_s / alpha_prod_t_prev) * d;

        if (eta > 0 && sigma_to > 0.0f) {
            x = std::sqrt(alpha_prod_t_prev / alpha_prod_s) * x +
@@ -1649,6 +1740,112 @@ static sd::Tensor<float> sample_tcd(denoise_cb_t model,
    return x;
 }

+static sd::Tensor<float> sample_euler_cfg_pp(denoise_cb_t model,
+                                             sd::Tensor<float> x,
+                                             const std::vector<float>& sigmas) {
+    int steps = static_cast<int>(sigmas.size()) - 1;
+    for (int i = 0; i < steps; i++) {
+        float sigma       = sigmas[i];
+        auto denoised_opt = model(x, sigma, i + 1);
+        if (denoised_opt.pred.empty() || denoised_opt.pred_uncond.empty()) {
+            return {};
+        }
+
+        sd::Tensor<float> denoised        = std::move(denoised_opt.pred);
+        sd::Tensor<float> uncond_denoised = std::move(denoised_opt.pred_uncond);
+        sd::Tensor<float> d               = (x - uncond_denoised) / sigma;
+
+        x = denoised + d * sigmas[i + 1];
+    }
+    return x;
+}
+
+static sd::Tensor<float> sample_euler_ancestral_cfg_pp(denoise_cb_t model,
+                                                       sd::Tensor<float> x,
+                                                       const std::vector<float>& sigmas,
+                                                       std::shared_ptr<RNG> rng,
+                                                       float eta) {
+    int steps = static_cast<int>(sigmas.size()) - 1;
+    for (int i = 0; i < steps; i++) {
+        float sigma       = sigmas[i];
+        auto denoised_opt = model(x, sigma, i + 1);
+        if (denoised_opt.pred.empty() || denoised_opt.pred_uncond.empty()) {
+            return {};
+        }
+
+        sd::Tensor<float> denoised        = std::move(denoised_opt.pred);
+        sd::Tensor<float> uncond_denoised = std::move(denoised_opt.pred_uncond);
+        sd::Tensor<float> d               = (x - uncond_denoised) / sigma;
+
+        auto [sigma_down, sigma_up] = get_ancestral_step(sigmas[i], sigmas[i + 1], eta);
+
+        x = denoised + d * sigma_down;
+
+        if (sigmas[i + 1] > 0) {
+            x += sd::Tensor<float>::randn_like(x, rng) * sigma_up;
+        }
+    }
+    return x;
+}
+
+// https://github.com/ToyotaResearchInstitute/gradient-estimation-sampler
+static sd::Tensor<float> sample_gradient_estimation(denoise_cb_t model,
+                                                    sd::Tensor<float> x,
+                                                    const std::vector<float>& sigmas,
+                                                    std::shared_ptr<RNG> rng,
+                                                    bool is_flow_denoiser,
+                                                    float eta,
+                                                    const SamplerExtraArgs& extra_sample_args) {
+    float ge_gamma = 2.0f;
+
+    for (const auto& [key, value] : extra_sample_args) {
+        if (key == "gamma") {
+            float parsed = 0.0f;
+            if (!parse_strict_float(value, parsed)) {
+                LOG_WARN("ignoring invalid euler_ge extra sample arg '%s=%s'", key.c_str(), value.c_str());
+                continue;
+            }
+            LOG_DEBUG("setting euler_ge gamma to %.2f", parsed);
+            ge_gamma = parsed;
+        }
+    }
+
+    int steps = static_cast<int>(sigmas.size()) - 1;
+    sd::Tensor<float> old_d;
+    bool has_old_d = false;
+    for (int i = 0; i < steps; i++) {
+        float sigma       = sigmas[i];
+        float sigma_to    = sigmas[i + 1];
+        auto denoised_opt = model(x, sigma, i + 1);
+        if (denoised_opt.pred.empty()) {
+            return {};
+        }
+        sd::Tensor<float> denoised = std::move(denoised_opt.pred);
+        if (sigma_to == 0.f) {
+            x = denoised;
+        } else {
+            auto [sigma_down, sigma_up, alpha_scale] = get_ancestral_step(sigma, sigma_to, eta, is_flow_denoiser);
+            sd::Tensor<float> d                      = (x - denoised) / sigma;
+            float dt                                 = sigma_down - sigma;
+            if (has_old_d) {
+                sd::Tensor<float> d_bar = d * ge_gamma + old_d * (1.0f - ge_gamma);
+                x += d_bar * dt;
+            } else {
+                x += d * dt;
+            }
+            old_d     = std::move(d);
+            has_old_d = true;
+            if (sigma_up > 0.f) {
+                if (is_flow_denoiser) {
+                    x *= alpha_scale;
+                }
+                x += sd::Tensor<float>::randn_like(x, rng) * sigma_up;
+            }
+        }
+    }
+    return x;
+}
+
 // k diffusion reverse ODE: dx = (x - D(x;\sigma)) / \sigma dt; \sigma(t) = t
 static sd::Tensor<float> sample_k_diffusion(sample_method_t method,
                                            denoise_cb_t model,
@@ -1656,13 +1853,12 @@ static sd::Tensor<float> sample_k_diffusion(sample_method_t method,
                                            std::vector<float> sigmas,
                                            std::shared_ptr<RNG> rng,
                                            float eta,
-                                            bool is_flow_denoiser) {
+                                            bool is_flow_denoiser,
+                                            const char* extra_sample_args) {
+    SamplerExtraArgs extra_args = parse_key_value_args(extra_sample_args, "extra sample arg");
    switch (method) {
        case EULER_A_SAMPLE_METHOD:
-            if (is_flow_denoiser)
-                return sample_euler_flow(model, std::move(x), sigmas, rng, eta);
-            else
-                return sample_euler_ancestral(model, std::move(x), sigmas, rng, eta);
+            return sample_euler_ancestral(model, std::move(x), sigmas, rng, is_flow_denoiser, eta);
        case EULER_SAMPLE_METHOD:
            return sample_euler(model, std::move(x), sigmas);
        case HEUN_SAMPLE_METHOD:
@@ -1679,7 +1875,7 @@ static sd::Tensor<float> sample_k_diffusion(sample_method_t method,
        case DPMPP2Mv2_SAMPLE_METHOD:
            return sample_dpmpp_2m_v2(model, std::move(x), sigmas);
        case LCM_SAMPLE_METHOD:
-            return sample_lcm(model, std::move(x), sigmas, rng, is_flow_denoiser);
+            return sample_lcm(model, std::move(x), sigmas, rng, is_flow_denoiser, extra_args);
        case IPNDM_SAMPLE_METHOD:
            return sample_ipndm(model, std::move(x), sigmas);
        case IPNDM_V_SAMPLE_METHOD:
@@ -1691,9 +1887,16 @@ static sd::Tensor<float> sample_k_diffusion(sample_method_t method,
        case ER_SDE_SAMPLE_METHOD:
            return sample_er_sde(model, std::move(x), sigmas, rng, is_flow_denoiser, eta);
        case DDIM_TRAILING_SAMPLE_METHOD:
-            return sample_ddim_trailing(model, std::move(x), sigmas, rng, eta);
+            // DDIM is equivalent to Euler Ancestral with the Simple scheduler
+            return sample_euler_ancestral(model, std::move(x), sigmas, rng, is_flow_denoiser, eta);
        case TCD_SAMPLE_METHOD:
            return sample_tcd(model, std::move(x), sigmas, rng, eta);
+        case EULER_CFG_PP_SAMPLE_METHOD:
+            return sample_euler_cfg_pp(model, std::move(x), sigmas);
+        case EULER_A_CFG_PP_SAMPLE_METHOD:
+            return sample_euler_ancestral_cfg_pp(model, std::move(x), sigmas, rng, eta);
+        case EULER_GE_SAMPLE_METHOD:
+            return sample_gradient_estimation(model, std::move(x), sigmas, rng, is_flow_denoiser, eta, extra_args);
        default:
            return {};
    }
@@ -1,16 +1,62 @@
 #ifndef __DIFFUSION_MODEL_H__
 #define __DIFFUSION_MODEL_H__

-#include <optional>
-#include "anima.hpp"
-#include "ernie_image.hpp"
-#include "flux.hpp"
-#include "mmdit.hpp"
-#include "qwen_image.hpp"
+#include <string>
+#include <utility>
+#include <variant>
+
+#include "ggml_extend.hpp"
 #include "tensor_ggml.hpp"
-#include "unet.hpp"
-#include "wan.hpp"
-#include "z_image.hpp"
+
+struct UNetDiffusionExtra {
+    int num_video_frames                           = -1;
+    const std::vector<sd::Tensor<float>>* controls = nullptr;
+    float control_strength                         = 0.f;
+};
+
+struct SkipLayerDiffusionExtra {
+    const std::vector<int>* skip_layers = nullptr;
+};
+
+struct FluxDiffusionExtra {
+    const sd::Tensor<float>* guidance   = nullptr;
+    const std::vector<int>* skip_layers = nullptr;
+};
+
+struct AnimaDiffusionExtra {
+    const sd::Tensor<int32_t>* t5_ids   = nullptr;
+    const sd::Tensor<float>* t5_weights = nullptr;
+};
+
+struct WanDiffusionExtra {
+    const sd::Tensor<float>* vace_context = nullptr;
+    float vace_strength                   = 1.f;
+};
+
+struct HiDreamO1DiffusionExtra {
+    const sd::Tensor<int32_t>* input_ids                               = nullptr;
+    const sd::Tensor<int32_t>* input_pos                               = nullptr;
+    const sd::Tensor<int32_t>* token_types                             = nullptr;
+    const sd::Tensor<int32_t>* vinput_mask                             = nullptr;
+    const std::vector<std::pair<int, sd::Tensor<float>>>* image_embeds = nullptr;
+};
+
+struct LTXAVDiffusionExtra {
+    const sd::Tensor<float>* audio_x         = nullptr;
+    const sd::Tensor<float>* audio_timesteps = nullptr;
+    int audio_length                         = 0;
+    float frame_rate                         = 24.f;
+    const sd::Tensor<float>* video_positions = nullptr;
+};
+
+using DiffusionExtraParams = std::variant<std::monostate,
+                                          UNetDiffusionExtra,
+                                          SkipLayerDiffusionExtra,
+                                          FluxDiffusionExtra,
+                                          AnimaDiffusionExtra,
+                                          WanDiffusionExtra,
+                                          HiDreamO1DiffusionExtra,
+                                          LTXAVDiffusionExtra>;

 struct DiffusionParams {
    const sd::Tensor<float>* x                        = nullptr;
@@ -18,598 +64,44 @@ struct DiffusionParams {
    const sd::Tensor<float>* context                  = nullptr;
    const sd::Tensor<float>* c_concat                 = nullptr;
    const sd::Tensor<float>* y                        = nullptr;
-    const sd::Tensor<int32_t>* t5_ids                 = nullptr;
-    const sd::Tensor<float>* t5_weights               = nullptr;
-    const sd::Tensor<float>* guidance                 = nullptr;
    const std::vector<sd::Tensor<float>>* ref_latents = nullptr;
    bool increase_ref_index                           = false;
-    int num_video_frames                              = -1;
-    const std::vector<sd::Tensor<float>>* controls    = nullptr;
-    float control_strength                            = 0.f;
-    const sd::Tensor<float>* vace_context             = nullptr;
-    float vace_strength                               = 1.f;
-    const std::vector<int>* skip_layers               = nullptr;
+    DiffusionExtraParams extra                        = std::monostate{};
 };

+template <typename T>
+static inline const T* diffusion_extra_as(const DiffusionParams& params) {
+    const auto* extra = std::get_if<T>(&params.extra);
+    GGML_ASSERT(extra != nullptr);
+    return extra;
+}
+
 template <typename T>
 static inline const sd::Tensor<T>& tensor_or_empty(const sd::Tensor<T>* tensor) {
    static const sd::Tensor<T> kEmpty;
    return tensor != nullptr ? *tensor : kEmpty;
 }

-struct DiffusionModel {
-    virtual std::string get_desc()                                               = 0;
+struct DiffusionModelRunner : public GGMLRunner {
+protected:
+    std::string prefix;
+
+public:
+    DiffusionModelRunner(ggml_backend_t backend,
+                         ggml_backend_t params_backend,
+                         const std::string& prefix)
+        : GGMLRunner(backend, params_backend),
+          prefix(prefix) {}
+
    virtual sd::Tensor<float> compute(int n_threads,
-                                      const DiffusionParams& diffusion_params)   = 0;
-    virtual void alloc_params_buffer()                                           = 0;
-    virtual void free_params_buffer()                                            = 0;
-    virtual void free_compute_buffer()                                           = 0;
-    virtual void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) = 0;
-    virtual size_t get_params_buffer_size()                                      = 0;
-    virtual void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter){};
-    virtual int64_t get_adm_in_channels()                            = 0;
-    virtual void set_flash_attention_enabled(bool enabled)           = 0;
-    virtual void set_max_graph_vram_bytes(size_t max_vram_bytes)     = 0;
-    virtual void set_circular_axes(bool circular_x, bool circular_y) = 0;
-};
+                                      const DiffusionParams& diffusion_params) = 0;

-struct UNetModel : public DiffusionModel {
-    UNetModelRunner unet;
-
-    UNetModel(ggml_backend_t backend,
-              bool offload_params_to_cpu,
-              const String2TensorStorage& tensor_storage_map = {},
-              SDVersion version                              = VERSION_SD1)
-        : unet(backend, offload_params_to_cpu, tensor_storage_map, "model.diffusion_model", version) {
-    }
-
-    std::string get_desc() override {
-        return unet.get_desc();
-    }
-
-    void alloc_params_buffer() override {
-        unet.alloc_params_buffer();
-    }
-
-    void free_params_buffer() override {
-        unet.free_params_buffer();
-    }
-
-    void free_compute_buffer() override {
-        unet.free_compute_buffer();
-    }
-
-    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
-        unet.get_param_tensors(tensors, "model.diffusion_model");
-    }
-
-    size_t get_params_buffer_size() override {
-        return unet.get_params_buffer_size();
-    }
-
-    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
-        unet.set_weight_adapter(adapter);
-    }
-
-    int64_t get_adm_in_channels() override {
-        return unet.unet.adm_in_channels;
-    }
-
-    void set_flash_attention_enabled(bool enabled) {
-        unet.set_flash_attention_enabled(enabled);
-    }
-
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        unet.set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
-    void set_circular_axes(bool circular_x, bool circular_y) override {
-        unet.set_circular_axes(circular_x, circular_y);
-    }
-
-    sd::Tensor<float> compute(int n_threads,
-                              const DiffusionParams& diffusion_params) override {
-        GGML_ASSERT(diffusion_params.x != nullptr);
-        GGML_ASSERT(diffusion_params.timesteps != nullptr);
-        static const std::vector<sd::Tensor<float>> empty_controls;
-        return unet.compute(n_threads,
-                            *diffusion_params.x,
-                            *diffusion_params.timesteps,
-                            tensor_or_empty(diffusion_params.context),
-                            tensor_or_empty(diffusion_params.c_concat),
-                            tensor_or_empty(diffusion_params.y),
-                            diffusion_params.num_video_frames,
-                            diffusion_params.controls ? *diffusion_params.controls : empty_controls,
-                            diffusion_params.control_strength);
-    }
-};
-
-struct MMDiTModel : public DiffusionModel {
-    MMDiTRunner mmdit;
-
-    MMDiTModel(ggml_backend_t backend,
-               bool offload_params_to_cpu,
-               const String2TensorStorage& tensor_storage_map = {})
-        : mmdit(backend, offload_params_to_cpu, tensor_storage_map, "model.diffusion_model") {
-    }
-
-    std::string get_desc() override {
-        return mmdit.get_desc();
-    }
-
-    void alloc_params_buffer() override {
-        mmdit.alloc_params_buffer();
-    }
-
-    void free_params_buffer() override {
-        mmdit.free_params_buffer();
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) {
+        get_param_tensors(tensors, prefix);
    }

-    void free_compute_buffer() override {
-        mmdit.free_compute_buffer();
-    }
-
-    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
-        mmdit.get_param_tensors(tensors, "model.diffusion_model");
-    }
-
-    size_t get_params_buffer_size() override {
-        return mmdit.get_params_buffer_size();
-    }
-
-    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
-        mmdit.set_weight_adapter(adapter);
-    }
-
-    int64_t get_adm_in_channels() override {
-        return 768 + 1280;
-    }
-
-    void set_flash_attention_enabled(bool enabled) {
-        mmdit.set_flash_attention_enabled(enabled);
-    }
-
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        mmdit.set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
-    void set_circular_axes(bool circular_x, bool circular_y) override {
-        mmdit.set_circular_axes(circular_x, circular_y);
-    }
-
-    sd::Tensor<float> compute(int n_threads,
-                              const DiffusionParams& diffusion_params) override {
-        GGML_ASSERT(diffusion_params.x != nullptr);
-        GGML_ASSERT(diffusion_params.timesteps != nullptr);
-        static const std::vector<int> empty_skip_layers;
-        return mmdit.compute(n_threads,
-                             *diffusion_params.x,
-                             *diffusion_params.timesteps,
-                             tensor_or_empty(diffusion_params.context),
-                             tensor_or_empty(diffusion_params.y),
-                             diffusion_params.skip_layers ? *diffusion_params.skip_layers : empty_skip_layers);
-    }
-};
-
-struct FluxModel : public DiffusionModel {
-    Flux::FluxRunner flux;
-
-    FluxModel(ggml_backend_t backend,
-              bool offload_params_to_cpu,
-              const String2TensorStorage& tensor_storage_map = {},
-              SDVersion version                              = VERSION_FLUX,
-              bool use_mask                                  = false)
-        : flux(backend, offload_params_to_cpu, tensor_storage_map, "model.diffusion_model", version, use_mask) {
-    }
-
-    std::string get_desc() override {
-        return flux.get_desc();
-    }
-
-    void alloc_params_buffer() override {
-        flux.alloc_params_buffer();
-    }
-
-    void free_params_buffer() override {
-        flux.free_params_buffer();
-    }
-
-    void free_compute_buffer() override {
-        flux.free_compute_buffer();
-    }
-
-    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
-        flux.get_param_tensors(tensors, "model.diffusion_model");
-    }
-
-    size_t get_params_buffer_size() override {
-        return flux.get_params_buffer_size();
-    }
-
-    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
-        flux.set_weight_adapter(adapter);
-    }
-
-    int64_t get_adm_in_channels() override {
-        return 768;
-    }
-
-    void set_flash_attention_enabled(bool enabled) {
-        flux.set_flash_attention_enabled(enabled);
-    }
-
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        flux.set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
-    void set_circular_axes(bool circular_x, bool circular_y) override {
-        flux.set_circular_axes(circular_x, circular_y);
-    }
-
-    sd::Tensor<float> compute(int n_threads,
-                              const DiffusionParams& diffusion_params) override {
-        GGML_ASSERT(diffusion_params.x != nullptr);
-        GGML_ASSERT(diffusion_params.timesteps != nullptr);
-        static const std::vector<sd::Tensor<float>> empty_ref_latents;
-        static const std::vector<int> empty_skip_layers;
-        return flux.compute(n_threads,
-                            *diffusion_params.x,
-                            *diffusion_params.timesteps,
-                            tensor_or_empty(diffusion_params.context),
-                            tensor_or_empty(diffusion_params.c_concat),
-                            tensor_or_empty(diffusion_params.y),
-                            tensor_or_empty(diffusion_params.guidance),
-                            diffusion_params.ref_latents ? *diffusion_params.ref_latents : empty_ref_latents,
-                            diffusion_params.increase_ref_index,
-                            diffusion_params.skip_layers ? *diffusion_params.skip_layers : empty_skip_layers);
-    }
-};
-
-struct AnimaModel : public DiffusionModel {
-    std::string prefix;
-    Anima::AnimaRunner anima;
-
-    AnimaModel(ggml_backend_t backend,
-               bool offload_params_to_cpu,
-               const String2TensorStorage& tensor_storage_map = {},
-               const std::string prefix                       = "model.diffusion_model")
-        : prefix(prefix), anima(backend, offload_params_to_cpu, tensor_storage_map, prefix) {
-    }
-
-    std::string get_desc() override {
-        return anima.get_desc();
-    }
-
-    void alloc_params_buffer() override {
-        anima.alloc_params_buffer();
-    }
-
-    void free_params_buffer() override {
-        anima.free_params_buffer();
-    }
-
-    void free_compute_buffer() override {
-        anima.free_compute_buffer();
-    }
-
-    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
-        anima.get_param_tensors(tensors, prefix);
-    }
-
-    size_t get_params_buffer_size() override {
-        return anima.get_params_buffer_size();
-    }
-
-    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
-        anima.set_weight_adapter(adapter);
-    }
-
-    int64_t get_adm_in_channels() override {
-        return 768;
-    }
-
-    void set_flash_attention_enabled(bool enabled) {
-        anima.set_flash_attention_enabled(enabled);
-    }
-
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        anima.set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
-    void set_circular_axes(bool circular_x, bool circular_y) override {
-        anima.set_circular_axes(circular_x, circular_y);
-    }
-
-    sd::Tensor<float> compute(int n_threads,
-                              const DiffusionParams& diffusion_params) override {
-        GGML_ASSERT(diffusion_params.x != nullptr);
-        GGML_ASSERT(diffusion_params.timesteps != nullptr);
-        return anima.compute(n_threads,
-                             *diffusion_params.x,
-                             *diffusion_params.timesteps,
-                             tensor_or_empty(diffusion_params.context),
-                             tensor_or_empty(diffusion_params.t5_ids),
-                             tensor_or_empty(diffusion_params.t5_weights));
-    }
-};
-
-struct WanModel : public DiffusionModel {
-    std::string prefix;
-    WAN::WanRunner wan;
-
-    WanModel(ggml_backend_t backend,
-             bool offload_params_to_cpu,
-             const String2TensorStorage& tensor_storage_map = {},
-             const std::string prefix                       = "model.diffusion_model",
-             SDVersion version                              = VERSION_WAN2)
-        : prefix(prefix), wan(backend, offload_params_to_cpu, tensor_storage_map, prefix, version) {
-    }
-
-    std::string get_desc() override {
-        return wan.get_desc();
-    }
-
-    void alloc_params_buffer() override {
-        wan.alloc_params_buffer();
-    }
-
-    void free_params_buffer() override {
-        wan.free_params_buffer();
-    }
-
-    void free_compute_buffer() override {
-        wan.free_compute_buffer();
-    }
-
-    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
-        wan.get_param_tensors(tensors, prefix);
-    }
-
-    size_t get_params_buffer_size() override {
-        return wan.get_params_buffer_size();
-    }
-
-    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
-        wan.set_weight_adapter(adapter);
-    }
-
-    int64_t get_adm_in_channels() override {
-        return 768;
-    }
-
-    void set_flash_attention_enabled(bool enabled) {
-        wan.set_flash_attention_enabled(enabled);
-    }
-
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        wan.set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
-    void set_circular_axes(bool circular_x, bool circular_y) override {
-        wan.set_circular_axes(circular_x, circular_y);
-    }
-
-    sd::Tensor<float> compute(int n_threads,
-                              const DiffusionParams& diffusion_params) override {
-        GGML_ASSERT(diffusion_params.x != nullptr);
-        GGML_ASSERT(diffusion_params.timesteps != nullptr);
-        return wan.compute(n_threads,
-                           *diffusion_params.x,
-                           *diffusion_params.timesteps,
-                           tensor_or_empty(diffusion_params.context),
-                           tensor_or_empty(diffusion_params.y),
-                           tensor_or_empty(diffusion_params.c_concat),
-                           sd::Tensor<float>(),
-                           tensor_or_empty(diffusion_params.vace_context),
-                           diffusion_params.vace_strength);
-    }
-};
-
-struct QwenImageModel : public DiffusionModel {
-    std::string prefix;
-    Qwen::QwenImageRunner qwen_image;
-
-    QwenImageModel(ggml_backend_t backend,
-                   bool offload_params_to_cpu,
-                   const String2TensorStorage& tensor_storage_map = {},
-                   const std::string prefix                       = "model.diffusion_model",
-                   SDVersion version                              = VERSION_QWEN_IMAGE,
-                   bool zero_cond_t                               = false)
-        : prefix(prefix), qwen_image(backend, offload_params_to_cpu, tensor_storage_map, prefix, version, zero_cond_t) {
-    }
-
-    std::string get_desc() override {
-        return qwen_image.get_desc();
-    }
-
-    void alloc_params_buffer() override {
-        qwen_image.alloc_params_buffer();
-    }
-
-    void free_params_buffer() override {
-        qwen_image.free_params_buffer();
-    }
-
-    void free_compute_buffer() override {
-        qwen_image.free_compute_buffer();
-    }
-
-    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
-        qwen_image.get_param_tensors(tensors, prefix);
-    }
-
-    size_t get_params_buffer_size() override {
-        return qwen_image.get_params_buffer_size();
-    }
-
-    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
-        qwen_image.set_weight_adapter(adapter);
-    }
-
-    int64_t get_adm_in_channels() override {
-        return 768;
-    }
-
-    void set_flash_attention_enabled(bool enabled) {
-        qwen_image.set_flash_attention_enabled(enabled);
-    }
-
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        qwen_image.set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
-    void set_circular_axes(bool circular_x, bool circular_y) override {
-        qwen_image.set_circular_axes(circular_x, circular_y);
-    }
-
-    sd::Tensor<float> compute(int n_threads,
-                              const DiffusionParams& diffusion_params) override {
-        GGML_ASSERT(diffusion_params.x != nullptr);
-        GGML_ASSERT(diffusion_params.timesteps != nullptr);
-        static const std::vector<sd::Tensor<float>> empty_ref_latents;
-        return qwen_image.compute(n_threads,
-                                  *diffusion_params.x,
-                                  *diffusion_params.timesteps,
-                                  tensor_or_empty(diffusion_params.context),
-                                  diffusion_params.ref_latents ? *diffusion_params.ref_latents : empty_ref_latents,
-                                  true);
-    }
-};
-
-struct ZImageModel : public DiffusionModel {
-    std::string prefix;
-    ZImage::ZImageRunner z_image;
-
-    ZImageModel(ggml_backend_t backend,
-                bool offload_params_to_cpu,
-                const String2TensorStorage& tensor_storage_map = {},
-                const std::string prefix                       = "model.diffusion_model",
-                SDVersion version                              = VERSION_Z_IMAGE)
-        : prefix(prefix), z_image(backend, offload_params_to_cpu, tensor_storage_map, prefix, version) {
-    }
-
-    std::string get_desc() override {
-        return z_image.get_desc();
-    }
-
-    void alloc_params_buffer() override {
-        z_image.alloc_params_buffer();
-    }
-
-    void free_params_buffer() override {
-        z_image.free_params_buffer();
-    }
-
-    void free_compute_buffer() override {
-        z_image.free_compute_buffer();
-    }
-
-    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
-        z_image.get_param_tensors(tensors, prefix);
-    }
-
-    size_t get_params_buffer_size() override {
-        return z_image.get_params_buffer_size();
-    }
-
-    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
-        z_image.set_weight_adapter(adapter);
-    }
-
-    int64_t get_adm_in_channels() override {
-        return 768;
-    }
-
-    void set_flash_attention_enabled(bool enabled) {
-        z_image.set_flash_attention_enabled(enabled);
-    }
-
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        z_image.set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
-    void set_circular_axes(bool circular_x, bool circular_y) override {
-        z_image.set_circular_axes(circular_x, circular_y);
-    }
-
-    sd::Tensor<float> compute(int n_threads,
-                              const DiffusionParams& diffusion_params) override {
-        GGML_ASSERT(diffusion_params.x != nullptr);
-        GGML_ASSERT(diffusion_params.timesteps != nullptr);
-        static const std::vector<sd::Tensor<float>> empty_ref_latents;
-        return z_image.compute(n_threads,
-                               *diffusion_params.x,
-                               *diffusion_params.timesteps,
-                               tensor_or_empty(diffusion_params.context),
-                               diffusion_params.ref_latents ? *diffusion_params.ref_latents : empty_ref_latents,
-                               true);
-    }
-};
-
-struct ErnieImageModel : public DiffusionModel {
-    std::string prefix;
-    ErnieImage::ErnieImageRunner ernie_image;
-
-    ErnieImageModel(ggml_backend_t backend,
-                    bool offload_params_to_cpu,
-                    const String2TensorStorage& tensor_storage_map = {},
-                    const std::string prefix                       = "model.diffusion_model")
-        : prefix(prefix), ernie_image(backend, offload_params_to_cpu, tensor_storage_map, prefix) {
-    }
-
-    std::string get_desc() override {
-        return ernie_image.get_desc();
-    }
-
-    void alloc_params_buffer() override {
-        ernie_image.alloc_params_buffer();
-    }
-
-    void free_params_buffer() override {
-        ernie_image.free_params_buffer();
-    }
-
-    void free_compute_buffer() override {
-        ernie_image.free_compute_buffer();
-    }
-
-    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
-        ernie_image.get_param_tensors(tensors, prefix);
-    }
-
-    size_t get_params_buffer_size() override {
-        return ernie_image.get_params_buffer_size();
-    }
-
-    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
-        ernie_image.set_weight_adapter(adapter);
-    }
-
-    int64_t get_adm_in_channels() override {
-        return 768;
-    }
-
-    void set_flash_attention_enabled(bool enabled) {
-        ernie_image.set_flash_attention_enabled(enabled);
-    }
-
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        ernie_image.set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
-    void set_circular_axes(bool circular_x, bool circular_y) override {
-        ernie_image.set_circular_axes(circular_x, circular_y);
-    }
-
-    sd::Tensor<float> compute(int n_threads,
-                              const DiffusionParams& diffusion_params) override {
-        GGML_ASSERT(diffusion_params.x != nullptr);
-        GGML_ASSERT(diffusion_params.timesteps != nullptr);
-        return ernie_image.compute(n_threads,
-                                   *diffusion_params.x,
-                                   *diffusion_params.timesteps,
-                                   tensor_or_empty(diffusion_params.context));
-    }
+    virtual void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors,
+                                   const std::string& prefix) = 0;
 };

 #endif
@@ -5,6 +5,7 @@
 #include <vector>

 #include "common_dit.hpp"
+#include "diffusion_model.hpp"
 #include "flux.hpp"
 #include "qwen_image.hpp"
 #include "rope.hpp"
@@ -325,16 +326,16 @@ namespace ErnieImage {
        }
    };

-    struct ErnieImageRunner : public GGMLRunner {
+    struct ErnieImageRunner : public DiffusionModelRunner {
        ErnieImageParams ernie_params;
        ErnieImageModel ernie_image;
        std::vector<float> pe_vec;

        ErnieImageRunner(ggml_backend_t backend,
-                         bool offload_params_to_cpu,
+                         ggml_backend_t params_backend,
                         const String2TensorStorage& tensor_storage_map = {},
                         const std::string prefix                       = "")
-            : GGMLRunner(backend, offload_params_to_cpu) {
+            : DiffusionModelRunner(backend, params_backend, prefix) {
            ernie_params.num_layers = 0;
            for (const auto& [name, tensor_storage] : tensor_storage_map) {
                if (!starts_with(name, prefix)) {
@@ -393,7 +394,7 @@ namespace ErnieImage {
            return "ernie_image";
        }

-        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
            ernie_image.get_param_tensors(tensors, prefix);
        }

@@ -435,6 +436,16 @@ namespace ErnieImage {
            };
            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), x.dim());
        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const DiffusionParams& diffusion_params) override {
+            GGML_ASSERT(diffusion_params.x != nullptr);
+            GGML_ASSERT(diffusion_params.timesteps != nullptr);
+            return compute(n_threads,
+                           *diffusion_params.x,
+                           *diffusion_params.timesteps,
+                           tensor_or_empty(diffusion_params.context));
+        }
    };
 }  // namespace ErnieImage

@@ -161,10 +161,10 @@ struct ESRGAN : public GGMLRunner {
    int tile_size = 128;  // avoid cuda OOM for 4gb VRAM

    ESRGAN(ggml_backend_t backend,
-           bool offload_params_to_cpu,
+           ggml_backend_t params_backend,
           int tile_size                                  = 128,
           const String2TensorStorage& tensor_storage_map = {})
-        : GGMLRunner(backend, offload_params_to_cpu) {
+        : GGMLRunner(backend, params_backend) {
        this->tile_size = tile_size;
    }

@@ -270,7 +270,11 @@ struct ESRGAN : public GGMLRunner {
        rrdb_net = std::make_unique<RRDBNet>(detected_scale, detected_num_block, detected_num_in_ch, detected_num_out_ch, detected_num_feat, detected_num_grow_ch);
        rrdb_net->init(params_ctx, {}, "");

-        alloc_params_buffer();
+        if (!alloc_params_buffer()) {
+            LOG_ERROR("esrgan model buffer allocation failed");
+            return false;
+        }
+
        std::map<std::string, ggml_tensor*> esrgan_tensors;
        rrdb_net->get_param_tensors(esrgan_tensors);

@@ -5,6 +5,7 @@
 #include <vector>

 #include "common_dit.hpp"
+#include "diffusion_model.hpp"
 #include "model.h"
 #include "rope.hpp"

@@ -446,7 +447,6 @@ namespace Flux {
            if (use_yak_mlp || use_mlp_silu_act) {
                mlp_mult_factor = 2;
            }
-
            blocks["linear1"]  = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, hidden_size * 3 + mlp_hidden_dim * mlp_mult_factor, mlp_proj_bias));
            blocks["linear2"]  = std::shared_ptr<GGMLBlock>(new Linear(hidden_size + mlp_hidden_dim, hidden_size, mlp_proj_bias));
            blocks["norm"]     = std::shared_ptr<GGMLBlock>(new QKNorm(head_dim));
@@ -1177,7 +1177,7 @@ namespace Flux {
        }
    };

-    struct FluxRunner : public GGMLRunner {
+    struct FluxRunner : public DiffusionModelRunner {
    public:
        FluxParams flux_params;
        Flux flux;
@@ -1189,12 +1189,12 @@ namespace Flux {
        bool use_mask = false;

        FluxRunner(ggml_backend_t backend,
-                   bool offload_params_to_cpu,
+                   ggml_backend_t params_backend,
                   const String2TensorStorage& tensor_storage_map = {},
                   const std::string prefix                       = "",
                   SDVersion version                              = VERSION_FLUX,
                   bool use_mask                                  = false)
-            : GGMLRunner(backend, offload_params_to_cpu), version(version), use_mask(use_mask) {
+            : DiffusionModelRunner(backend, params_backend, prefix), version(version), use_mask(use_mask) {
            flux_params.version             = version;
            flux_params.guidance_embed      = false;
            flux_params.depth               = 0;
@@ -1225,6 +1225,9 @@ namespace Flux {
                flux_params.share_modulation = true;
                flux_params.ref_index_scale  = 10.f;
                flux_params.use_mlp_silu_act = true;
+            } else if (sd_version_is_longcat(version)) {
+                flux_params.context_in_dim = 3584;
+                flux_params.vec_in_dim     = 0;
            }
            int64_t head_dim                   = 0;
            int64_t actual_radiance_patch_size = -1;
@@ -1306,7 +1309,7 @@ namespace Flux {
            return "flux";
        }

-        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
            flux.get_param_tensors(tensors, prefix);
        }

@@ -1412,7 +1415,6 @@ namespace Flux {
            } else if (version == VERSION_OVIS_IMAGE) {
                txt_arange_dims = {1, 2};
            }
-
            pe_vec      = Rope::gen_flux_pe(static_cast<int>(x->ne[1]),
                                            static_cast<int>(x->ne[0]),
                                            flux_params.patch_size,
@@ -1425,7 +1427,8 @@ namespace Flux {
                                            flux_params.theta,
                                            circular_y_enabled,
                                            circular_x_enabled,
-                                            flux_params.axes_dim);
+                                            flux_params.axes_dim,
+                                            sd_version_is_longcat(version));
            int pos_len = static_cast<int>(pe_vec.size() / flux_params.axes_dim_sum / 2);
            // LOG_DEBUG("pos_len %d", pos_len);
            auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, flux_params.axes_dim_sum / 2, pos_len);
@@ -1488,6 +1491,25 @@ namespace Flux {
            return result;
        }

+        sd::Tensor<float> compute(int n_threads,
+                                  const DiffusionParams& diffusion_params) override {
+            GGML_ASSERT(diffusion_params.x != nullptr);
+            GGML_ASSERT(diffusion_params.timesteps != nullptr);
+            const auto* extra = diffusion_extra_as<FluxDiffusionExtra>(diffusion_params);
+            static const std::vector<sd::Tensor<float>> empty_ref_latents;
+            static const std::vector<int> empty_skip_layers;
+            return compute(n_threads,
+                           *diffusion_params.x,
+                           *diffusion_params.timesteps,
+                           tensor_or_empty(diffusion_params.context),
+                           tensor_or_empty(diffusion_params.c_concat),
+                           tensor_or_empty(diffusion_params.y),
+                           tensor_or_empty(extra->guidance),
+                           diffusion_params.ref_latents ? *diffusion_params.ref_latents : empty_ref_latents,
+                           diffusion_params.increase_ref_index,
+                           extra->skip_layers ? *extra->skip_layers : empty_skip_layers);
+        }
+
        void test() {
            ggml_init_params params;
            params.mem_size   = static_cast<size_t>(1024 * 1024) * 1024;  // 1GB
@@ -1545,7 +1567,7 @@ namespace Flux {

        static void load_from_file_and_test(const std::string& file_path) {
            // ggml_backend_t backend = ggml_backend_cuda_init(0);
-            ggml_backend_t backend    = ggml_backend_cpu_init();
+            ggml_backend_t backend    = sd_backend_cpu_init();
            ggml_type model_data_type = GGML_TYPE_COUNT;

            ModelLoader model_loader;
@@ -1564,13 +1586,17 @@ namespace Flux {
            }

            std::shared_ptr<FluxRunner> flux = std::make_shared<FluxRunner>(backend,
-                                                                            false,
+                                                                            backend,
                                                                            tensor_storage_map,
                                                                            "model.diffusion_model",
                                                                            VERSION_FLUX2,
                                                                            false);

-            flux->alloc_params_buffer();
+            if (!flux->alloc_params_buffer()) {
+                LOG_ERROR("flux model allocation failed");
+                return;
+            }
+
            std::map<std::string, ggml_tensor*> tensors;
            flux->get_param_tensors(tensors, "model.diffusion_model");

@@ -0,0 +1,656 @@
+#include "ggml_extend_backend.h"
+
+#include <algorithm>
+#include <cctype>
+#include <cstdlib>
+#include <mutex>
+#include <sstream>
+#include <stdexcept>
+#include <vector>
+
+#include "stable-diffusion.h"
+#include "util.h"
+
+static std::string trim_copy(const std::string& value) {
+    size_t begin = 0;
+    while (begin < value.size() && std::isspace(static_cast<unsigned char>(value[begin]))) {
+        ++begin;
+    }
+    size_t end = value.size();
+    while (end > begin && std::isspace(static_cast<unsigned char>(value[end - 1]))) {
+        --end;
+    }
+    return value.substr(begin, end - begin);
+}
+
+static std::string lower_copy(std::string value) {
+    std::transform(value.begin(), value.end(), value.begin(), [](unsigned char c) {
+        return static_cast<char>(std::tolower(c));
+    });
+    return value;
+}
+
+static std::vector<std::string> split_copy(const std::string& value, char delimiter) {
+    std::vector<std::string> parts;
+    std::string part;
+    std::istringstream stream(value);
+    while (std::getline(stream, part, delimiter)) {
+        parts.push_back(part);
+    }
+    return parts;
+}
+
+static bool is_default_backend_token(const std::string& name) {
+    const std::string lower = lower_copy(trim_copy(name));
+    return lower.empty() || lower == "default" || lower == "auto";
+}
+
+static bool parse_backend_module(const std::string& raw_name, SDBackendModule* module) {
+    std::string name = lower_copy(trim_copy(raw_name));
+    name.erase(std::remove(name.begin(), name.end(), '-'), name.end());
+    name.erase(std::remove(name.begin(), name.end(), '_'), name.end());
+
+    if (name == "diffusion" || name == "model" || name == "unet" || name == "dit") {
+        *module = SDBackendModule::DIFFUSION;
+        return true;
+    }
+    if (name == "te" || name == "clip" || name == "text" || name == "textencoder" || name == "textencoders" || name == "conditioner" || name == "cond" || name == "llm" || name == "t5" || name == "t5xxl") {
+        *module = SDBackendModule::TE;
+        return true;
+    }
+    if (name == "clipvision" || name == "vision") {
+        *module = SDBackendModule::CLIP_VISION;
+        return true;
+    }
+    if (name == "vae" || name == "firststage" || name == "autoencoder" || name == "tae") {
+        *module = SDBackendModule::VAE;
+        return true;
+    }
+    if (name == "controlnet" || name == "control") {
+        *module = SDBackendModule::CONTROL_NET;
+        return true;
+    }
+    if (name == "photomaker" || name == "photomakerid" || name == "pmid" || name == "photo") {
+        *module = SDBackendModule::PHOTOMAKER;
+        return true;
+    }
+    if (name == "upscaler" || name == "esrgan" || name == "hires") {
+        *module = SDBackendModule::UPSCALER;
+        return true;
+    }
+    return false;
+}
+
+static std::string module_assignment_name(const SDBackendAssignment& assignment, SDBackendModule module) {
+    auto it = assignment.module_names.find(module);
+    if (it != assignment.module_names.end()) {
+        return it->second;
+    }
+    return assignment.default_name;
+}
+
+static std::string backend_cache_key(ggml_backend_t backend) {
+    if (backend == nullptr) {
+        return "";
+    }
+    ggml_backend_dev_t dev = ggml_backend_get_device(backend);
+    if (dev != nullptr) {
+        return lower_copy(ggml_backend_dev_name(dev));
+    }
+    const char* backend_name = ggml_backend_name(backend);
+    return backend_name != nullptr ? lower_copy(backend_name) : "";
+}
+
+static std::string resolve_first_device_by_type(enum ggml_backend_dev_type type) {
+    ggml_backend_dev_t dev = ggml_backend_dev_by_type(type);
+    if (dev == nullptr) {
+        return "";
+    }
+    return ggml_backend_dev_name(dev);
+}
+
+static ggml_backend_buffer_t ggml_backend_tensor_buffer(const struct ggml_tensor* tensor) {
+    if (tensor == nullptr) {
+        return nullptr;
+    }
+
+    return tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+}
+
+static bool ggml_backend_tensor_is_host_accessible(const struct ggml_tensor* tensor) {
+    if (tensor == nullptr || tensor->data == nullptr) {
+        return false;
+    }
+
+    ggml_backend_buffer_t buffer = ggml_backend_tensor_buffer(tensor);
+    return buffer == nullptr || ggml_backend_buffer_is_host(buffer);
+}
+
+static size_t ggml_backend_tensor_offset(const struct ggml_tensor* tensor, int64_t i0, int64_t i1, int64_t i2, int64_t i3) {
+    return static_cast<size_t>(i0 * tensor->nb[0] + i1 * tensor->nb[1] + i2 * tensor->nb[2] + i3 * tensor->nb[3]);
+}
+
+template <typename T>
+static void ggml_backend_tensor_write_scalar(const struct ggml_tensor* tensor, int64_t i0, int64_t i1, int64_t i2, int64_t i3, T value) {
+    const size_t offset = ggml_backend_tensor_offset(tensor, i0, i1, i2, i3);
+
+    if (ggml_backend_tensor_is_host_accessible(tensor)) {
+        auto* dst = reinterpret_cast<T*>(reinterpret_cast<char*>(tensor->data) + offset);
+        *dst      = value;
+        return;
+    }
+
+    ggml_backend_tensor_set(const_cast<struct ggml_tensor*>(tensor), &value, offset, sizeof(T));
+}
+
+static void ggml_set_f32_nd(const struct ggml_tensor* tensor, int64_t i0, int64_t i1, int64_t i2, int64_t i3, float value) {
+    switch (tensor->type) {
+        case GGML_TYPE_I8:
+            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, static_cast<int8_t>(value));
+            break;
+        case GGML_TYPE_I16:
+            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, static_cast<int16_t>(value));
+            break;
+        case GGML_TYPE_I32:
+            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, static_cast<int32_t>(value));
+            break;
+        case GGML_TYPE_F16:
+            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, ggml_fp32_to_fp16(value));
+            break;
+        case GGML_TYPE_BF16:
+            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, ggml_fp32_to_bf16(value));
+            break;
+        case GGML_TYPE_F32:
+            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, value);
+            break;
+        default:
+            GGML_ABORT("fatal error");
+    }
+}
+
+void ggml_ext_im_set_f32_1d(const struct ggml_tensor* tensor, int i, float value) {
+    if (!ggml_is_contiguous(tensor)) {
+        int64_t id[4] = {0, 0, 0, 0};
+        ggml_unravel_index(tensor, i, &id[0], &id[1], &id[2], &id[3]);
+        ggml_set_f32_nd(tensor, id[0], id[1], id[2], id[3], value);
+        return;
+    }
+
+    switch (tensor->type) {
+        case GGML_TYPE_I8:
+            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, static_cast<int8_t>(value));
+            break;
+        case GGML_TYPE_I16:
+            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, static_cast<int16_t>(value));
+            break;
+        case GGML_TYPE_I32:
+            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, static_cast<int32_t>(value));
+            break;
+        case GGML_TYPE_F16:
+            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, ggml_fp32_to_fp16(value));
+            break;
+        case GGML_TYPE_BF16:
+            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, ggml_fp32_to_bf16(value));
+            break;
+        case GGML_TYPE_F32:
+            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, value);
+            break;
+        default:
+            GGML_ABORT("fatal error");
+    }
+}
+
+static void ggml_backend_load_all_once() {
+    // If the registry already has devices and the CPU backend is present,
+    // assume either static registration or explicit host-side preloading has
+    // completed and avoid rescanning the default paths.
+    if (ggml_backend_dev_count() > 0 && ggml_backend_reg_by_name("CPU") != nullptr) {
+        return;
+    }
+    // In dynamic-backend mode the backend modules are discovered at runtime,
+    // so we must load them before asking for the CPU backend or its proc table.
+    // If the host preloaded only a subset of backends, allow one default-path
+    // scan so missing modules can still be discovered.
+    static std::once_flag once;
+    std::call_once(once, []() {
+        if (ggml_backend_dev_count() > 0 && ggml_backend_reg_by_name("CPU") != nullptr) {
+            return;
+        }
+        ggml_backend_load_all();
+    });
+}
+
+bool sd_backend_is(ggml_backend_t backend, const std::string& name) {
+    if (!backend) {
+        return false;
+    }
+    ggml_backend_dev_t dev = ggml_backend_get_device(backend);
+    if (!dev) {
+        return false;
+    }
+    std::string dev_name = ggml_backend_dev_name(dev);
+    return lower_copy(dev_name).find(lower_copy(name)) != std::string::npos;
+}
+
+static std::string get_default_backend_name() {
+    ggml_backend_load_all_once();
+    // should pick the same backend preference as ggml_backend_init_best
+    std::string name = resolve_first_device_by_type(GGML_BACKEND_DEVICE_TYPE_GPU);
+    if (!name.empty()) {
+        return name;
+    }
+    name = resolve_first_device_by_type(GGML_BACKEND_DEVICE_TYPE_IGPU);
+    if (!name.empty()) {
+        return name;
+    }
+    return resolve_first_device_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+}
+
+static std::string sd_resolve_backend_name(const std::string& name) {
+    ggml_backend_load_all_once();
+    std::string requested = trim_copy(name);
+    std::string lower     = lower_copy(requested);
+
+    if (is_default_backend_token(lower)) {
+        return get_default_backend_name();
+    }
+    if (lower == "gpu") {
+        std::string result = resolve_first_device_by_type(GGML_BACKEND_DEVICE_TYPE_GPU);
+        if (!result.empty()) {
+            return result;
+        }
+        return resolve_first_device_by_type(GGML_BACKEND_DEVICE_TYPE_IGPU);
+    }
+
+    const size_t device_count = ggml_backend_dev_count();
+    for (size_t i = 0; i < device_count; ++i) {
+        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+        std::string dev_name   = ggml_backend_dev_name(dev);
+        if (lower_copy(dev_name) == lower) {
+            return dev_name;
+        }
+    }
+
+    for (size_t i = 0; i < device_count; ++i) {
+        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+        std::string dev_name   = ggml_backend_dev_name(dev);
+        std::string dev_lower  = lower_copy(dev_name);
+        if (dev_lower.rfind(lower, 0) == 0) {
+            return dev_name;
+        }
+    }
+
+    return "";
+}
+
+static bool backend_name_exists(const std::string& name) {
+    return !sd_resolve_backend_name(name).empty();
+}
+
+static ggml_backend_t init_named_backend(const std::string& name) {
+    ggml_backend_load_all_once();
+    LOG_DEBUG("Initializing backend: %s", name.c_str());
+    if (trim_copy(name).empty()) {
+        return ggml_backend_init_best();
+    }
+
+    std::string resolved = sd_resolve_backend_name(name);
+    if (resolved.empty()) {
+        return nullptr;
+    }
+    return ggml_backend_init_by_name(resolved.c_str(), nullptr);
+}
+
+bool sd_backend_is_cpu(ggml_backend_t backend) {
+    if (backend == nullptr) {
+        return false;
+    }
+    auto dev = ggml_backend_get_device(backend);
+    return dev != nullptr && ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_CPU;
+}
+
+ggml_backend_t sd_backend_cpu_init() {
+    ggml_backend_load_all_once();
+    return ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
+}
+
+bool sd_backend_cpu_set_n_threads(ggml_backend_t backend, int n_threads) {
+    if (backend == nullptr) {
+        return false;
+    }
+    auto dev = ggml_backend_get_device(backend);
+    if (dev != nullptr && ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_CPU) {
+        auto reg                           = ggml_backend_dev_backend_reg(dev);
+        auto ggml_backend_set_n_threads_fn = (ggml_backend_set_n_threads_t)ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_n_threads");
+        if (ggml_backend_set_n_threads_fn != nullptr) {
+            ggml_backend_set_n_threads_fn(backend, n_threads);
+            return true;
+        }
+    }
+    return false;
+}
+
+const char* sd_get_system_info() {
+    static std::string cache_info = []() -> std::string {
+        ggml_backend_load_all_once();
+        std::stringstream ss;
+        ss << "System Info: \n";
+        auto dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+        if (dev != nullptr) {
+            auto reg                          = ggml_backend_dev_backend_reg(dev);
+            auto ggml_backend_get_features_fn = (ggml_backend_get_features_t)ggml_backend_reg_get_proc_address(reg, "ggml_backend_get_features");
+            if (ggml_backend_get_features_fn != nullptr) {
+                ggml_backend_feature* feat = ggml_backend_get_features_fn(reg);
+                while (feat->name && feat->value) {
+                    ss << "   " << feat->name << " = " << feat->value << " | ";
+                    feat++;
+                }
+            } else {
+                LOG_WARN("unable to get CPU features");
+            }
+        } else {
+            LOG_WARN("unable to get CPU features");
+        }
+        return ss.str();
+    }();
+    return cache_info.c_str();
+}
+
+static ggml_backend_t sd_get_default_backend() {
+    ggml_backend_load_all_once();
+    static std::once_flag once;
+    std::call_once(once, []() {
+        size_t dev_count = ggml_backend_dev_count();
+        if (dev_count == 0) {
+            LOG_ERROR("No devices found!");
+        } else {
+            LOG_DEBUG("Found %zu backend devices:", dev_count);
+            for (size_t i = 0; i < dev_count; ++i) {
+                auto dev = ggml_backend_dev_get(i);
+                LOG_DEBUG("#%zu: %s", i, ggml_backend_dev_name(dev));
+            }
+        }
+    });
+
+    ggml_backend_t backend   = nullptr;
+    const char* SD_VK_DEVICE = getenv("SD_VK_DEVICE");
+    if (SD_VK_DEVICE != nullptr) {
+        std::string sd_vk_device_str = SD_VK_DEVICE;
+        try {
+            unsigned long long device  = std::stoull(sd_vk_device_str);
+            std::string vk_device_name = "Vulkan" + std::to_string(device);
+            if (backend_name_exists(vk_device_name)) {
+                LOG_INFO("Selecting %s as main device by env var SD_VK_DEVICE", vk_device_name.c_str());
+                backend = init_named_backend(vk_device_name);
+                if (!backend) {
+                    LOG_WARN("Device %s requested by SD_VK_DEVICE failed to init. Falling back to the default device.", vk_device_name.c_str());
+                }
+            } else {
+                LOG_WARN("Device %s requested by SD_VK_DEVICE was not found. Falling back to the default device.", vk_device_name.c_str());
+            }
+        } catch (const std::invalid_argument&) {
+            LOG_WARN("SD_VK_DEVICE environment variable is not a valid integer (%s). Falling back to the default device.", SD_VK_DEVICE);
+        } catch (const std::out_of_range&) {
+            LOG_WARN("SD_VK_DEVICE environment variable value is out of range for `unsigned long long` type (%s). Falling back to the default device.", SD_VK_DEVICE);
+        }
+    }
+
+    if (!backend) {
+        std::string dev_name = get_default_backend_name();
+        backend              = init_named_backend(dev_name);
+        if (!backend && !dev_name.empty()) {
+            LOG_WARN("device %s failed to init", dev_name.c_str());
+        }
+    }
+
+    if (!backend) {
+        LOG_WARN("loading CPU backend");
+        backend = sd_backend_cpu_init();
+    }
+
+    if (sd_backend_is_cpu(backend)) {
+        LOG_DEBUG("Using CPU backend");
+    }
+
+    return backend;
+}
+
+static bool sd_parse_backend_assignment(const std::string& spec, SDBackendAssignment* assignment, std::string* error) {
+    if (assignment == nullptr) {
+        return false;
+    }
+
+    *assignment          = {};
+    const std::string in = trim_copy(spec);
+    if (in.empty()) {
+        return true;
+    }
+
+    for (const std::string& raw_part : split_copy(in, ',')) {
+        const std::string part = trim_copy(raw_part);
+        if (part.empty()) {
+            continue;
+        }
+
+        const size_t eq = part.find('=');
+        if (eq == std::string::npos) {
+            assignment->set_default(part);
+            continue;
+        }
+
+        const std::string key   = trim_copy(part.substr(0, eq));
+        const std::string value = trim_copy(part.substr(eq + 1));
+        if (key.empty() || value.empty()) {
+            if (error != nullptr) {
+                *error = "invalid backend assignment '" + part + "'";
+            }
+            return false;
+        }
+
+        const std::string key_lower = lower_copy(key);
+        if (key_lower == "all" || key_lower == "default" || key_lower == "*") {
+            assignment->set_default(value);
+            continue;
+        }
+
+        SDBackendModule module = SDBackendModule::DIFFUSION;
+        if (!parse_backend_module(key, &module)) {
+            if (error != nullptr) {
+                *error = "unknown backend module '" + key + "'";
+            }
+            return false;
+        }
+        assignment->set_module(module, value);
+    }
+    return true;
+}
+
+bool SDBackendAssignment::empty() const {
+    return default_name.empty() && module_names.empty();
+}
+
+std::string SDBackendAssignment::get(SDBackendModule module) const {
+    return module_assignment_name(*this, module);
+}
+
+void SDBackendAssignment::set_default(const std::string& name) {
+    default_name = trim_copy(name);
+}
+
+void SDBackendAssignment::set_module(SDBackendModule module, const std::string& name) {
+    module_names[module] = trim_copy(name);
+}
+
+void SDBackendHandleDeleter::operator()(ggml_backend_t backend) const {
+    ggml_backend_free(backend);
+}
+
+SDBackendManager::~SDBackendManager() {
+    reset();
+}
+
+void SDBackendManager::reset() {
+    backends_.clear();
+    runtime_assignment_ = {};
+    params_assignment_  = {};
+}
+
+ggml_backend_t SDBackendManager::runtime_backend(SDBackendModule module) {
+    return init_cached_backend(runtime_assignment_.get(module));
+}
+
+ggml_backend_t SDBackendManager::params_backend(SDBackendModule module) {
+    std::string name = params_assignment_.get(module);
+    if (name.empty()) {
+        return runtime_backend(module);
+    }
+    return init_cached_backend(name);
+}
+
+bool SDBackendManager::runtime_backend_is_cpu(SDBackendModule module) {
+    return sd_backend_is_cpu(runtime_backend(module));
+}
+
+bool SDBackendManager::params_backend_is_cpu(SDBackendModule module) {
+    return sd_backend_is_cpu(params_backend(module));
+}
+
+bool SDBackendManager::runtime_backend_supports_host_buffer(SDBackendModule module) {
+    ggml_backend_t backend = runtime_backend(module);
+    if (backend == nullptr) {
+        return false;
+    }
+    if (sd_backend_is_cpu(backend)) {
+        return true;
+    }
+    ggml_backend_dev_t dev = ggml_backend_get_device(backend);
+    if (dev == nullptr) {
+        return false;
+    }
+    ggml_backend_dev_props props;
+    ggml_backend_dev_get_props(dev, &props);
+    return props.caps.buffer_from_host_ptr;
+}
+
+bool SDBackendManager::init(const char* backend_spec,
+                            const char* params_backend_spec,
+                            bool offload_params_to_cpu,
+                            bool keep_clip_on_cpu,
+                            bool keep_vae_on_cpu,
+                            bool keep_control_net_on_cpu,
+                            std::string* error) {
+    reset();
+
+    if (!sd_parse_backend_assignment(SAFE_STR(backend_spec), &runtime_assignment_, error)) {
+        return false;
+    }
+    if (!sd_parse_backend_assignment(SAFE_STR(params_backend_spec), &params_assignment_, error)) {
+        return false;
+    }
+
+    if (runtime_assignment_.empty()) {
+        if (keep_clip_on_cpu) {
+            runtime_assignment_.set_module(SDBackendModule::TE, "cpu");
+        }
+        if (keep_vae_on_cpu) {
+            runtime_assignment_.set_module(SDBackendModule::VAE, "cpu");
+        }
+        if (keep_control_net_on_cpu) {
+            runtime_assignment_.set_module(SDBackendModule::CONTROL_NET, "cpu");
+        }
+    }
+
+    if (params_assignment_.empty() && offload_params_to_cpu) {
+        params_assignment_.set_default("cpu");
+    }
+
+    return validate(error);
+}
+
+bool SDBackendManager::validate(std::string* error) const {
+    auto validate_name = [&](const std::string& name) -> bool {
+        if (is_default_backend_token(name)) {
+            return true;
+        }
+        if (!sd_resolve_backend_name(name).empty()) {
+            return true;
+        }
+        if (error != nullptr) {
+            *error = "backend '" + name + "' was not found";
+        }
+        return false;
+    };
+
+    if (!validate_name(runtime_assignment_.default_name) ||
+        !validate_name(params_assignment_.default_name)) {
+        return false;
+    }
+    for (const auto& kv : runtime_assignment_.module_names) {
+        if (!validate_name(kv.second)) {
+            return false;
+        }
+    }
+    for (const auto& kv : params_assignment_.module_names) {
+        if (!validate_name(kv.second)) {
+            return false;
+        }
+    }
+    return true;
+}
+
+ggml_backend_t SDBackendManager::init_cached_backend(const std::string& name) {
+    std::string resolved   = sd_resolve_backend_name(name);
+    std::string key        = lower_copy(resolved);
+    ggml_backend_t backend = nullptr;
+
+    if (!key.empty()) {
+        auto it = backends_.find(key);
+        if (it != backends_.end()) {
+            return it->second.get();
+        }
+    } else if (!is_default_backend_token(name)) {
+        LOG_ERROR("backend '%s' was not found", name.c_str());
+        return nullptr;
+    }
+
+    backend = is_default_backend_token(name) ? sd_get_default_backend() : init_named_backend(resolved);
+    if (backend == nullptr) {
+        LOG_ERROR("failed to initialize backend '%s'", name.c_str());
+        return nullptr;
+    }
+
+    std::string actual_key = backend_cache_key(backend);
+    if (actual_key.empty()) {
+        actual_key = !key.empty() ? key : lower_copy(trim_copy(name));
+    }
+
+    auto it = backends_.find(actual_key);
+    if (it != backends_.end()) {
+        ggml_backend_free(backend);
+        return it->second.get();
+    }
+
+    SDBackendHandle handle(backend);
+    backends_.emplace(actual_key, std::move(handle));
+    return backend;
+}
+
+const char* sd_backend_module_name(SDBackendModule module) {
+    switch (module) {
+        case SDBackendModule::DIFFUSION:
+            return "diffusion";
+        case SDBackendModule::TE:
+            return "te";
+        case SDBackendModule::CLIP_VISION:
+            return "clip_vision";
+        case SDBackendModule::VAE:
+            return "vae";
+        case SDBackendModule::CONTROL_NET:
+            return "controlnet";
+        case SDBackendModule::PHOTOMAKER:
+            return "photomaker";
+        case SDBackendModule::UPSCALER:
+            return "upscaler";
+    }
+    return "unknown";
+}
@@ -0,0 +1,79 @@
+#ifndef __SD_GGML_EXTEND_BACKEND_H__
+#define __SD_GGML_EXTEND_BACKEND_H__
+
+#include <cstdint>
+#include <cstring>
+#include <memory>
+#include <string>
+#include <unordered_map>
+
+#include "ggml-backend.h"
+#include "ggml.h"
+
+enum class SDBackendModule {
+    DIFFUSION,
+    TE,
+    CLIP_VISION,
+    VAE,
+    CONTROL_NET,
+    PHOTOMAKER,
+    UPSCALER,
+};
+
+struct SDBackendAssignment {
+    std::string default_name;
+    std::unordered_map<SDBackendModule, std::string> module_names;
+
+    bool empty() const;
+    std::string get(SDBackendModule module) const;
+    void set_default(const std::string& name);
+    void set_module(SDBackendModule module, const std::string& name);
+};
+
+struct SDBackendHandleDeleter {
+    void operator()(ggml_backend_t backend) const;
+};
+
+using SDBackendHandle = std::unique_ptr<struct ggml_backend, SDBackendHandleDeleter>;
+
+class SDBackendManager {
+private:
+    SDBackendAssignment runtime_assignment_;
+    SDBackendAssignment params_assignment_;
+    std::unordered_map<std::string, SDBackendHandle> backends_;
+
+public:
+    SDBackendManager() = default;
+    ~SDBackendManager();
+
+    SDBackendManager(const SDBackendManager&)            = delete;
+    SDBackendManager& operator=(const SDBackendManager&) = delete;
+
+    bool init(const char* backend_spec,
+              const char* params_backend_spec,
+              bool offload_params_to_cpu,
+              bool keep_clip_on_cpu,
+              bool keep_vae_on_cpu,
+              bool keep_control_net_on_cpu,
+              std::string* error);
+    void reset();
+
+    ggml_backend_t runtime_backend(SDBackendModule module);
+    ggml_backend_t params_backend(SDBackendModule module);
+
+    bool runtime_backend_is_cpu(SDBackendModule module);
+    bool params_backend_is_cpu(SDBackendModule module);
+    bool runtime_backend_supports_host_buffer(SDBackendModule module);
+
+private:
+    bool validate(std::string* error) const;
+    ggml_backend_t init_cached_backend(const std::string& name);
+};
+
+bool sd_backend_is(ggml_backend_t backend, const std::string& name);
+bool sd_backend_is_cpu(ggml_backend_t backend);
+ggml_backend_t sd_backend_cpu_init();
+bool sd_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads);
+const char* sd_backend_module_name(SDBackendModule module);
+void ggml_ext_im_set_f32_1d(const struct ggml_tensor* tensor, int i, float value);
+#endif
@@ -1,298 +0,0 @@
-#ifndef __GGML_EXTEND_BACKEND_HPP__
-#define __GGML_EXTEND_BACKEND_HPP__
-
-#include <cstring>
-#include <mutex>
-
-#include "ggml-backend.h"
-#include "ggml.h"
-
-#ifndef __STATIC_INLINE__
-#define __STATIC_INLINE__ static inline
-#endif
-
-inline void ggml_backend_load_all_once() {
-    // If the registry already has devices and the CPU backend is present,
-    // assume either static registration or explicit host-side preloading has
-    // completed and avoid rescanning the default paths.
-    if (ggml_backend_dev_count() > 0 && ggml_backend_reg_by_name("CPU") != nullptr) {
-        return;
-    }
-    // In dynamic-backend mode the backend modules are discovered at runtime,
-    // so we must load them before asking for the CPU backend or its proc table.
-    // If the host preloaded only a subset of backends, allow one default-path
-    // scan so missing modules can still be discovered.
-    static std::once_flag once;
-    std::call_once(once, []() {
-        if (ggml_backend_dev_count() > 0 && ggml_backend_reg_by_name("CPU") != nullptr) {
-            return;
-        }
-        ggml_backend_load_all();
-    });
-}
-
-// Do not gate this branch on GGML_CPU or GGML_CPU_ALL_VARIANTS:
-// those are CMake options used to configure ggml itself, but they are not
-// exported as PUBLIC compile definitions to stable-diffusion in backend-DL mode.
-// In practice, this target can reliably see GGML_BACKEND_DL, but not whether
-// the CPU backend was compiled as a loadable module. We therefore use runtime
-// backend discovery instead of compile-time assumptions.
-
-__STATIC_INLINE__ ggml_backend_reg_t ggml_backend_cpu_reg() {
-    ggml_backend_reg_t reg = ggml_backend_reg_by_name("CPU");
-    if (reg != nullptr) {
-        return reg;
-    }
-
-    ggml_backend_load_all_once();
-    return ggml_backend_reg_by_name("CPU");
-}
-
-__STATIC_INLINE__ ggml_backend_reg_t ggml_backend_reg_from_backend(ggml_backend_t backend) {
-    if (backend != nullptr) {
-        ggml_backend_dev_t device = ggml_backend_get_device(backend);
-        if (device != nullptr) {
-            return ggml_backend_dev_backend_reg(device);
-        }
-    }
-
-    return ggml_backend_cpu_reg();
-}
-
-__STATIC_INLINE__ ggml_backend_t ggml_backend_cpu_init() {
-    ggml_backend_t backend = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
-    if (backend != nullptr) {
-        return backend;
-    }
-
-    ggml_backend_load_all_once();
-    return ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
-}
-
-__STATIC_INLINE__ bool ggml_backend_is_cpu(ggml_backend_t backend) {
-    if (backend == nullptr) {
-        return false;
-    }
-
-    ggml_backend_dev_t device = ggml_backend_get_device(backend);
-    if (device != nullptr) {
-        return ggml_backend_dev_type(device) == GGML_BACKEND_DEVICE_TYPE_CPU;
-    }
-
-    const char* backend_name = ggml_backend_name(backend);
-    return backend_name != nullptr && std::strcmp(backend_name, "CPU") == 0;
-}
-
-__STATIC_INLINE__ void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
-    ggml_backend_reg_t reg = ggml_backend_reg_from_backend(backend_cpu);
-    if (reg == nullptr) {
-        return;
-    }
-
-    auto fn = reinterpret_cast<ggml_backend_set_n_threads_t>(ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_n_threads"));
-    if (fn != nullptr) {
-        fn(backend_cpu, n_threads);
-    }
-}
-
-using __ggml_backend_cpu_set_threadpool_t = void (*)(ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
-
-__STATIC_INLINE__ void ggml_backend_cpu_set_threadpool(ggml_backend_t backend_cpu, ggml_threadpool_t threadpool) {
-    ggml_backend_reg_t reg = ggml_backend_reg_from_backend(backend_cpu);
-    if (reg == nullptr) {
-        return;
-    }
-
-    auto fn = reinterpret_cast<__ggml_backend_cpu_set_threadpool_t>(ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_set_threadpool"));
-    if (fn != nullptr) {
-        fn(backend_cpu, threadpool);
-    }
-}
-
-__STATIC_INLINE__ void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void* abort_callback_data) {
-    ggml_backend_reg_t reg = ggml_backend_reg_from_backend(backend_cpu);
-    if (reg == nullptr) {
-        return;
-    }
-
-    auto fn = reinterpret_cast<ggml_backend_set_abort_callback_t>(ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_abort_callback"));
-    if (fn != nullptr) {
-        fn(backend_cpu, abort_callback, abort_callback_data);
-    }
-}
-
-__STATIC_INLINE__ ggml_backend_buffer_t ggml_backend_tensor_buffer(const struct ggml_tensor* tensor) {
-    if (tensor == nullptr) {
-        return nullptr;
-    }
-
-    return tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
-}
-
-__STATIC_INLINE__ bool ggml_backend_tensor_is_host_accessible(const struct ggml_tensor* tensor) {
-    if (tensor == nullptr || tensor->data == nullptr) {
-        return false;
-    }
-
-    ggml_backend_buffer_t buffer = ggml_backend_tensor_buffer(tensor);
-    return buffer == nullptr || ggml_backend_buffer_is_host(buffer);
-}
-
-__STATIC_INLINE__ size_t ggml_backend_tensor_offset(const struct ggml_tensor* tensor, int64_t i0, int64_t i1, int64_t i2, int64_t i3) {
-    return (size_t)(i0 * tensor->nb[0] + i1 * tensor->nb[1] + i2 * tensor->nb[2] + i3 * tensor->nb[3]);
-}
-
-template <typename T>
-__STATIC_INLINE__ void ggml_backend_tensor_write_scalar(const struct ggml_tensor* tensor, int64_t i0, int64_t i1, int64_t i2, int64_t i3, T value) {
-    const size_t offset = ggml_backend_tensor_offset(tensor, i0, i1, i2, i3);
-
-    if (ggml_backend_tensor_is_host_accessible(tensor)) {
-        auto* dst = reinterpret_cast<T*>(reinterpret_cast<char*>(tensor->data) + offset);
-        *dst      = value;
-        return;
-    }
-
-    ggml_backend_tensor_set(const_cast<struct ggml_tensor*>(tensor), &value, offset, sizeof(T));
-}
-
-__STATIC_INLINE__ void ggml_set_f32_nd(const struct ggml_tensor* tensor, int64_t i0, int64_t i1, int64_t i2, int64_t i3, float value) {
-    switch (tensor->type) {
-        case GGML_TYPE_I8:
-            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, static_cast<int8_t>(value));
-            break;
-        case GGML_TYPE_I16:
-            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, static_cast<int16_t>(value));
-            break;
-        case GGML_TYPE_I32:
-            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, static_cast<int32_t>(value));
-            break;
-        case GGML_TYPE_F16:
-            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, ggml_fp32_to_fp16(value));
-            break;
-        case GGML_TYPE_BF16:
-            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, ggml_fp32_to_bf16(value));
-            break;
-        case GGML_TYPE_F32:
-            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, value);
-            break;
-        default:
-            GGML_ABORT("fatal error");
-    }
-}
-
-__STATIC_INLINE__ void ggml_set_f32_1d(const struct ggml_tensor* tensor, int i, float value) {
-    if (!ggml_is_contiguous(tensor)) {
-        int64_t id[4] = {0, 0, 0, 0};
-        ggml_unravel_index(tensor, i, &id[0], &id[1], &id[2], &id[3]);
-        ggml_set_f32_nd(tensor, id[0], id[1], id[2], id[3], value);
-        return;
-    }
-
-    switch (tensor->type) {
-        case GGML_TYPE_I8:
-            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, static_cast<int8_t>(value));
-            break;
-        case GGML_TYPE_I16:
-            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, static_cast<int16_t>(value));
-            break;
-        case GGML_TYPE_I32:
-            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, static_cast<int32_t>(value));
-            break;
-        case GGML_TYPE_F16:
-            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, ggml_fp32_to_fp16(value));
-            break;
-        case GGML_TYPE_BF16:
-            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, ggml_fp32_to_bf16(value));
-            break;
-        case GGML_TYPE_F32:
-            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, value);
-            break;
-        default:
-            GGML_ABORT("fatal error");
-    }
-}
-
-__STATIC_INLINE__ enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context* ctx, struct ggml_cgraph* cgraph, int n_threads) {
-    (void)ctx;
-
-    // The legacy ggml_graph_compute_with_ctx() symbol lives in ggml-cpu, but
-    // the backend proc table does not expose it in GGML_BACKEND_DL mode.
-    // Recreate the old behavior by initializing the CPU backend explicitly and
-    // executing the graph through the generic backend API.
-    ggml_backend_t backend = ggml_backend_cpu_init();
-    if (backend == nullptr) {
-        return GGML_STATUS_ALLOC_FAILED;
-    }
-
-    ggml_backend_cpu_set_n_threads(backend, n_threads);
-
-    const enum ggml_status status = ggml_backend_graph_compute(backend, cgraph);
-    ggml_backend_free(backend);
-
-    return status;
-}
-
-__STATIC_INLINE__ ggml_tensor* ggml_set_f32(struct ggml_tensor* tensor, float value) {
-    GGML_ASSERT(tensor != nullptr);
-
-    if (ggml_backend_tensor_is_host_accessible(tensor) && ggml_is_contiguous(tensor)) {
-        const int64_t nelements = ggml_nelements(tensor);
-
-        switch (tensor->type) {
-            case GGML_TYPE_I8: {
-                auto* data     = reinterpret_cast<int8_t*>(tensor->data);
-                const int8_t v = static_cast<int8_t>(value);
-                for (int64_t i = 0; i < nelements; ++i) {
-                    data[i] = v;
-                }
-            } break;
-            case GGML_TYPE_I16: {
-                auto* data      = reinterpret_cast<int16_t*>(tensor->data);
-                const int16_t v = static_cast<int16_t>(value);
-                for (int64_t i = 0; i < nelements; ++i) {
-                    data[i] = v;
-                }
-            } break;
-            case GGML_TYPE_I32: {
-                auto* data      = reinterpret_cast<int32_t*>(tensor->data);
-                const int32_t v = static_cast<int32_t>(value);
-                for (int64_t i = 0; i < nelements; ++i) {
-                    data[i] = v;
-                }
-            } break;
-            case GGML_TYPE_F16: {
-                auto* data          = reinterpret_cast<ggml_fp16_t*>(tensor->data);
-                const ggml_fp16_t v = ggml_fp32_to_fp16(value);
-                for (int64_t i = 0; i < nelements; ++i) {
-                    data[i] = v;
-                }
-            } break;
-            case GGML_TYPE_BF16: {
-                auto* data          = reinterpret_cast<ggml_bf16_t*>(tensor->data);
-                const ggml_bf16_t v = ggml_fp32_to_bf16(value);
-                for (int64_t i = 0; i < nelements; ++i) {
-                    data[i] = v;
-                }
-            } break;
-            case GGML_TYPE_F32: {
-                auto* data = reinterpret_cast<float*>(tensor->data);
-                for (int64_t i = 0; i < nelements; ++i) {
-                    data[i] = value;
-                }
-            } break;
-            default:
-                GGML_ABORT("fatal error");
-        }
-
-        return tensor;
-    }
-
-    const int64_t nelements = ggml_nelements(tensor);
-    for (int64_t i = 0; i < nelements; ++i) {
-        ggml_set_f32_1d(tensor, static_cast<int>(i), value);
-    }
-
-    return tensor;
-}
-
-#endif
@@ -16,6 +16,8 @@

 namespace sd::ggml_graph_cut {

+    static constexpr double MAX_VRAM_BYTES_PER_GIB = 1024.0 * 1024.0 * 1024.0;
+
    static std::string graph_cut_tensor_display_name(const ggml_tensor* tensor) {
        if (tensor == nullptr) {
            return "<null>";
@@ -45,6 +47,21 @@ namespace sd::ggml_graph_cut {
        return params_tensor_set.find(tensor) != params_tensor_set.end();
    }

+    static int graph_node_index_by_name(ggml_cgraph* gf, const char* name) {
+        GGML_ASSERT(gf != nullptr);
+        if (name == nullptr || name[0] == '\0') {
+            return -1;
+        }
+        const int n_nodes = ggml_graph_n_nodes(gf);
+        for (int i = 0; i < n_nodes; ++i) {
+            ggml_tensor* node = ggml_graph_node(gf, i);
+            if (node != nullptr && std::strcmp(node->name, name) == 0) {
+                return i;
+            }
+        }
+        return -1;
+    }
+
    static Plan::InputShape input_shape(const ggml_tensor* tensor) {
        Plan::InputShape shape;
        if (tensor == nullptr) {
@@ -64,6 +81,60 @@ namespace sd::ggml_graph_cut {
               segment.output_bytes;
    }

+    size_t max_vram_gib_to_bytes(float max_vram) {
+        if (max_vram <= 0.f) {
+            return 0;
+        }
+        return static_cast<size_t>(static_cast<double>(max_vram) * MAX_VRAM_BYTES_PER_GIB);
+    }
+
+    static float max_vram_bytes_to_gib(size_t max_vram_bytes) {
+        return static_cast<float>(static_cast<double>(max_vram_bytes) / MAX_VRAM_BYTES_PER_GIB);
+    }
+
+    static size_t resolve_auto_max_vram_bytes(float spare_vram, ggml_backend_t backend) {
+        if (backend == nullptr) {
+            LOG_WARN("--max-vram < 0 requested, but no backend is available; disabling graph splitting");
+            return 0;
+        }
+
+        ggml_backend_dev_t dev = ggml_backend_get_device(backend);
+        if (dev == nullptr) {
+            LOG_WARN("--max-vram < 0 requested, but no backend device is available; disabling graph splitting");
+            return 0;
+        }
+        if (ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_CPU) {
+            LOG_WARN("--max-vram < 0 requested, but the main backend is CPU; disabling graph splitting");
+            return 0;
+        }
+
+        size_t free_vram  = 0;
+        size_t total_vram = 0;
+        ggml_backend_dev_memory(dev, &free_vram, &total_vram);
+        size_t spare_bytes = static_cast<size_t>(MAX_VRAM_BYTES_PER_GIB * spare_vram);
+
+        if (free_vram <= spare_bytes) {
+            LOG_WARN("--max-vram < 0 requested, but free VRAM is %.2f GiB; reserving %.2f GiB leaves no graph budget",
+                     free_vram / MAX_VRAM_BYTES_PER_GIB, spare_vram);
+            return 0;
+        }
+
+        const size_t max_vram_bytes = free_vram - spare_bytes;
+        LOG_INFO("--max-vram < 0 auto-detected %.2f GiB free VRAM (%.2f GiB total), reserving %.2f GiB; using %.2f GiB",
+                 free_vram / MAX_VRAM_BYTES_PER_GIB,
+                 total_vram / MAX_VRAM_BYTES_PER_GIB,
+                 spare_vram,
+                 max_vram_bytes / MAX_VRAM_BYTES_PER_GIB);
+        return max_vram_bytes;
+    }
+
+    float resolve_max_vram_gib(float max_vram, ggml_backend_t backend) {
+        if (max_vram >= 0.f) {
+            return max_vram;
+        }
+        return max_vram_bytes_to_gib(resolve_auto_max_vram_bytes(-max_vram, backend));
+    }
+
    static Segment make_segment_seed(const Plan& plan,
                                     size_t start_segment_index,
                                     size_t end_segment_index) {
@@ -244,6 +315,11 @@ namespace sd::ggml_graph_cut {
        if (tensor == nullptr) {
            return nullptr;
        }
+        if (tensor_buffer(tensor) == nullptr && tensor->src[0] != nullptr &&
+            ggml_nelements(tensor->src[0]) == ggml_nelements(tensor) &&
+            ggml_nbytes(tensor->src[0]) == ggml_nbytes(tensor)) {
+            return cache_source_tensor(tensor->src[0]);
+        }
        return tensor->view_src ? tensor->view_src : tensor;
    }

@@ -503,11 +579,15 @@ namespace sd::ggml_graph_cut {
                          log_desc);
        }

-        ggml_tensor* final_output = ggml_graph_node(gf, -1);
-        if (final_output != nullptr && available_cut_output_node_indices.find(n_nodes - 1) == available_cut_output_node_indices.end()) {
+        int final_output_index = graph_node_index_by_name(gf, "ggml_runner_final_result_tensor");
+        if (final_output_index < 0) {
+            final_output_index = n_nodes - 1;
+        }
+        ggml_tensor* final_output = final_output_index >= 0 ? ggml_graph_node(gf, final_output_index) : nullptr;
+        if (final_output != nullptr && available_cut_output_node_indices.find(final_output_index) == available_cut_output_node_indices.end()) {
            Segment final_segment;
            final_segment.group_name = "ggml_runner.final";
-            final_segment.output_node_indices.push_back(n_nodes - 1);
+            final_segment.output_node_indices.push_back(final_output_index);
            build_segment(gf,
                          plan,
                          final_segment,
@@ -619,9 +699,9 @@ namespace sd::ggml_graph_cut {
        }

        if (log_desc != nullptr) {
-            LOG_INFO("%s graph cut max_vram budget merge took %lld ms",
-                     log_desc,
-                     ggml_time_ms() - t_budget_begin);
+            LOG_DEBUG("%s graph cut max_vram budget merge took %lld ms",
+                      log_desc,
+                      ggml_time_ms() - t_budget_begin);
        }

        return merged_plan;
@@ -673,4 +753,54 @@ namespace sd::ggml_graph_cut {
        return resolved_plan;
    }

+    void annotate_residency(Plan& plan, size_t max_graph_vram_bytes) {
+        // Cached plans may be reused with a smaller live budget.
+        for (auto& seg : plan.segments) {
+            seg.residency = SegmentResidency::STREAMED;
+        }
+        if (max_graph_vram_bytes == 0 || plan.segments.size() < 2) {
+            return;
+        }
+
+        bool any_param_bearing = false;
+        for (const auto& seg : plan.segments) {
+            if (seg.input_param_bytes > 0) {
+                any_param_bearing = true;
+                break;
+            }
+        }
+        if (!any_param_bearing) {
+            return;
+        }
+
+        // Leave room for the largest active streamed segment.
+        size_t worst_streamed_footprint = 0;
+        for (const auto& seg : plan.segments) {
+            const size_t seg_footprint = seg.input_param_bytes +
+                                         seg.compute_buffer_size +
+                                         seg.output_bytes +
+                                         seg.input_previous_cut_bytes +
+                                         seg.input_external_bytes;
+            if (seg_footprint > worst_streamed_footprint) {
+                worst_streamed_footprint = seg_footprint;
+            }
+        }
+        constexpr size_t safety = 512ull * 1024 * 1024;
+        const size_t reserved   = safety + worst_streamed_footprint;
+
+        if (max_graph_vram_bytes <= reserved) {
+            return;
+        }
+        const size_t available = max_graph_vram_bytes - reserved;
+
+        size_t cumulative = 0;
+        for (auto& seg : plan.segments) {
+            if (cumulative + seg.input_param_bytes > available) {
+                break;
+            }
+            seg.residency = SegmentResidency::RESIDENT;
+            cumulative += seg.input_param_bytes;
+        }
+    }
+
 }  // namespace sd::ggml_graph_cut
@@ -2,6 +2,7 @@
 #define __SD_GGML_GRAPH_CUT_H__

 #include <array>
+#include <cstdint>
 #include <string>
 #include <unordered_set>
 #include <vector>
@@ -11,6 +12,12 @@

 namespace sd::ggml_graph_cut {

+    // Streaming residency for a segment's params.
+    enum class SegmentResidency : uint8_t {
+        STREAMED = 0,
+        RESIDENT = 1,
+    };
+
    struct Segment {
        enum InputType {
            INPUT_EXTERNAL = 0,
@@ -34,6 +41,7 @@ namespace sd::ggml_graph_cut {
        std::vector<int> internal_node_indices;
        std::vector<int> output_node_indices;
        std::vector<InputRef> input_refs;
+        SegmentResidency residency = SegmentResidency::STREAMED;
    };

    struct Plan {
@@ -83,6 +91,8 @@ namespace sd::ggml_graph_cut {
                                          ggml_cgraph* gf,
                                          const Segment& segment,
                                          const char* log_desc);
+    size_t max_vram_gib_to_bytes(float max_vram);
+    float resolve_max_vram_gib(float max_vram, ggml_backend_t backend);
    Plan build_plan(ggml_backend_t backend,
                    ggml_cgraph* gf,
                    const std::unordered_set<const ggml_tensor*>& params_tensor_set,
@@ -99,6 +109,9 @@ namespace sd::ggml_graph_cut {
                      size_t max_graph_vram_bytes,
                      const std::unordered_set<const ggml_tensor*>& params_tensor_set,
                      const char* log_desc);
+
+    // Mark leading segments resident when they fit after streamed-segment headroom.
+    void annotate_residency(Plan& plan, size_t max_graph_vram_bytes);
 }  // namespace sd::ggml_graph_cut

 #endif
@@ -0,0 +1,259 @@
+#include "guidance.h"
+
+#include <algorithm>
+#include <cmath>
+#include <cstdlib>
+#include <string>
+#include <utility>
+
+#include "util.h"
+
+namespace sd::guidance {
+
+    static bool has_tensor(const sd::Tensor<float>* tensor) {
+        return tensor != nullptr && !tensor->empty();
+    }
+
+    bool is_adaptive_projected_guidance_enabled(const AdaptiveProjectedGuidanceParams& params) {
+        return params.eta != 1.0f || params.momentum != 0.0f || params.norm_threshold > 0.0f;
+    }
+
+    AdaptiveProjectedGuidanceParams parse_adaptive_projected_guidance_args(const char* extra_sample_args) {
+        AdaptiveProjectedGuidanceParams params;
+        for (const auto& [key, value] : parse_key_value_args(extra_sample_args, "extra sample arg")) {
+            float parsed = 0.0f;
+            if (key == "apg_eta") {
+                if (!parse_strict_float(value, parsed)) {
+                    LOG_WARN("ignoring invalid APG extra sample arg '%s=%s'", key.c_str(), value.c_str());
+                    continue;
+                }
+                params.eta = parsed;
+            } else if (key == "apg_momentum") {
+                if (!parse_strict_float(value, parsed)) {
+                    LOG_WARN("ignoring invalid APG extra sample arg '%s=%s'", key.c_str(), value.c_str());
+                    continue;
+                }
+                params.momentum = parsed;
+            } else if (key == "apg_norm_threshold") {
+                if (!parse_strict_float(value, parsed)) {
+                    LOG_WARN("ignoring invalid APG extra sample arg '%s=%s'", key.c_str(), value.c_str());
+                    continue;
+                }
+                params.norm_threshold = parsed;
+            } else if (key == "apg_norm_threshold_smoothing") {
+                if (!parse_strict_float(value, parsed)) {
+                    LOG_WARN("ignoring invalid APG extra sample arg '%s=%s'", key.c_str(), value.c_str());
+                    continue;
+                }
+                params.norm_threshold_smoothing = parsed;
+            }
+        }
+        return params;
+    }
+
+    bool parse_skip_layer_guidance_uncond_arg(const char* extra_sample_args) {
+        bool uncond = false;
+        for (const auto& [key, value] : parse_key_value_args(extra_sample_args, "extra sample arg")) {
+            if (key == "slg_uncond") {
+                if (!parse_strict_bool(value, uncond)) {
+                    LOG_WARN("ignoring invalid SLG extra sample arg '%s=%s'", key.c_str(), value.c_str());
+                }
+            }
+        }
+        return uncond;
+    }
+
+    ClassifierFreeGuidance::ClassifierFreeGuidance(float guidance_scale,
+                                                   float image_guidance_scale)
+        : guidance_scale_(guidance_scale),
+          image_guidance_scale_(image_guidance_scale) {
+    }
+
+    GuiderOutput ClassifierFreeGuidance::forward(const GuidanceInput& input,
+                                                 GuiderOutput previous) const {
+        (void)previous;
+
+        GuiderOutput output;
+        if (!has_tensor(input.pred_cond)) {
+            return output;
+        }
+
+        const sd::Tensor<float>& pred_cond = *input.pred_cond;
+        output.pred                        = pred_cond;
+        if (has_tensor(input.pred_uncond)) {
+            const sd::Tensor<float>& pred_uncond = *input.pred_uncond;
+            if (has_tensor(input.pred_img_uncond)) {
+                const sd::Tensor<float>& pred_img_uncond = *input.pred_img_uncond;
+                output.pred                              = pred_img_uncond +
+                              image_guidance_scale_ * (pred_uncond - pred_img_uncond) +
+                              guidance_scale_ * (pred_cond - pred_uncond);
+
+            } else {
+                output.pred = pred_uncond + guidance_scale_ * (pred_cond - pred_uncond);
+            }
+        } else if (has_tensor(input.pred_img_uncond)) {
+            const sd::Tensor<float>& pred_img_uncond = *input.pred_img_uncond;
+            output.pred                              = pred_img_uncond + guidance_scale_ * (pred_cond - pred_img_uncond);
+        }
+
+        return output;
+    }
+
+    AdaptiveProjectedGuidance::AdaptiveProjectedGuidance(float guidance_scale,
+                                                         float image_guidance_scale,
+                                                         AdaptiveProjectedGuidanceParams params)
+        : guidance_scale_(guidance_scale),
+          image_guidance_scale_(image_guidance_scale),
+          params_(params) {
+    }
+
+    static sd::Tensor<float> calculate_guidance_delta(const sd::Tensor<float>& pred_cond,
+                                                      const sd::Tensor<float>* pred_uncond,
+                                                      const sd::Tensor<float>* pred_img_uncond,
+                                                      float guidance_scale,
+                                                      float image_guidance_scale) {
+        if (pred_img_uncond != nullptr) {
+            if (pred_uncond != nullptr && guidance_scale == 1.0f) {
+                return *pred_uncond - *pred_img_uncond;
+            }
+            if (pred_uncond != nullptr) {
+                return pred_cond +
+                       (*pred_uncond * (image_guidance_scale - guidance_scale) +
+                        *pred_img_uncond * (1.0f - image_guidance_scale)) /
+                           (guidance_scale - 1.0f);
+            }
+            return pred_cond - *pred_img_uncond;
+        }
+        return pred_cond - *pred_uncond;
+    }
+
+    GuiderOutput AdaptiveProjectedGuidance::forward(const GuidanceInput& input,
+                                                    GuiderOutput previous) const {
+        (void)previous;
+
+        GuiderOutput output;
+        if (!has_tensor(input.pred_cond)) {
+            return output;
+        }
+
+        const sd::Tensor<float>& pred_cond = *input.pred_cond;
+        output.pred                        = pred_cond;
+        if (has_tensor(input.pred_uncond)) {
+            const sd::Tensor<float>& pred_uncond = *input.pred_uncond;
+            if (has_tensor(input.pred_img_uncond)) {
+                const sd::Tensor<float>& pred_img_uncond = *input.pred_img_uncond;
+                output.pred                              = pred_img_uncond +
+                              image_guidance_scale_ * (pred_uncond - pred_img_uncond) +
+                              guidance_scale_ * (pred_cond - pred_uncond);
+            } else {
+                output.pred = pred_uncond + guidance_scale_ * (pred_cond - pred_uncond);
+            }
+        } else if (has_tensor(input.pred_img_uncond)) {
+            const sd::Tensor<float>& pred_img_uncond = *input.pred_img_uncond;
+            output.pred                              = pred_img_uncond + guidance_scale_ * (pred_cond - pred_img_uncond);
+        }
+        if (!has_tensor(input.pred_uncond) && !has_tensor(input.pred_img_uncond)) {
+            return output;
+        }
+
+        const sd::Tensor<float>* pred_uncond     = input.pred_uncond;
+        const sd::Tensor<float>* pred_img_uncond = input.pred_img_uncond;
+
+        sd::Tensor<float> deltas = calculate_guidance_delta(pred_cond,
+                                                            pred_uncond,
+                                                            pred_img_uncond,
+                                                            guidance_scale_,
+                                                            image_guidance_scale_);
+        if (params_.momentum != 0.0f) {
+            if (momentum_buffer_.shape() != deltas.shape()) {
+                momentum_buffer_ = sd::Tensor<float>::zeros_like(deltas);
+            }
+            deltas += params_.momentum * momentum_buffer_;
+            momentum_buffer_ = deltas;
+        }
+
+        float diff_norm = 0.0f;
+        if (params_.norm_threshold > 0.0f) {
+            diff_norm = std::sqrt((deltas * deltas).sum());
+        }
+
+        float apg_scale_factor = 1.0f;
+        if (params_.norm_threshold > 0.0f) {
+            if (diff_norm > 0.0f) {
+                if (params_.norm_threshold_smoothing <= 0.0f) {
+                    apg_scale_factor = std::min(1.0f, params_.norm_threshold / diff_norm);
+                } else {
+                    float x          = params_.norm_threshold / diff_norm;
+                    apg_scale_factor = x / std::pow(1.0f + std::pow(x, 1.0f / params_.norm_threshold_smoothing),
+                                                    params_.norm_threshold_smoothing);
+                }
+            }
+        }
+
+        deltas *= apg_scale_factor;
+        if (params_.eta != 1.0f) {
+            float cond_norm_sq = (pred_cond * pred_cond).sum();
+            if (cond_norm_sq != 0.0f) {
+                float projection_scale = (pred_cond * deltas).sum() / cond_norm_sq;
+                deltas += (params_.eta - 1.0f) * (projection_scale * pred_cond);
+            }
+        }
+
+        output.pred = pred_cond;
+        if (pred_uncond != nullptr) {
+            if (guidance_scale_ != 1.0f) {
+                output.pred = pred_cond + (guidance_scale_ - 1.0f) * deltas;
+            } else if (pred_img_uncond != nullptr) {
+                output.pred = pred_cond + (image_guidance_scale_ - 1.0f) * deltas;
+            }
+        } else if (pred_img_uncond != nullptr) {
+            output.pred = *pred_img_uncond + guidance_scale_ * deltas;
+        }
+
+        return output;
+    }
+
+    SkipLayerGuidance::SkipLayerGuidance(std::vector<int> layers,
+                                         float scale,
+                                         float start,
+                                         float stop)
+        : layers_(std::move(layers)),
+          scale_(scale),
+          start_(start),
+          stop_(stop) {
+    }
+
+    bool SkipLayerGuidance::is_enabled_for_step(const GuidanceInput& input) const {
+        if (layers_.empty() || input.schedule_size == 0) {
+            return false;
+        }
+
+        int start_step = static_cast<int>(start_ * static_cast<float>(input.schedule_size));
+        int stop_step  = static_cast<int>(stop_ * static_cast<float>(input.schedule_size));
+        return input.step > start_step && input.step < stop_step;
+    }
+
+    const std::vector<int>& SkipLayerGuidance::layers() const {
+        return layers_;
+    }
+
+    GuiderOutput SkipLayerGuidance::forward(const GuidanceInput& input,
+                                            GuiderOutput output) const {
+        if (scale_ == 0.0f || !is_enabled_for_step(input) || !input.predict_skip_layer) {
+            return output;
+        }
+
+        if (output.pred.empty() || !has_tensor(input.pred_cond)) {
+            return GuiderOutput();
+        }
+
+        output.pred_skip_layer = input.predict_skip_layer();
+        if (output.pred_skip_layer.empty()) {
+            return GuiderOutput();
+        }
+
+        output.pred += (*input.pred_cond - output.pred_skip_layer) * scale_;
+        return output;
+    }
+
+}  // namespace sd::guidance
@@ -0,0 +1,96 @@
+#ifndef __SD_GUIDANCE_H__
+#define __SD_GUIDANCE_H__
+
+#include <cstddef>
+#include <functional>
+#include <vector>
+
+#include "tensor.hpp"
+
+namespace sd::guidance {
+
+    struct GuiderOutput {
+        sd::Tensor<float> pred;
+        sd::Tensor<float> pred_cond;
+        sd::Tensor<float> pred_uncond;
+        sd::Tensor<float> pred_img_cond;
+        sd::Tensor<float> pred_skip_layer;
+    };
+
+    struct AdaptiveProjectedGuidanceParams {
+        float eta                      = 1.0f;
+        float momentum                 = 0.0f;
+        float norm_threshold           = 0.0f;
+        float norm_threshold_smoothing = 0.0f;
+    };
+
+    AdaptiveProjectedGuidanceParams parse_adaptive_projected_guidance_args(const char* extra_sample_args);
+    bool is_adaptive_projected_guidance_enabled(const AdaptiveProjectedGuidanceParams& params);
+    bool parse_skip_layer_guidance_uncond_arg(const char* extra_sample_args);
+
+    struct GuidanceInput {
+        int step                                 = 0;
+        size_t schedule_size                     = 0;
+        const sd::Tensor<float>* pred_cond       = nullptr;
+        const sd::Tensor<float>* pred_uncond     = nullptr;
+        const sd::Tensor<float>* pred_img_uncond = nullptr;
+
+        std::function<sd::Tensor<float>()> predict_skip_layer;
+    };
+
+    class BaseGuidance {
+    public:
+        virtual ~BaseGuidance()                                   = default;
+        virtual GuiderOutput forward(const GuidanceInput& input,
+                                     GuiderOutput previous) const = 0;
+    };
+
+    class ClassifierFreeGuidance : public BaseGuidance {
+        float guidance_scale_       = 1.0f;
+        float image_guidance_scale_ = 1.0f;
+
+    public:
+        ClassifierFreeGuidance(float guidance_scale,
+                               float image_guidance_scale);
+
+        GuiderOutput forward(const GuidanceInput& input,
+                             GuiderOutput previous) const override;
+    };
+
+    class AdaptiveProjectedGuidance : public BaseGuidance {
+        float guidance_scale_       = 1.0f;
+        float image_guidance_scale_ = 1.0f;
+        AdaptiveProjectedGuidanceParams params_;
+        mutable sd::Tensor<float> momentum_buffer_;
+
+    public:
+        AdaptiveProjectedGuidance(float guidance_scale,
+                                  float image_guidance_scale,
+                                  AdaptiveProjectedGuidanceParams params);
+
+        GuiderOutput forward(const GuidanceInput& input,
+                             GuiderOutput previous) const override;
+    };
+
+    class SkipLayerGuidance : public BaseGuidance {
+        std::vector<int> layers_;
+        float scale_ = 0.0f;
+        float start_ = 0.0f;
+        float stop_  = 1.0f;
+
+    public:
+        SkipLayerGuidance(std::vector<int> layers,
+                          float scale,
+                          float start,
+                          float stop);
+
+        bool is_enabled_for_step(const GuidanceInput& input) const;
+        const std::vector<int>& layers() const;
+
+        GuiderOutput forward(const GuidanceInput& input,
+                             GuiderOutput previous) const override;
+    };
+
+}  // namespace sd::guidance
+
+#endif  // __SD_GUIDANCE_H__
@@ -0,0 +1,679 @@
+#ifndef __SD_HIDREAM_O1_H__
+#define __SD_HIDREAM_O1_H__
+
+#include <algorithm>
+#include <array>
+#include <cmath>
+#include <cstring>
+#include <memory>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "common_dit.hpp"
+#include "conditioner.hpp"
+#include "diffusion_model.hpp"
+#include "llm.hpp"
+#include "util.h"
+
+namespace HiDreamO1 {
+    constexpr int HIDREAM_O1_GRAPH_SIZE = 32768;
+    constexpr int PATCH_SIZE            = 32;
+    constexpr int TIMESTEP_TOKEN_NUM    = 1;
+    constexpr int IMAGE_TOKEN_ID        = 151655;
+    constexpr int VISION_START_TOKEN_ID = 151652;
+
+    static inline std::string repeat_special_token(const std::string& token, int64_t count) {
+        std::string out;
+        out.reserve(static_cast<size_t>(count) * token.size());
+        for (int64_t i = 0; i < count; ++i) {
+            out += token;
+        }
+        return out;
+    }
+
+    static inline std::pair<int, int> calculate_dimensions(int max_size, double ratio) {
+        int width  = static_cast<int>(std::sqrt(max_size * max_size * ratio));
+        int height = static_cast<int>(width / ratio);
+        width      = (width / PATCH_SIZE) * PATCH_SIZE;
+        height     = (height / PATCH_SIZE) * PATCH_SIZE;
+        width      = std::max(width, PATCH_SIZE);
+        height     = std::max(height, PATCH_SIZE);
+        return {width, height};
+    }
+
+    static inline sd::Tensor<float> resize_to_area(const sd::Tensor<float>& image, int image_size) {
+        int64_t width  = image.shape()[0];
+        int64_t height = image.shape()[1];
+        int64_t s_max  = static_cast<int64_t>(image_size) * image_size;
+        double scale   = std::sqrt(static_cast<double>(s_max) / static_cast<double>(width * height));
+
+        std::vector<std::pair<int64_t, int64_t>> sizes = {
+            {(static_cast<int64_t>(std::llround(width * scale)) / PATCH_SIZE) * PATCH_SIZE, (static_cast<int64_t>(std::llround(height * scale)) / PATCH_SIZE) * PATCH_SIZE},
+            {(static_cast<int64_t>(std::llround(width * scale)) / PATCH_SIZE) * PATCH_SIZE, (static_cast<int64_t>(std::floor(height * scale)) / PATCH_SIZE) * PATCH_SIZE},
+            {(static_cast<int64_t>(std::floor(width * scale)) / PATCH_SIZE) * PATCH_SIZE, (static_cast<int64_t>(std::llround(height * scale)) / PATCH_SIZE) * PATCH_SIZE},
+            {(static_cast<int64_t>(std::floor(width * scale)) / PATCH_SIZE) * PATCH_SIZE, (static_cast<int64_t>(std::floor(height * scale)) / PATCH_SIZE) * PATCH_SIZE},
+        };
+        std::sort(sizes.begin(), sizes.end(), [](const auto& a, const auto& b) {
+            return a.first * a.second > b.first * b.second;
+        });
+
+        std::pair<int64_t, int64_t> new_size = sizes.back();
+        for (const auto& size : sizes) {
+            if (size.first > 0 && size.second > 0 && size.first * size.second <= s_max) {
+                new_size = size;
+                break;
+            }
+        }
+
+        double s1 = static_cast<double>(width) / static_cast<double>(new_size.first);
+        double s2 = static_cast<double>(height) / static_cast<double>(new_size.second);
+        sd::Tensor<float> resized;
+        if (s1 < s2) {
+            int64_t resized_h = static_cast<int64_t>(std::llround(height / s1));
+            resized           = sd::ops::interpolate(image,
+                                                     {new_size.first, resized_h, image.shape()[2], image.shape()[3]},
+                                                     sd::ops::InterpolateMode::Bicubic);
+            int64_t top       = (resized_h - new_size.second) / 2;
+            resized           = sd::ops::slice(resized, 1, top, top + new_size.second);
+        } else {
+            int64_t resized_w = static_cast<int64_t>(std::llround(width / s2));
+            resized           = sd::ops::interpolate(image,
+                                                     {resized_w, new_size.second, image.shape()[2], image.shape()[3]},
+                                                     sd::ops::InterpolateMode::Bicubic);
+            int64_t left      = (resized_w - new_size.first) / 2;
+            resized           = sd::ops::slice(resized, 0, left, left + new_size.first);
+        }
+        return resized;
+    }
+
+    static inline std::vector<int32_t> build_position_ids(const std::vector<int32_t>& input_ids,
+                                                          const std::vector<std::array<int32_t, 3>>& image_grids,
+                                                          const std::vector<int32_t>& skip_vision_start_token) {
+        std::vector<int32_t> position_ids(4 * input_ids.size(), 0);
+        int image_index = 0;
+        int st          = 0;
+        int fix_point   = 4096;
+        std::vector<int32_t> out_t;
+        std::vector<int32_t> out_h;
+        std::vector<int32_t> out_w;
+
+        while (st < static_cast<int>(input_ids.size())) {
+            int ed = st;
+            while (ed < static_cast<int>(input_ids.size()) && input_ids[ed] != IMAGE_TOKEN_ID) {
+                ed++;
+            }
+
+            if (ed >= static_cast<int>(input_ids.size())) {
+                int st_idx = out_t.empty() ? 0 : (*std::max_element(out_t.begin(), out_t.end()) + 1);
+                for (int i = 0; i < static_cast<int>(input_ids.size()) - st; ++i) {
+                    out_t.push_back(st_idx + i);
+                    out_h.push_back(st_idx + i);
+                    out_w.push_back(st_idx + i);
+                }
+                break;
+            }
+
+            int text_len = std::max(0, ed - st - skip_vision_start_token[image_index]);
+            int st_idx   = out_t.empty() ? 0 : (*std::max_element(out_t.begin(), out_t.end()) + 1);
+            for (int i = 0; i < text_len; ++i) {
+                out_t.push_back(st_idx + i);
+                out_h.push_back(st_idx + i);
+                out_w.push_back(st_idx + i);
+            }
+
+            auto grid = image_grids[image_index];
+            int base;
+            if (skip_vision_start_token[image_index]) {
+                if (fix_point > 0) {
+                    base      = fix_point;
+                    fix_point = 0;
+                } else {
+                    base = st_idx;
+                }
+            } else {
+                base = text_len + st_idx;
+            }
+            for (int32_t ti = 0; ti < grid[0]; ++ti) {
+                for (int32_t hi = 0; hi < grid[1]; ++hi) {
+                    for (int32_t wi = 0; wi < grid[2]; ++wi) {
+                        out_t.push_back(base + ti);
+                        out_h.push_back(base + hi);
+                        out_w.push_back(base + wi);
+                    }
+                }
+            }
+
+            st = ed + grid[0] * grid[1] * grid[2];
+            image_index++;
+        }
+
+        GGML_ASSERT(out_t.size() == input_ids.size());
+        for (size_t i = 0; i < input_ids.size(); ++i) {
+            // ggml IMROPE consumes 4 flattened position streams:
+            //   [t, h, w, e]
+            // llama.cpp's generic Qwen-VL fallback expands text positions as
+            // [pos, pos, pos, 0]. Keep the extra stream zeroed here too.
+            position_ids[i]                        = out_t[i];
+            position_ids[input_ids.size() + i]     = out_h[i];
+            position_ids[input_ids.size() * 2 + i] = out_w[i];
+            position_ids[input_ids.size() * 3 + i] = 0;
+        }
+        return position_ids;
+    }
+
+    struct TimestepEmbedder : public GGMLBlock {
+        int frequency_embedding_size = 256;
+
+        TimestepEmbedder(int64_t hidden_size) {
+            blocks["mlp.0"] = std::make_shared<Linear>(frequency_embedding_size, hidden_size, true);
+            blocks["mlp.2"] = std::make_shared<Linear>(hidden_size, hidden_size, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* t) {
+            auto mlp_0 = std::dynamic_pointer_cast<Linear>(blocks["mlp.0"]);
+            auto mlp_2 = std::dynamic_pointer_cast<Linear>(blocks["mlp.2"]);
+            auto emb   = ggml_ext_timestep_embedding(ctx->ggml_ctx, t, frequency_embedding_size, 10000, 1000.0f);
+            emb        = mlp_0->forward(ctx, emb);
+            emb        = ggml_silu_inplace(ctx->ggml_ctx, emb);
+            emb        = mlp_2->forward(ctx, emb);
+            return emb;
+        }
+    };
+
+    struct BottleneckPatchEmbed : public GGMLBlock {
+        BottleneckPatchEmbed(int64_t in_dim, int64_t pca_dim, int64_t embed_dim) {
+            blocks["proj1"] = std::make_shared<Linear>(in_dim, pca_dim, false);
+            blocks["proj2"] = std::make_shared<Linear>(pca_dim, embed_dim, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto proj1 = std::dynamic_pointer_cast<Linear>(blocks["proj1"]);
+            auto proj2 = std::dynamic_pointer_cast<Linear>(blocks["proj2"]);
+            return proj2->forward(ctx, proj1->forward(ctx, x));
+        }
+    };
+
+    struct FinalLayer : public GGMLBlock {
+        FinalLayer(int64_t hidden_size, int64_t out_dim) {
+            blocks["linear"] = std::make_shared<Linear>(hidden_size, out_dim, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto linear = std::dynamic_pointer_cast<Linear>(blocks["linear"]);
+            return linear->forward(ctx, x);
+        }
+    };
+
+    struct HiDreamO1Params {
+        LLM::LLMParams llm;
+        int patch_size = PATCH_SIZE;
+    };
+
+    static inline HiDreamO1Params make_hidream_o1_params() {
+        HiDreamO1Params params;
+        params.llm.arch                           = LLM::LLMArch::QWEN3_VL;
+        params.llm.hidden_size                    = 4096;
+        params.llm.intermediate_size              = 12288;
+        params.llm.num_layers                     = 36;
+        params.llm.num_heads                      = 32;
+        params.llm.num_kv_heads                   = 8;
+        params.llm.head_dim                       = 128;
+        params.llm.qkv_bias                       = false;
+        params.llm.qk_norm                        = true;
+        params.llm.vocab_size                     = 151936;
+        params.llm.rms_norm_eps                   = 1e-6f;
+        params.llm.vision.arch                    = LLM::LLMVisionArch::QWEN3_VL;
+        params.llm.vision.num_layers              = 27;
+        params.llm.vision.hidden_size             = 1152;
+        params.llm.vision.intermediate_size       = 4304;
+        params.llm.vision.num_heads               = 16;
+        params.llm.vision.out_hidden_size         = 4096;
+        params.llm.vision.patch_size              = 16;
+        params.llm.vision.spatial_merge_size      = 2;
+        params.llm.vision.temporal_patch_size     = 2;
+        params.llm.vision.num_position_embeddings = 2304;
+        return params;
+    }
+
+    struct HiDreamO1Model : public GGMLBlock {
+        HiDreamO1Params params;
+
+        HiDreamO1Model() = default;
+        explicit HiDreamO1Model(HiDreamO1Params params)
+            : params(std::move(params)) {
+            blocks["language_model"] = std::make_shared<LLM::TextModel>(this->params.llm);
+            blocks["t_embedder1"]    = std::make_shared<TimestepEmbedder>(this->params.llm.hidden_size);
+            blocks["x_embedder"]     = std::make_shared<BottleneckPatchEmbed>(this->params.patch_size * this->params.patch_size * 3,
+                                                                          this->params.llm.hidden_size / 4,
+                                                                          this->params.llm.hidden_size);
+            blocks["final_layer2"]   = std::make_shared<FinalLayer>(this->params.llm.hidden_size,
+                                                                  this->params.patch_size * this->params.patch_size * 3);
+        }
+
+        std::shared_ptr<LLM::TextModel> text_model() {
+            return std::dynamic_pointer_cast<LLM::TextModel>(blocks["language_model"]);
+        }
+
+        std::shared_ptr<TimestepEmbedder> timestep_embedder() {
+            return std::dynamic_pointer_cast<TimestepEmbedder>(blocks["t_embedder1"]);
+        }
+
+        std::shared_ptr<BottleneckPatchEmbed> patch_embedder() {
+            return std::dynamic_pointer_cast<BottleneckPatchEmbed>(blocks["x_embedder"]);
+        }
+
+        std::shared_ptr<FinalLayer> final_layer() {
+            return std::dynamic_pointer_cast<FinalLayer>(blocks["final_layer2"]);
+        }
+    };
+
+    struct HiDreamO1VisionRunner : public GGMLRunner {
+        HiDreamO1Params params;
+        std::shared_ptr<LLM::VisionModel> model;
+
+        std::vector<int> window_index_vec;
+        std::vector<int> window_inverse_index_vec;
+        std::vector<float> window_mask_vec;
+        std::vector<float> pe_vec;
+        std::array<std::vector<int32_t>, 4> pos_embed_idx_data_;
+        std::array<std::vector<float>, 4> pos_embed_weight_data_;
+
+        HiDreamO1VisionRunner(ggml_backend_t backend,
+                              ggml_backend_t params_backend,
+                              const String2TensorStorage& tensor_storage_map = {},
+                              const std::string& prefix                      = "model.visual")
+            : GGMLRunner(backend, params_backend),
+              params(make_hidream_o1_params()),
+              model(std::make_shared<LLM::VisionModel>(false, params.llm.vision)) {
+            model->init(params_ctx, tensor_storage_map, prefix);
+        }
+
+        std::string get_desc() override {
+            return "hidream_o1_vision";
+        }
+
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix = "model.visual") {
+            model->get_param_tensors(tensors, prefix);
+        }
+
+        ggml_tensor* encode_image(GGMLRunnerContext* runner_ctx, ggml_tensor* image) {
+            return LLM::LLMRunner::encode_image_common(this,
+                                                       compute_ctx,
+                                                       runner_ctx,
+                                                       image,
+                                                       params.llm.vision,
+                                                       model,
+                                                       window_index_vec,
+                                                       window_inverse_index_vec,
+                                                       window_mask_vec,
+                                                       pe_vec,
+                                                       pos_embed_idx_data_,
+                                                       pos_embed_weight_data_);
+        }
+
+        ggml_cgraph* build_graph(const sd::Tensor<float>& image_tensor) {
+            ggml_cgraph* gf    = new_graph_custom(HIDREAM_O1_GRAPH_SIZE);
+            ggml_tensor* image = make_input(image_tensor);
+            auto runner_ctx    = get_context();
+            auto image_embeds  = encode_image(&runner_ctx, image);
+            ggml_build_forward_expand(gf, image_embeds);
+            return gf;
+        }
+
+        sd::Tensor<float> compute(int n_threads, const sd::Tensor<float>& image) {
+            auto get_graph = [&]() {
+                return build_graph(image);
+            };
+            auto output = GGMLRunner::compute<float>(get_graph, n_threads, false);
+            return output.has_value() ? std::move(output.value()) : sd::Tensor<float>();
+        }
+    };
+
+    struct HiDreamO1Runner : public DiffusionModelRunner {
+        HiDreamO1Params params;
+        HiDreamO1Model model;
+
+        std::vector<float> attention_mask_vec;
+
+        HiDreamO1Runner(ggml_backend_t backend,
+                        ggml_backend_t params_backend,
+                        const String2TensorStorage& tensor_storage_map = {},
+                        const std::string& prefix                      = "model")
+            : DiffusionModelRunner(backend, params_backend, prefix),
+              params(make_hidream_o1_params()) {
+            model = HiDreamO1Model(params);
+            model.init(params_ctx, tensor_storage_map, prefix);
+        }
+
+        std::string get_desc() override {
+            return "hidream_o1";
+        }
+
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
+            model.get_param_tensors(tensors, prefix);
+        }
+
+        ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,
+                                 const sd::Tensor<float>& timestep_tensor,
+                                 const sd::Tensor<int32_t>& input_ids_tensor,
+                                 const sd::Tensor<int32_t>& input_pos_tensor,
+                                 const sd::Tensor<int32_t>& token_types_tensor,
+                                 const sd::Tensor<int32_t>& vinput_mask_tensor,
+                                 const std::vector<std::pair<int, sd::Tensor<float>>>& image_embeds_tensor,
+                                 const std::vector<sd::Tensor<float>>& ref_images) {
+            ggml_cgraph* gf        = new_graph_custom(HIDREAM_O1_GRAPH_SIZE);
+            ggml_tensor* x         = make_input(x_tensor);
+            ggml_tensor* timestep  = make_input(timestep_tensor);
+            ggml_tensor* input_ids = make_input(input_ids_tensor);
+            ggml_tensor* input_pos = make_input(input_pos_tensor);
+
+            auto text_model   = model.text_model();
+            auto t_embedder1  = model.timestep_embedder();
+            auto x_embedder   = model.patch_embedder();
+            auto final_layer2 = model.final_layer();
+
+            std::vector<ggml_tensor*> ref_image_tensors;
+            for (const auto& image : ref_images) {
+                ref_image_tensors.push_back(make_input(image));
+            }
+
+            attention_mask_vec    = std::vector<float>(static_cast<size_t>(token_types_tensor.shape()[0] * token_types_tensor.shape()[0]), 0.0f);
+            int64_t total_seq_len = token_types_tensor.shape()[0];
+            for (int64_t query = 0; query < total_seq_len; ++query) {
+                bool is_gen = token_types_tensor.values()[static_cast<size_t>(query)] > 0;
+                for (int64_t key = 0; key < total_seq_len; ++key) {
+                    if (!is_gen && key > query) {
+                        attention_mask_vec[static_cast<size_t>(query * total_seq_len + key)] = -INFINITY;
+                    }
+                }
+            }
+            auto attention_mask = ggml_new_tensor_2d(compute_ctx, GGML_TYPE_F32, total_seq_len, total_seq_len);
+            set_backend_tensor_data(attention_mask, attention_mask_vec.data());
+
+            auto runner_ctx = get_context();
+            auto txt        = text_model->embed(&runner_ctx, input_ids);
+            std::vector<std::pair<int, ggml_tensor*>> image_embeds;
+            image_embeds.reserve(image_embeds_tensor.size());
+            for (const auto& image_embed : image_embeds_tensor) {
+                image_embeds.emplace_back(image_embed.first, make_input(image_embed.second));
+            }
+            txt = LLM::splice_image_embeds(&runner_ctx, txt, image_embeds);
+
+            auto t_emb          = t_embedder1->forward(&runner_ctx, timestep);
+            int64_t txt_seq_len = input_ids->ne[0];
+            if (txt_seq_len > 1) {
+                auto prefix = ggml_ext_slice(compute_ctx, txt, 1, 0, txt_seq_len - 1);
+                txt         = ggml_concat(compute_ctx, prefix, ggml_reshape_3d(compute_ctx, t_emb, t_emb->ne[0], 1, 1), 1);
+            } else {
+                txt = ggml_reshape_3d(compute_ctx, t_emb, t_emb->ne[0], 1, 1);
+            }
+
+            auto vinputs          = DiT::pad_and_patchify(&runner_ctx, x, PATCH_SIZE, PATCH_SIZE);
+            int64_t target_tokens = vinputs->ne[1];
+            for (ggml_tensor* ref_image : ref_image_tensors) {
+                auto ref = DiT::pad_and_patchify(&runner_ctx, ref_image, PATCH_SIZE, PATCH_SIZE);
+                vinputs  = ggml_concat(compute_ctx, vinputs, ref, 1);
+            }
+            auto vis = x_embedder->forward(&runner_ctx, vinputs);
+
+            auto inputs_embeds = ggml_concat(compute_ctx, txt, vis, 1);
+            auto hidden_states = text_model->forward_embeds(&runner_ctx, inputs_embeds, input_pos, attention_mask, {});
+            auto x_pred_all    = final_layer2->forward(&runner_ctx, hidden_states);
+
+            int64_t x_pred_start = txt_seq_len;
+            if (!vinput_mask_tensor.empty()) {
+                int64_t seq_len      = static_cast<int64_t>(vinput_mask_tensor.shape()[0]);
+                int64_t first_vinput = 0;
+                while (first_vinput < seq_len && vinput_mask_tensor.values()[static_cast<size_t>(first_vinput)] == 0) {
+                    first_vinput++;
+                }
+                x_pred_start = first_vinput;
+            }
+            auto x_pred = ggml_ext_slice(compute_ctx, x_pred_all, 1, x_pred_start, x_pred_start + target_tokens);
+            x_pred      = DiT::unpatchify_and_crop(compute_ctx, x_pred, x->ne[1], x->ne[0], PATCH_SIZE, PATCH_SIZE);
+
+            float sigma = 1.0f - timestep_tensor.values()[0];
+            sigma       = std::max(1e-6f, sigma);
+            auto out    = ggml_scale(compute_ctx, ggml_sub(compute_ctx, x, x_pred), 1.0f / sigma);
+
+            ggml_build_forward_expand(gf, out);
+            return gf;
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const sd::Tensor<float>& x,
+                                  const sd::Tensor<float>& timestep,
+                                  const sd::Tensor<int32_t>& input_ids,
+                                  const sd::Tensor<int32_t>& input_pos,
+                                  const sd::Tensor<int32_t>& token_types,
+                                  const sd::Tensor<int32_t>& vinput_mask,
+                                  const std::vector<std::pair<int, sd::Tensor<float>>>& image_embeds,
+                                  const std::vector<sd::Tensor<float>>& ref_images) {
+            auto get_graph = [&]() {
+                return build_graph(x, timestep, input_ids, input_pos, token_types, vinput_mask, image_embeds, ref_images);
+            };
+            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), x.dim());
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const DiffusionParams& diffusion_params) override {
+            GGML_ASSERT(diffusion_params.x != nullptr);
+            GGML_ASSERT(diffusion_params.timesteps != nullptr);
+            const auto* extra = diffusion_extra_as<HiDreamO1DiffusionExtra>(diffusion_params);
+            GGML_ASSERT(extra != nullptr);
+            GGML_ASSERT(extra->input_ids != nullptr);
+            GGML_ASSERT(extra->input_pos != nullptr);
+            GGML_ASSERT(extra->token_types != nullptr);
+            static const std::vector<sd::Tensor<float>> empty_images;
+            static const std::vector<std::pair<int, sd::Tensor<float>>> empty_image_embeds;
+            return compute(n_threads,
+                           *diffusion_params.x,
+                           *diffusion_params.timesteps,
+                           *extra->input_ids,
+                           *extra->input_pos,
+                           *extra->token_types,
+                           tensor_or_empty(extra->vinput_mask),
+                           extra->image_embeds ? *extra->image_embeds : empty_image_embeds,
+                           diffusion_params.ref_latents ? *diffusion_params.ref_latents : empty_images);
+        }
+    };
+
+    struct HiDreamO1Conditioner : public Conditioner {
+        Qwen2Tokenizer tokenizer;
+        std::shared_ptr<HiDreamO1VisionRunner> vision_runner;
+
+        HiDreamO1Conditioner(ggml_backend_t backend,
+                             ggml_backend_t params_backend,
+                             const String2TensorStorage& tensor_storage_map = {})
+            : vision_runner(std::make_shared<HiDreamO1VisionRunner>(backend, params_backend, tensor_storage_map)) {}
+
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+            vision_runner->get_param_tensors(tensors);
+        }
+
+        bool alloc_params_buffer() override {
+            if (!vision_runner->alloc_params_buffer()) {
+                return false;
+            }
+            return true;
+        }
+
+        void free_params_buffer() override {
+            vision_runner->free_params_buffer();
+        }
+
+        size_t get_params_buffer_size() override {
+            return vision_runner->get_params_buffer_size();
+        }
+
+        void set_max_graph_vram_bytes(size_t max_graph_vram_bytes) override {
+            vision_runner->set_max_graph_vram_bytes(max_graph_vram_bytes);
+        }
+
+        void set_flash_attention_enabled(bool enabled) override {
+            vision_runner->set_flash_attention_enabled(enabled);
+        }
+
+        void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
+            vision_runner->set_weight_adapter(adapter);
+        }
+
+        SDCondition get_learned_condition(int n_threads,
+                                          const ConditionerParams& conditioner_params) override {
+            SDCondition result;
+
+            int width                = conditioner_params.width;
+            int height               = conditioner_params.height;
+            int64_t target_image_len = static_cast<int64_t>(width / PATCH_SIZE) * static_cast<int64_t>(height / PATCH_SIZE);
+
+            std::vector<sd::Tensor<float>> ref_images;
+            if (conditioner_params.ref_images != nullptr) {
+                ref_images = *conditioner_params.ref_images;
+            }
+
+            std::vector<std::pair<int, sd::Tensor<float>>> vlm_images;
+            std::vector<std::array<int32_t, 3>> image_grids;
+            std::vector<int32_t> skip_vision_start;
+
+            std::string prompt = "<|im_start|>user\n";
+
+            if (ref_images.empty()) {
+                prompt += conditioner_params.text;
+                prompt += "<|im_end|>\n<|im_start|>assistant\n<|boi_token|><|tms_token|>";
+                auto input_ids = tokenizer.encode(prompt, nullptr);
+
+                std::vector<int32_t> input_ids_pad = input_ids;
+                input_ids_pad.push_back(VISION_START_TOKEN_ID);
+                input_ids_pad.insert(input_ids_pad.end(), target_image_len - 1, IMAGE_TOKEN_ID);
+
+                image_grids.push_back({1, static_cast<int32_t>(height / PATCH_SIZE), static_cast<int32_t>(width / PATCH_SIZE)});
+                skip_vision_start.push_back(1);
+
+                std::vector<int32_t> token_types(input_ids_pad.size(), 0);
+                int txt_seq_len = static_cast<int>(input_ids.size());
+                int bgn         = txt_seq_len - TIMESTEP_TOKEN_NUM;
+                for (int i = bgn; i < static_cast<int>(token_types.size()); ++i) {
+                    token_types[i] = 1;
+                }
+
+                auto position_ids = build_position_ids(input_ids_pad, image_grids, skip_vision_start);
+
+                std::vector<int64_t> input_shape{static_cast<int64_t>(input_ids.size())};
+                std::vector<int64_t> position_shape{static_cast<int64_t>(input_ids_pad.size() * 4)};
+                std::vector<int64_t> token_type_shape{static_cast<int64_t>(token_types.size())};
+                std::vector<int32_t> vinput_mask(token_types.size(), 0);
+                for (int64_t i = txt_seq_len; i < static_cast<int64_t>(vinput_mask.size()); ++i) {
+                    vinput_mask[static_cast<size_t>(i)] = 1;
+                }
+                std::vector<int64_t> vinput_mask_shape{static_cast<int64_t>(vinput_mask.size())};
+
+                result.c_input_ids    = sd::Tensor<int32_t>(input_shape, std::move(input_ids));
+                result.c_position_ids = sd::Tensor<int32_t>(position_shape, position_ids);
+                result.c_token_types  = sd::Tensor<int32_t>(token_type_shape, std::move(token_types));
+                result.c_vinput_mask  = sd::Tensor<int32_t>(vinput_mask_shape, std::move(vinput_mask));
+                return result;
+            }
+
+            int K = static_cast<int>(ref_images.size());
+            int max_size;
+            if (K == 1) {
+                max_size = std::max(height, width);
+            } else if (K == 2) {
+                max_size = std::max(height, width) * 48 / 64;
+            } else if (K <= 4) {
+                max_size = std::max(height, width) / 2;
+            } else if (K <= 8) {
+                max_size = std::max(height, width) * 24 / 64;
+            } else {
+                max_size = std::max(height, width) / 4;
+            }
+
+            int cond_img_size;
+            if (K <= 4) {
+                cond_img_size = 384;
+            } else if (K <= 8) {
+                cond_img_size = 384 * 48 / 64;
+            } else {
+                cond_img_size = 384 / 2;
+            }
+
+            for (const auto& ref_image : ref_images) {
+                auto resized_ref = resize_to_area(ref_image, max_size);
+                resized_ref      = sd::ops::clamp(resized_ref, 0.0f, 1.0f);
+
+                // VLM image: Qwen3-VL expects mean=[0.5]/std=[0.5] (i.e. range [-1,1]),
+                // not CLIP normalization. Resize the already-resized ref directly to
+                // (cond_w, cond_h) to match the Python pipeline's pil_r.resize().
+                auto dims                   = calculate_dimensions(cond_img_size,
+                                                                   static_cast<double>(resized_ref.shape()[0]) / static_cast<double>(resized_ref.shape()[1]));
+                sd::Tensor<float> vlm_image = sd::ops::interpolate(
+                    resized_ref,
+                    {dims.first, dims.second, resized_ref.shape()[2], resized_ref.shape()[3]});
+                vlm_image            = vlm_image * 2.0f - 1.0f;
+                int64_t image_tokens = static_cast<int64_t>(dims.first / PATCH_SIZE) * static_cast<int64_t>(dims.second / PATCH_SIZE);
+
+                auto patch_img = resized_ref * 2.0f - 1.0f;
+                result.c_ref_images.push_back(std::move(patch_img));
+                int64_t prompt_start = static_cast<int64_t>(tokenizer.encode(prompt + "<|vision_start|>", nullptr).size());
+                prompt += "<|vision_start|>";
+                prompt += repeat_special_token("<|image_pad|>", image_tokens);
+                prompt += "<|vision_end|>";
+                vlm_images.emplace_back(static_cast<int>(prompt_start), std::move(vlm_image));
+                image_grids.push_back({1, dims.second / PATCH_SIZE, dims.first / PATCH_SIZE});
+                skip_vision_start.push_back(0);
+            }
+
+            prompt += conditioner_params.text;
+            prompt += "<|im_end|>\n<|im_start|>assistant\n<|boi_token|><|tms_token|>";
+            auto input_ids = tokenizer.encode(prompt, nullptr);
+
+            std::vector<int32_t> input_ids_pad = input_ids;
+            input_ids_pad.push_back(VISION_START_TOKEN_ID);
+            input_ids_pad.insert(input_ids_pad.end(), target_image_len - 1, IMAGE_TOKEN_ID);
+            image_grids.push_back({1, static_cast<int32_t>(height / PATCH_SIZE), static_cast<int32_t>(width / PATCH_SIZE)});
+            skip_vision_start.push_back(1);
+
+            for (const auto& ref_image : result.c_ref_images) {
+                int64_t ref_len = static_cast<int64_t>(ref_image.shape()[0] / PATCH_SIZE) * static_cast<int64_t>(ref_image.shape()[1] / PATCH_SIZE);
+                input_ids_pad.push_back(VISION_START_TOKEN_ID);
+                input_ids_pad.insert(input_ids_pad.end(), ref_len - 1, IMAGE_TOKEN_ID);
+                image_grids.push_back({1, static_cast<int32_t>(ref_image.shape()[1] / PATCH_SIZE), static_cast<int32_t>(ref_image.shape()[0] / PATCH_SIZE)});
+                skip_vision_start.push_back(1);
+            }
+
+            std::vector<int32_t> token_types(input_ids_pad.size(), 0);
+            int txt_seq_len = static_cast<int>(input_ids.size());
+            int bgn         = txt_seq_len - TIMESTEP_TOKEN_NUM;
+            for (int i = bgn; i < static_cast<int>(token_types.size()); ++i) {
+                token_types[i] = 1;
+            }
+
+            std::vector<int64_t> input_shape{static_cast<int64_t>(input_ids.size())};
+            std::vector<int64_t> position_shape{static_cast<int64_t>(input_ids_pad.size() * 4)};
+            std::vector<int64_t> token_type_shape{static_cast<int64_t>(token_types.size())};
+            std::vector<int32_t> vinput_mask(token_types.size(), 0);
+            for (int i = txt_seq_len; i < static_cast<int>(vinput_mask.size()); ++i) {
+                vinput_mask[static_cast<size_t>(i)] = 1;
+            }
+            std::vector<int64_t> vinput_mask_shape{static_cast<int64_t>(vinput_mask.size())};
+
+            result.c_input_ids    = sd::Tensor<int32_t>(input_shape, std::move(input_ids));
+            result.c_position_ids = sd::Tensor<int32_t>(position_shape, build_position_ids(input_ids_pad, image_grids, skip_vision_start));
+            result.c_token_types  = sd::Tensor<int32_t>(token_type_shape, std::move(token_types));
+            result.c_vinput_mask  = sd::Tensor<int32_t>(vinput_mask_shape, std::move(vinput_mask));
+            result.c_image_embeds.reserve(vlm_images.size());
+            for (const auto& vlm_image : vlm_images) {
+                auto image_embed = vision_runner->compute(n_threads, vlm_image.second);
+                if (image_embed.empty()) {
+                    LOG_ERROR("hidream_o1 conditioner: encode VLM image failed");
+                    return SDCondition();
+                }
+                result.c_image_embeds.emplace_back(vlm_image.first, std::move(image_embed));
+            }
+            return result;
+        }
+    };
+}  // namespace HiDreamO1
+
+#endif  // __SD_HIDREAM_O1_H__
@@ -0,0 +1,527 @@
+#ifndef __IDEOGRAM4_HPP__
+#define __IDEOGRAM4_HPP__
+
+#include <algorithm>
+#include <cmath>
+#include <cstdlib>
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "diffusion_model.hpp"
+#include "ggml_extend.hpp"
+#include "ggml_graph_cut.h"
+#include "rope.hpp"
+
+namespace Ideogram4 {
+    constexpr int IDEOGRAM4_GRAPH_SIZE    = 65536;
+    constexpr int OUTPUT_IMAGE_INDICATOR  = 2;
+    constexpr int IMAGE_POSITION_OFFSET   = 65536;
+    constexpr int DEFAULT_MROPE_SECTION_T = 24;
+    constexpr int DEFAULT_MROPE_SECTION_H = 20;
+    constexpr int DEFAULT_MROPE_SECTION_W = 20;
+    constexpr int TIMESTEP_MAX_PERIOD     = 10000;
+    constexpr int LLM_HIDDEN_STATE_LAYERS = 13;
+
+    struct Ideogram4Config {
+        int64_t emb_dim                = 4608;
+        int64_t num_layers             = 34;
+        int64_t num_heads              = 18;
+        int64_t intermediate_size      = 12288;
+        int64_t adanln_dim             = 512;
+        int64_t in_channels            = 128;
+        int64_t llm_features_dim       = 53248;
+        int64_t rope_theta             = 5000000;
+        float norm_eps                 = 1e-5f;
+        int patch_size                 = 2;
+        int ae_channels                = 32;
+        std::vector<int> mrope_section = {DEFAULT_MROPE_SECTION_T,
+                                          DEFAULT_MROPE_SECTION_H,
+                                          DEFAULT_MROPE_SECTION_W};
+    };
+
+    __STATIC_INLINE__ ggml_tensor* timestep_embedding_sin_cos(ggml_context* ctx,
+                                                              ggml_tensor* timesteps,
+                                                              int dim) {
+        GGML_ASSERT(dim % 2 == 0);
+        auto embedding = ggml_ext_timestep_embedding(ctx, timesteps, dim, TIMESTEP_MAX_PERIOD, 10.f);
+        auto chunks    = ggml_ext_chunk(ctx, embedding, 2, 0);
+        return ggml_concat(ctx, chunks[1], chunks[0], 0);
+    }
+
+    __STATIC_INLINE__ ggml_tensor* to_token_modulation(ggml_context* ctx, ggml_tensor* x) {
+        // [N, C] -> [N, 1, C] in PyTorch layout.
+        if (ggml_n_dims(x) < 3 || x->ne[1] != 1) {
+            x = ggml_reshape_3d(ctx, x, x->ne[0], 1, x->ne[1]);
+        }
+        return x;
+    }
+
+    __STATIC_INLINE__ ggml_tensor* interleave_hidden_state_layers(ggml_context* ctx, ggml_tensor* x) {
+        // Match upstream stack(...).permute(1, 2, 3, 0).reshape(...):
+        // [layers * hidden, tokens, batch] -> [hidden * layers, tokens, batch].
+        GGML_ASSERT(x->ne[0] % LLM_HIDDEN_STATE_LAYERS == 0);
+        const int64_t hidden_size = x->ne[0] / LLM_HIDDEN_STATE_LAYERS;
+        const int64_t token_count = x->ne[1];
+        const int64_t batch_count = x->ne[2];
+
+        x = ggml_reshape_4d(ctx, x, hidden_size, LLM_HIDDEN_STATE_LAYERS, token_count, batch_count);
+        x = ggml_cont(ctx, ggml_permute(ctx, x, 1, 0, 2, 3));
+        return ggml_reshape_3d(ctx, x, hidden_size * LLM_HIDDEN_STATE_LAYERS, token_count, batch_count);
+    }
+
+    __STATIC_INLINE__ ggml_tensor* modulate(ggml_context* ctx, ggml_tensor* x, ggml_tensor* scale) {
+        scale = to_token_modulation(ctx, scale);
+        return ggml_add(ctx, x, ggml_mul(ctx, x, scale));
+    }
+
+    __STATIC_INLINE__ ggml_tensor* patchify(ggml_context* ctx, ggml_tensor* x, const Ideogram4Config& config) {
+        // x: [N, 128, H, W] with channel order [ae, ph, pw].
+        // return: [N, H*W, 128] with token channel order [ph, pw, ae].
+        const int64_t W = x->ne[0];
+        const int64_t H = x->ne[1];
+        const int64_t C = x->ne[2];
+        const int64_t N = x->ne[3];
+
+        GGML_ASSERT(N == 1);
+        GGML_ASSERT(C == config.ae_channels * config.patch_size * config.patch_size);
+
+        x = ggml_cont(ctx, x);
+        x = ggml_reshape_4d(ctx, x, W * H, config.patch_size, config.patch_size, config.ae_channels);
+        x = ggml_cont(ctx, ggml_permute(ctx, x, 3, 1, 2, 0));
+        x = ggml_reshape_3d(ctx, x, C, W * H, N);
+        return x;
+    }
+
+    __STATIC_INLINE__ ggml_tensor* unpatchify(ggml_context* ctx,
+                                              ggml_tensor* x,
+                                              int64_t H,
+                                              int64_t W,
+                                              const Ideogram4Config& config) {
+        const int64_t C = x->ne[0];
+        const int64_t N = x->ne[2];
+
+        GGML_ASSERT(N == 1);
+        GGML_ASSERT(C == config.ae_channels * config.patch_size * config.patch_size);
+        GGML_ASSERT(x->ne[1] == H * W);
+
+        x = ggml_reshape_4d(ctx, x, config.ae_channels, config.patch_size, config.patch_size, H * W);
+        x = ggml_cont(ctx, ggml_permute(ctx, x, 3, 1, 2, 0));
+        x = ggml_reshape_4d(ctx, x, W, H, C, N);
+        return x;
+    }
+
+    __STATIC_INLINE__ std::shared_ptr<Linear> make_linear(int64_t in_features,
+                                                          int64_t out_features,
+                                                          bool bias = true) {
+        return std::make_shared<Linear>(in_features, out_features, bias, false, false, 1.f, true);
+    }
+
+    __STATIC_INLINE__ std::vector<float> gen_ideogram4_pe(int grid_h,
+                                                          int grid_w,
+                                                          int bs,
+                                                          int context_len,
+                                                          int head_dim,
+                                                          int rope_theta,
+                                                          const std::vector<int>& mrope_section) {
+        GGML_ASSERT(bs == 1);
+        std::vector<std::vector<float>> ids(static_cast<size_t>(bs) * (context_len + grid_h * grid_w),
+                                            std::vector<float>(3, 0.f));
+
+        for (int i = 0; i < context_len; ++i) {
+            ids[i] = {static_cast<float>(i), static_cast<float>(i), static_cast<float>(i)};
+        }
+
+        int cursor = context_len;
+        for (int y = 0; y < grid_h; ++y) {
+            for (int x = 0; x < grid_w; ++x) {
+                ids[cursor++] = {static_cast<float>(IMAGE_POSITION_OFFSET),
+                                 static_cast<float>(IMAGE_POSITION_OFFSET + y),
+                                 static_cast<float>(IMAGE_POSITION_OFFSET + x)};
+            }
+        }
+
+        return Rope::embed_interleaved_mrope(ids, bs, static_cast<float>(rope_theta), head_dim, mrope_section);
+    }
+
+    class Ideogram4Attention : public GGMLBlock {
+    protected:
+        int64_t hidden_size;
+        int64_t num_heads;
+        int64_t head_dim;
+
+    public:
+        Ideogram4Attention(int64_t hidden_size, int64_t num_heads, float eps)
+            : hidden_size(hidden_size), num_heads(num_heads), head_dim(hidden_size / num_heads) {
+            GGML_ASSERT(hidden_size % num_heads == 0);
+            blocks["qkv"]    = make_linear(hidden_size, hidden_size * 3, false);
+            blocks["norm_q"] = std::make_shared<RMSNorm>(head_dim, eps);
+            blocks["norm_k"] = std::make_shared<RMSNorm>(head_dim, eps);
+            blocks["o"]      = make_linear(hidden_size, hidden_size, false);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* pe,
+                             ggml_tensor* mask = nullptr) {
+            int64_t n_token = x->ne[1];
+            int64_t N       = x->ne[2];
+
+            auto qkv_proj = std::dynamic_pointer_cast<Linear>(blocks["qkv"]);
+            auto norm_q   = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_q"]);
+            auto norm_k   = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_k"]);
+            auto out_proj = std::dynamic_pointer_cast<Linear>(blocks["o"]);
+
+            auto qkv     = qkv_proj->forward(ctx, x);
+            auto qkv_vec = split_qkv(ctx->ggml_ctx, qkv);
+            auto q       = ggml_reshape_4d(ctx->ggml_ctx, qkv_vec[0], head_dim, num_heads, n_token, N);
+            auto k       = ggml_reshape_4d(ctx->ggml_ctx, qkv_vec[1], head_dim, num_heads, n_token, N);
+            auto v       = ggml_reshape_4d(ctx->ggml_ctx, qkv_vec[2], head_dim, num_heads, n_token, N);
+
+            q = norm_q->forward(ctx, q);
+            k = norm_k->forward(ctx, k);
+
+            x = Rope::attention(ctx, q, k, v, pe, mask, 1.f / 128.f, false);
+            x = out_proj->forward(ctx, x);
+            return x;
+        }
+    };
+
+    class Ideogram4MLP : public GGMLBlock {
+    public:
+        Ideogram4MLP(int64_t dim, int64_t hidden_dim) {
+            blocks["w1"] = make_linear(dim, hidden_dim, false);
+            blocks["w2"] = make_linear(hidden_dim, dim, false);
+            blocks["w3"] = make_linear(dim, hidden_dim, false);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto w1 = std::dynamic_pointer_cast<Linear>(blocks["w1"]);
+            auto w2 = std::dynamic_pointer_cast<Linear>(blocks["w2"]);
+            auto w3 = std::dynamic_pointer_cast<Linear>(blocks["w3"]);
+
+            auto x1 = ggml_silu(ctx->ggml_ctx, w1->forward(ctx, x));
+            auto x3 = w3->forward(ctx, x);
+            x       = ggml_mul(ctx->ggml_ctx, x1, x3);
+            x       = w2->forward(ctx, x);
+            return x;
+        }
+    };
+
+    class Ideogram4TransformerBlock : public GGMLBlock {
+    public:
+        Ideogram4TransformerBlock(const Ideogram4Config& config) {
+            blocks["attention"]        = std::make_shared<Ideogram4Attention>(config.emb_dim, config.num_heads, config.norm_eps);
+            blocks["feed_forward"]     = std::make_shared<Ideogram4MLP>(config.emb_dim, config.intermediate_size);
+            blocks["attention_norm1"]  = std::make_shared<RMSNorm>(config.emb_dim, config.norm_eps);
+            blocks["ffn_norm1"]        = std::make_shared<RMSNorm>(config.emb_dim, config.norm_eps);
+            blocks["attention_norm2"]  = std::make_shared<RMSNorm>(config.emb_dim, config.norm_eps);
+            blocks["ffn_norm2"]        = std::make_shared<RMSNorm>(config.emb_dim, config.norm_eps);
+            blocks["adaln_modulation"] = make_linear(config.adanln_dim, 4 * config.emb_dim, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* pe,
+                             ggml_tensor* adaln_input,
+                             ggml_tensor* mask = nullptr) {
+            auto attention        = std::dynamic_pointer_cast<Ideogram4Attention>(blocks["attention"]);
+            auto feed_forward     = std::dynamic_pointer_cast<Ideogram4MLP>(blocks["feed_forward"]);
+            auto attention_norm1  = std::dynamic_pointer_cast<RMSNorm>(blocks["attention_norm1"]);
+            auto ffn_norm1        = std::dynamic_pointer_cast<RMSNorm>(blocks["ffn_norm1"]);
+            auto attention_norm2  = std::dynamic_pointer_cast<RMSNorm>(blocks["attention_norm2"]);
+            auto ffn_norm2        = std::dynamic_pointer_cast<RMSNorm>(blocks["ffn_norm2"]);
+            auto adaln_modulation = std::dynamic_pointer_cast<Linear>(blocks["adaln_modulation"]);
+
+            auto mod       = adaln_modulation->forward(ctx, adaln_input);
+            auto mods      = ggml_ext_chunk(ctx->ggml_ctx, mod, 4, 0);
+            auto scale_msa = mods[0];
+            auto gate_msa  = to_token_modulation(ctx->ggml_ctx, ggml_tanh(ctx->ggml_ctx, mods[1]));
+            auto scale_mlp = mods[2];
+            auto gate_mlp  = to_token_modulation(ctx->ggml_ctx, ggml_tanh(ctx->ggml_ctx, mods[3]));
+
+            auto attn_out = attention_norm1->forward(ctx, x);
+            attn_out      = modulate(ctx->ggml_ctx, attn_out, scale_msa);
+            attn_out      = attention->forward(ctx, attn_out, pe, mask);
+            attn_out      = attention_norm2->forward(ctx, attn_out);
+            x             = ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, attn_out, gate_msa));
+
+            auto ffn_out = ffn_norm1->forward(ctx, x);
+            ffn_out      = modulate(ctx->ggml_ctx, ffn_out, scale_mlp);
+            ffn_out      = feed_forward->forward(ctx, ffn_out);
+            ffn_out      = ffn_norm2->forward(ctx, ffn_out);
+            x            = ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, ffn_out, gate_mlp));
+
+            return x;
+        }
+    };
+
+    class Ideogram4EmbedScalar : public GGMLBlock {
+    protected:
+        int64_t dim;
+
+    public:
+        Ideogram4EmbedScalar(int64_t dim)
+            : dim(dim) {
+            blocks["mlp_in"]  = make_linear(dim, dim, true);
+            blocks["mlp_out"] = make_linear(dim, dim, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto mlp_in  = std::dynamic_pointer_cast<Linear>(blocks["mlp_in"]);
+            auto mlp_out = std::dynamic_pointer_cast<Linear>(blocks["mlp_out"]);
+
+            x = timestep_embedding_sin_cos(ctx->ggml_ctx, x, static_cast<int>(dim));
+            x = ggml_silu(ctx->ggml_ctx, mlp_in->forward(ctx, x));
+            x = mlp_out->forward(ctx, x);
+            return x;
+        }
+    };
+
+    class Ideogram4FinalLayer : public GGMLBlock {
+    public:
+        Ideogram4FinalLayer(const Ideogram4Config& config) {
+            blocks["norm_final"]       = std::make_shared<LayerNorm>(config.emb_dim, 1e-6f, false);
+            blocks["linear"]           = make_linear(config.emb_dim, config.in_channels, true);
+            blocks["adaln_modulation"] = make_linear(config.adanln_dim, config.emb_dim, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* c) {
+            auto norm_final       = std::dynamic_pointer_cast<LayerNorm>(blocks["norm_final"]);
+            auto linear           = std::dynamic_pointer_cast<Linear>(blocks["linear"]);
+            auto adaln_modulation = std::dynamic_pointer_cast<Linear>(blocks["adaln_modulation"]);
+
+            auto scale = adaln_modulation->forward(ctx, ggml_silu(ctx->ggml_ctx, c));
+            x          = norm_final->forward(ctx, x);
+            x          = modulate(ctx->ggml_ctx, x, scale);
+            x          = linear->forward(ctx, x);
+            return x;
+        }
+    };
+
+    class Ideogram4Transformer : public GGMLBlock {
+    protected:
+        Ideogram4Config config;
+
+    public:
+        Ideogram4Transformer() = default;
+        explicit Ideogram4Transformer(Ideogram4Config config)
+            : config(std::move(config)) {
+            blocks["input_proj"]            = make_linear(this->config.in_channels, this->config.emb_dim, true);
+            blocks["llm_cond_norm"]         = std::make_shared<RMSNorm>(this->config.llm_features_dim, 1e-6f);
+            blocks["llm_cond_proj"]         = make_linear(this->config.llm_features_dim, this->config.emb_dim, true);
+            blocks["t_embedding"]           = std::make_shared<Ideogram4EmbedScalar>(this->config.emb_dim);
+            blocks["adaln_proj"]            = make_linear(this->config.emb_dim, this->config.adanln_dim, true);
+            blocks["embed_image_indicator"] = std::make_shared<Embedding>(2, this->config.emb_dim);
+
+            for (int i = 0; i < this->config.num_layers; ++i) {
+                blocks["layers." + std::to_string(i)] = std::make_shared<Ideogram4TransformerBlock>(this->config);
+            }
+            blocks["final_layer"] = std::make_shared<Ideogram4FinalLayer>(this->config);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* timestep,
+                             ggml_tensor* context,
+                             ggml_tensor* pe,
+                             ggml_tensor* image_indicator_ids) {
+            int64_t W = x->ne[0];
+            int64_t H = x->ne[1];
+            int64_t N = x->ne[3];
+            GGML_ASSERT(N == 1);
+
+            auto input_proj            = std::dynamic_pointer_cast<Linear>(blocks["input_proj"]);
+            auto llm_cond_norm         = std::dynamic_pointer_cast<RMSNorm>(blocks["llm_cond_norm"]);
+            auto llm_cond_proj         = std::dynamic_pointer_cast<Linear>(blocks["llm_cond_proj"]);
+            auto t_embedding           = std::dynamic_pointer_cast<Ideogram4EmbedScalar>(blocks["t_embedding"]);
+            auto adaln_proj            = std::dynamic_pointer_cast<Linear>(blocks["adaln_proj"]);
+            auto embed_image_indicator = std::dynamic_pointer_cast<Embedding>(blocks["embed_image_indicator"]);
+            auto final_layer           = std::dynamic_pointer_cast<Ideogram4FinalLayer>(blocks["final_layer"]);
+
+            auto img = patchify(ctx->ggml_ctx, x, config);
+            img      = input_proj->forward(ctx, img);
+
+            ggml_tensor* h      = img;
+            int64_t context_len = 0;
+            if (context != nullptr) {
+                if (ggml_n_dims(context) < 3) {
+                    context = ggml_reshape_3d(ctx->ggml_ctx, context, context->ne[0], context->ne[1], 1);
+                }
+                context     = interleave_hidden_state_layers(ctx->ggml_ctx, context);
+                context_len = context->ne[1];
+                auto txt    = llm_cond_norm->forward(ctx, context);
+                txt         = llm_cond_proj->forward(ctx, txt);
+                h           = ggml_concat(ctx->ggml_ctx, txt, img, 1);
+            }
+
+            auto indicator_embedding = embed_image_indicator->forward(ctx, image_indicator_ids);
+            h                        = ggml_add(ctx->ggml_ctx, h, indicator_embedding);
+
+            auto t_cond      = t_embedding->forward(ctx, timestep);
+            auto adaln_input = ggml_silu(ctx->ggml_ctx, adaln_proj->forward(ctx, t_cond));
+
+            for (int i = 0; i < config.num_layers; ++i) {
+                auto block = std::dynamic_pointer_cast<Ideogram4TransformerBlock>(blocks["layers." + std::to_string(i)]);
+                h          = block->forward(ctx, h, pe, adaln_input, nullptr);
+                sd::ggml_graph_cut::mark_graph_cut(h, "ideogram4.layers." + std::to_string(i), "hidden");
+            }
+
+            h = final_layer->forward(ctx, h, adaln_input);
+            if (context_len > 0) {
+                h = ggml_ext_slice(ctx->ggml_ctx, h, 1, context_len, h->ne[1]);
+            }
+
+            h = unpatchify(ctx->ggml_ctx, h, H, W, config);
+            h = ggml_ext_scale(ctx->ggml_ctx, h, -1.f);
+            return h;
+        }
+    };
+
+    class Ideogram4Runner : public DiffusionModelRunner {
+    protected:
+        static int64_t detect_num_layers(const String2TensorStorage& tensor_storage_map,
+                                         const std::string& prefix) {
+            int64_t detected_layers  = 0;
+            std::string layer_prefix = prefix.empty() ? "layers." : prefix + ".layers.";
+            for (const auto& pair : tensor_storage_map) {
+                const std::string& name = pair.first;
+                if (name.find(layer_prefix) != 0) {
+                    continue;
+                }
+                std::string tail = name.substr(layer_prefix.size());
+                size_t dot       = tail.find('.');
+                if (dot == std::string::npos) {
+                    continue;
+                }
+                int layer_idx   = std::atoi(tail.substr(0, dot).c_str());
+                detected_layers = std::max<int64_t>(detected_layers, layer_idx + 1);
+            }
+            return detected_layers;
+        }
+
+        bool should_use_uncond_model(const DiffusionParams& diffusion_params) const {
+            return has_uncond_model &&
+                   diffusion_params.context == nullptr &&
+                   diffusion_params.y != nullptr &&
+                   !diffusion_params.y->empty();
+        }
+
+    public:
+        Ideogram4Config config;
+        Ideogram4Transformer model;
+        Ideogram4Transformer uncond_model;
+        bool has_uncond_model = false;
+        std::string uncond_prefix;
+        std::vector<float> pe_vec;
+        std::vector<int32_t> image_indicator_vec;
+
+        Ideogram4Runner(ggml_backend_t backend,
+                        ggml_backend_t params_backend,
+                        const String2TensorStorage& tensor_storage_map = {},
+                        const std::string prefix                       = "")
+            : DiffusionModelRunner(backend, params_backend, prefix),
+              uncond_prefix(prefix + ".uncond") {
+            int64_t detected_layers = detect_num_layers(tensor_storage_map, prefix);
+            if (detected_layers > 0) {
+                config.num_layers = detected_layers;
+            }
+
+            model = Ideogram4Transformer(config);
+            model.init(params_ctx, tensor_storage_map, prefix);
+            for (const auto& pair : tensor_storage_map) {
+                const std::string& name = pair.first;
+                if (starts_with(name, uncond_prefix)) {
+                    has_uncond_model = true;
+                    break;
+                }
+            }
+            if (has_uncond_model) {
+                LOG_DEBUG("using uncond model");
+                uncond_model = Ideogram4Transformer(config);
+                uncond_model.init(params_ctx, tensor_storage_map, uncond_prefix);
+            }
+        }
+
+        std::string get_desc() override {
+            return "ideogram4";
+        }
+
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
+            model.get_param_tensors(tensors, prefix);
+            if (has_uncond_model) {
+                uncond_model.get_param_tensors(tensors, this->uncond_prefix);
+            }
+        }
+
+        ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,
+                                 const sd::Tensor<float>& timesteps_tensor,
+                                 const sd::Tensor<float>& context_tensor,
+                                 bool use_uncond_model = false) {
+            ggml_cgraph* gf        = new_graph_custom(IDEOGRAM4_GRAPH_SIZE);
+            ggml_tensor* x         = make_input(x_tensor);
+            ggml_tensor* timesteps = make_input(timesteps_tensor);
+            GGML_ASSERT(x->ne[3] == 1);
+            Ideogram4Transformer& active_model = use_uncond_model ? uncond_model : model;
+
+            ggml_tensor* context = nullptr;
+            int64_t context_len  = 0;
+            if (!context_tensor.empty()) {
+                context     = make_input(context_tensor);
+                context_len = context->ne[1];
+            }
+
+            int64_t grid_w   = x->ne[0];
+            int64_t grid_h   = x->ne[1];
+            int64_t pos_len  = context_len + grid_h * grid_w;
+            int64_t head_dim = config.emb_dim / config.num_heads;
+
+            pe_vec  = gen_ideogram4_pe(static_cast<int>(grid_h),
+                                       static_cast<int>(grid_w),
+                                       static_cast<int>(x->ne[3]),
+                                       static_cast<int>(context_len),
+                                       static_cast<int>(head_dim),
+                                       static_cast<int>(config.rope_theta),
+                                       config.mrope_section);
+            auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, head_dim / 2, pos_len);
+            set_backend_tensor_data(pe, pe_vec.data());
+
+            image_indicator_vec.assign(static_cast<size_t>(pos_len), 1);
+            for (int64_t i = 0; i < context_len; ++i) {
+                image_indicator_vec[static_cast<size_t>(i)] = 0;
+            }
+            auto indicator = ggml_new_tensor_2d(compute_ctx, GGML_TYPE_I32, pos_len, x->ne[3]);
+            set_backend_tensor_data(indicator, image_indicator_vec.data());
+
+            auto runner_ctx  = get_context();
+            ggml_tensor* out = active_model.forward(&runner_ctx, x, timesteps, context, pe, indicator);
+            ggml_build_forward_expand(gf, out);
+            return gf;
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const sd::Tensor<float>& x,
+                                  const sd::Tensor<float>& timesteps,
+                                  const sd::Tensor<float>& context,
+                                  bool use_uncond_model = false) {
+            auto get_graph = [&]() -> ggml_cgraph* {
+                return build_graph(x, timesteps, context, use_uncond_model);
+            };
+            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), x.dim());
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const DiffusionParams& diffusion_params) override {
+            GGML_ASSERT(diffusion_params.x != nullptr);
+            GGML_ASSERT(diffusion_params.timesteps != nullptr);
+            bool use_uncond_model = should_use_uncond_model(diffusion_params);
+            return compute(n_threads,
+                           *diffusion_params.x,
+                           *diffusion_params.timesteps,
+                           tensor_or_empty(diffusion_params.context),
+                           use_uncond_model);
+        }
+    };
+}  // namespace Ideogram4
+
+#endif  // __IDEOGRAM4_HPP__
@@ -4,6 +4,138 @@
 #include "ggml.h"
 #include "tensor.hpp"

+const float ltxav_latent_rgb_proj[128][3] = {
+    {-0.0293802f, -0.0362516f, -0.0291386f},
+    {0.0117735f, 0.0223435f, 0.018856f},
+    {0.00922335f, 0.0145666f, 0.0038772f},
+    {0.0227299f, 0.0109122f, 0.0131384f},
+    {0.00192413f, 0.0024648f, 0.00689245f},
+    {-0.0105576f, -0.0135933f, -0.00873841f},
+    {-0.0310222f, -0.0396358f, -0.0408445f},
+    {0.0149737f, 0.0316323f, 0.03415f},
+    {0.0027752f, 0.00814889f, 0.0108575f},
+    {-0.000678017f, -0.00180589f, -0.0161684f},
+    {0.0153964f, 0.0159774f, 0.0186479f},
+    {-0.0222799f, -0.0202068f, -0.0181082f},
+    {0.0128696f, 0.00754416f, -0.00673279f},
+    {0.0142729f, 0.00448099f, -0.00193934f},
+    {-0.014066f, -0.0193755f, -0.0160104f},
+    {-0.0176785f, -0.015903f, -0.0152621f},
+    {0.0307381f, 0.0292082f, 0.0328668f},
+    {0.0332928f, 0.0368629f, 0.0440893f},
+    {0.0186304f, 0.0124069f, 0.0160734f},
+    {0.00477787f, -0.00315658f, -0.000145702f},
+    {0.0183099f, 0.0122593f, 0.00599732f},
+    {-0.0194551f, -0.0183924f, -0.0147465f},
+    {0.0025732f, 0.00442582f, 0.0173176f},
+    {-0.0169423f, -0.0293863f, -0.0225908f},
+    {-0.021228f, -0.0265094f, -0.0253049f},
+    {0.0327111f, 0.0187133f, 0.0266184f},
+    {-0.0226425f, -0.0313781f, -0.0414356f},
+    {-0.0163142f, -0.0146144f, -0.0171793f},
+    {0.0192183f, 0.0108411f, 0.00829186f},
+    {-0.032246f, -0.0274846f, -0.0287434f},
+    {0.00345399f, 0.0115567f, 0.015288f},
+    {0.000972292f, 0.00331303f, 0.0110501f},
+    {0.000939494f, -0.00705084f, -0.00979449f},
+    {0.0405155f, 0.0339534f, 0.0419513f},
+    {0.0198596f, 0.0186626f, 0.0213766f},
+    {-0.00982375f, -0.00880439f, -0.00470429f},
+    {-0.0313707f, -0.0258098f, -0.0211663f},
+    {0.0144159f, 0.0117896f, 0.0141573f},
+    {0.0164571f, 0.0149178f, 0.00921599f},
+    {0.0436184f, 0.0346583f, 0.0360647f},
+    {-0.00289744f, -0.000752502f, 0.000675415f},
+    {-0.00621715f, -0.000558851f, 0.0135814f},
+    {-0.00817579f, -0.0113584f, -0.00556793f},
+    {0.00965067f, 0.0178221f, 0.015821f},
+    {0.0211832f, 0.0180827f, 0.0154707f},
+    {-0.00412858f, -0.00374182f, 0.0029568f},
+    {-0.0175603f, -0.0226242f, -0.0279012f},
+    {-0.00437471f, -0.00668329f, 0.000164887f},
+    {-0.0355983f, -0.0419093f, -0.0383065f},
+    {0.0144314f, 0.0192514f, 0.0175639f},
+    {-0.0130693f, -0.00569884f, -0.00341647f},
+    {-0.00184689f, 0.00189034f, -0.00190561f},
+    {0.019457f, 0.00842282f, 0.0123738f},
+    {-0.00477146f, -0.00206932f, 0.00283336f},
+    {-0.0364544f, -0.0256141f, -0.0322336f},
+    {-0.0295634f, -0.0295048f, -0.021057f},
+    {0.0144484f, 0.0191862f, 0.0112445f},
+    {0.0536406f, 0.0582376f, 0.0570966f},
+    {0.0085178f, 0.00748455f, 0.00995162f},
+    {-0.0136637f, -0.0172914f, -0.0195978f},
+    {-0.0339128f, -0.0392692f, -0.0355216f},
+    {0.00612855f, 0.00568303f, -0.00212333f},
+    {-0.0029225f, 0.00668819f, 0.0122131f},
+    {0.00841843f, 0.000181587f, -0.00650644f},
+    {-0.00514432f, 0.0127043f, 0.0168049f},
+    {-0.00997384f, -0.00602262f, -0.0164031f},
+    {0.0233226f, 0.033254f, 0.0307266f},
+    {-0.0110201f, -0.0164169f, -0.0161829f},
+    {-0.0195952f, -0.0177943f, -0.0115377f},
+    {-0.00523918f, -0.00452043f, 0.00267397f},
+    {0.0313464f, 0.0288241f, 0.0262496f},
+    {0.0324018f, 0.0339792f, 0.0312209f},
+    {-0.0163247f, -0.0230503f, -0.0263239f},
+    {0.000420577f, -0.00535659f, -0.00663426f},
+    {-0.012897f, -0.00203767f, -0.000622678f},
+    {-0.0632956f, -0.0651325f, -0.0584479f},
+    {-0.00426634f, -0.0150098f, -0.00719348f},
+    {0.00476109f, 0.00674315f, 0.00895472f},
+    {0.0129384f, 0.0158352f, 0.00963773f},
+    {-0.0333379f, -0.0410522f, -0.0317462f},
+    {0.00344054f, 0.00275915f, 0.00355732f},
+    {0.0209062f, 0.0273453f, 0.0222967f},
+    {0.00827287f, 0.00223045f, 0.00325844f},
+    {-0.0149132f, -0.0183973f, -0.0199781f},
+    {-0.0100786f, -0.0103681f, -0.00218224f},
+    {-0.00791409f, -0.00405153f, -0.00599893f},
+    {0.0176126f, 0.00618342f, -6.6569e-05f},
+    {0.00942486f, -0.00206494f, -0.00580324f},
+    {0.00678093f, -0.00291742f, -0.000921195f},
+    {-0.0221992f, -0.00483162f, -0.000848514f},
+    {-0.0151587f, -0.0157166f, -0.0107302f},
+    {0.00909646f, 0.0171985f, 0.0169785f},
+    {0.0127224f, 0.0170612f, 0.0303428f},
+    {0.0196562f, 0.00212451f, 0.0127744f},
+    {0.0233013f, 0.0228994f, 0.0108387f},
+    {0.00520761f, 0.00992992f, 0.0066267f},
+    {-3.77736e-05f, 0.00460229f, -0.00475132f},
+    {-0.0311763f, -0.0453566f, -0.0486901f},
+    {0.0195798f, 0.0281246f, 0.0180102f},
+    {-0.0174149f, -0.0240867f, -0.0188785f},
+    {0.000104658f, 0.00659008f, 0.0144594f},
+    {-0.00311086f, -0.0241426f, -0.0244164f},
+    {0.0336462f, 0.0305173f, 0.0331101f},
+    {0.0613625f, 0.066561f, 0.0610198f},
+    {-0.0286757f, -0.0325401f, -0.0338036f},
+    {0.0141534f, 0.0188266f, 0.0253059f},
+    {-0.00548197f, -0.00170198f, 0.00561745f},
+    {-0.0117872f, -0.00763218f, -0.0145037f},
+    {-0.0253304f, -0.0245217f, -0.0144905f},
+    {-0.00393624f, 0.00350048f, 0.00765561f},
+    {0.0113625f, 0.00561576f, -0.0113672f},
+    {-0.0301278f, -0.0261472f, -0.0301903f},
+    {0.016863f, 0.0173781f, 0.0170916f},
+    {-0.00495108f, 0.00686749f, 0.00282767f},
+    {0.00125409f, -0.00378072f, -0.00264117f},
+    {-0.00264001f, -0.00529772f, -0.0113109f},
+    {-0.054888f, -0.0575461f, -0.0509146f},
+    {-0.019442f, -0.0232916f, -0.0258637f},
+    {0.0133362f, 0.0161808f, 0.00917951f},
+    {-0.0349002f, -0.0372642f, -0.0466206f},
+    {-0.00216926f, 0.00208738f, 0.00766492f},
+    {0.0268528f, 0.0301179f, 0.0228579f},
+    {0.0226176f, 0.021536f, 0.023152f},
+    {-0.0110646f, -0.00511349f, -0.0137346f},
+    {-0.0098424f, -0.00218176f, 0.00414545f},
+    {0.00200216f, 0.00441732f, -0.0136515f},
+    {0.00695946f, 0.00313109f, -0.00379435f},
+    {0.0188377f, 0.0144059f, 0.0229724f},
+};
+float ltxav_latent_rgb_bias[3] = {0.043849f, 0.0201085f, 0.0150286f};
+
 const float wan_21_latent_rgb_proj[16][3] = {
    {0.015123f, -0.148418f, 0.479828f},
    {0.003652f, -0.010680f, -0.037142f},
@@ -0,0 +1,132 @@
+#include "layer_registry.h"
+
+#include <utility>
+
+#include "util.h"
+
+namespace sd::layer_registry {
+
+    void LayerRegistry::register_layer(const std::string& name, ggml_tensor* tensor) {
+        auto& info = layers_[name];
+        info.tensors.push_back(tensor);
+        info.bytes += ggml_nbytes(tensor);
+    }
+
+    bool LayerRegistry::move_layer_to_gpu(const std::string& name) {
+        auto it = layers_.find(name);
+        if (it == layers_.end())
+            return false;
+
+        LayerInfo& info = it->second;
+        if (info.on_gpu)
+            return true;
+        if (gpu_backend_ == nullptr || cpu_backend_ == nullptr) {
+            LOG_ERROR("layer_registry: backends not set; cannot move '%s' to GPU",
+                      name.c_str());
+            return false;
+        }
+        if (info.tensors.empty()) {
+            info.on_gpu = true;
+            return true;
+        }
+
+        // 1. Build a no_alloc context big enough to hold one twin tensor per CPU
+        //    tensor, plus a little overhead.
+        const size_t ctx_size = info.tensors.size() * ggml_tensor_overhead() + 1024;
+        ggml_init_params ctx_params{ctx_size, /*mem_buffer=*/nullptr, /*no_alloc=*/true};
+        ggml_context* twin_ctx = ggml_init(ctx_params);
+        if (twin_ctx == nullptr) {
+            LOG_ERROR("layer_registry: failed to allocate twin context for '%s'",
+                      name.c_str());
+            return false;
+        }
+
+        // 2. Create one GPU twin per CPU tensor. The twin shares the original
+        //    name so any name-based lookup keeps working.
+        std::vector<ggml_tensor*> gpu_twins;
+        gpu_twins.reserve(info.tensors.size());
+        for (ggml_tensor* cpu_t : info.tensors) {
+            ggml_tensor* twin = ggml_dup_tensor(twin_ctx, cpu_t);
+            if (cpu_t->name[0] != '\0') {
+                ggml_set_name(twin, cpu_t->name);
+            }
+            gpu_twins.push_back(twin);
+        }
+
+        // 3. Back the twins with a GPU buffer in one alloc call.
+        ggml_backend_buffer_t gpu_buffer = ggml_backend_alloc_ctx_tensors(twin_ctx, gpu_backend_);
+        if (gpu_buffer == nullptr) {
+            LOG_ERROR("layer_registry: failed to allocate GPU buffer for '%s'",
+                      name.c_str());
+            ggml_free(twin_ctx);
+            return false;
+        }
+
+        // 4. H2D copy + sync.
+        for (size_t i = 0; i < info.tensors.size(); ++i) {
+            ggml_backend_tensor_copy(info.tensors[i], gpu_twins[i]);
+        }
+        ggml_backend_synchronize(gpu_backend_);
+
+        // 5. Swap buffer/data/extra so the originals now point at GPU memory.
+        for (size_t i = 0; i < info.tensors.size(); ++i) {
+            std::swap(info.tensors[i]->buffer, gpu_twins[i]->buffer);
+            std::swap(info.tensors[i]->data, gpu_twins[i]->data);
+            std::swap(info.tensors[i]->extra, gpu_twins[i]->extra);
+        }
+
+        info.gpu_twins  = std::move(gpu_twins);
+        info.twin_ctx   = twin_ctx;
+        info.gpu_buffer = gpu_buffer;
+        info.on_gpu     = true;
+        return true;
+    }
+
+    bool LayerRegistry::move_layer_to_cpu(const std::string& name) {
+        auto it = layers_.find(name);
+        if (it == layers_.end())
+            return false;
+
+        LayerInfo& info = it->second;
+        if (!info.on_gpu)
+            return true;
+        if (info.tensors.size() != info.gpu_twins.size()) {
+            LOG_ERROR("layer_registry: twin/tensor count mismatch for '%s'",
+                      name.c_str());
+            return false;
+        }
+
+        // 1. Swap back: originals point at CPU memory again.
+        for (size_t i = 0; i < info.tensors.size(); ++i) {
+            if (info.gpu_twins[i] == nullptr)
+                continue;
+            std::swap(info.tensors[i]->buffer, info.gpu_twins[i]->buffer);
+            std::swap(info.tensors[i]->data, info.gpu_twins[i]->data);
+            std::swap(info.tensors[i]->extra, info.gpu_twins[i]->extra);
+        }
+
+        // 2. Free the GPU buffer + twin context.
+        if (info.gpu_buffer != nullptr) {
+            ggml_backend_buffer_free(info.gpu_buffer);
+            info.gpu_buffer = nullptr;
+        }
+        if (info.twin_ctx != nullptr) {
+            ggml_free(info.twin_ctx);
+            info.twin_ctx = nullptr;
+        }
+        info.gpu_twins.clear();
+        info.on_gpu = false;
+        return true;
+    }
+
+    bool LayerRegistry::is_layer_on_gpu(const std::string& name) const {
+        auto it = layers_.find(name);
+        return it != layers_.end() && it->second.on_gpu;
+    }
+
+    size_t LayerRegistry::get_layer_size(const std::string& name) const {
+        auto it = layers_.find(name);
+        return it != layers_.end() ? it->second.bytes : 0;
+    }
+
+}  // namespace sd::layer_registry
@@ -0,0 +1,50 @@
+#ifndef __LAYER_REGISTRY_H__
+#define __LAYER_REGISTRY_H__
+
+#include <map>
+#include <set>
+#include <string>
+#include <vector>
+
+#include "ggml-backend.h"
+#include "ggml.h"
+
+namespace sd::layer_registry {
+
+    struct LayerInfo {
+        std::vector<ggml_tensor*> tensors;
+        std::vector<ggml_tensor*> gpu_twins;
+        ggml_context* twin_ctx           = nullptr;
+        ggml_backend_buffer_t gpu_buffer = nullptr;
+        bool on_gpu                      = false;
+        size_t bytes                     = 0;
+    };
+
+    class LayerRegistry {
+    public:
+        LayerRegistry() = default;
+        LayerRegistry(ggml_backend_t gpu_backend, ggml_backend_t cpu_backend)
+            : gpu_backend_(gpu_backend), cpu_backend_(cpu_backend) {}
+
+        void set_backends(ggml_backend_t gpu_backend, ggml_backend_t cpu_backend) {
+            gpu_backend_ = gpu_backend;
+            cpu_backend_ = cpu_backend;
+        }
+        void register_layer(const std::string& name, ggml_tensor* tensor);
+        bool move_layer_to_gpu(const std::string& name);
+        bool move_layer_to_cpu(const std::string& name);
+        bool is_layer_on_gpu(const std::string& name) const;
+        size_t get_layer_size(const std::string& name) const;
+        size_t get_layer_count() const { return layers_.size(); }
+
+        const std::map<std::string, LayerInfo>& layers() const { return layers_; }
+
+    private:
+        ggml_backend_t gpu_backend_ = nullptr;
+        ggml_backend_t cpu_backend_ = nullptr;
+        std::map<std::string, LayerInfo> layers_;
+    };
+
+}  // namespace sd::layer_registry
+
+#endif
@@ -0,0 +1,419 @@
+#ifndef __SD_LENS_HPP__
+#define __SD_LENS_HPP__
+
+#include <memory>
+#include <vector>
+
+#include "common_block.hpp"
+#include "diffusion_model.hpp"
+#include "flux.hpp"
+#include "qwen_image.hpp"
+#include "rope.hpp"
+
+namespace Lens {
+    constexpr int LENS_GRAPH_SIZE = 40960;
+
+    struct LensTimestepProjEmbeddings : public GGMLBlock {
+        LensTimestepProjEmbeddings(int64_t embedding_dim) {
+            blocks["timestep_embedder"] = std::make_shared<Qwen::TimestepEmbedding>(256, embedding_dim);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* timesteps) {
+            auto timestep_embedder = std::dynamic_pointer_cast<Qwen::TimestepEmbedding>(blocks["timestep_embedder"]);
+            auto timesteps_proj    = ggml_ext_timestep_embedding(ctx->ggml_ctx, timesteps, 256, 10000, 1000.f);
+            return timestep_embedder->forward(ctx, timesteps_proj);
+        }
+    };
+
+    struct LensGateMLP : public GGMLBlock {
+        LensGateMLP(int64_t dim, int64_t hidden_dim) {
+            blocks["w1"] = std::make_shared<Linear>(dim, hidden_dim, false);
+            blocks["w2"] = std::make_shared<Linear>(hidden_dim, dim, false);
+            blocks["w3"] = std::make_shared<Linear>(dim, hidden_dim, false);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto w1 = std::dynamic_pointer_cast<Linear>(blocks["w1"]);
+            auto w2 = std::dynamic_pointer_cast<Linear>(blocks["w2"]);
+            auto w3 = std::dynamic_pointer_cast<Linear>(blocks["w3"]);
+
+            auto gate = ggml_silu(ctx->ggml_ctx, w1->forward(ctx, x));
+            auto up   = w3->forward(ctx, x);
+            x         = ggml_mul(ctx->ggml_ctx, gate, up);
+            return w2->forward(ctx, x);
+        }
+    };
+
+    struct LensJointAttention : public GGMLBlock {
+        int64_t dim_head;
+        int64_t num_heads;
+
+        LensJointAttention(int64_t query_dim,
+                           int64_t dim_head,
+                           int64_t num_heads,
+                           float eps = 1e-5f)
+            : dim_head(dim_head), num_heads(num_heads) {
+            int64_t inner_dim = dim_head * num_heads;
+            blocks["img_qkv"] = std::make_shared<Linear>(query_dim, inner_dim * 3, true);
+            blocks["txt_qkv"] = std::make_shared<Linear>(query_dim, inner_dim * 3, true);
+
+            blocks["norm_q"]       = std::make_shared<RMSNorm>(dim_head, eps);
+            blocks["norm_k"]       = std::make_shared<RMSNorm>(dim_head, eps);
+            blocks["norm_added_q"] = std::make_shared<RMSNorm>(dim_head, eps);
+            blocks["norm_added_k"] = std::make_shared<RMSNorm>(dim_head, eps);
+
+            blocks["to_out.0"]   = std::make_shared<Linear>(inner_dim, query_dim, true);
+            blocks["to_add_out"] = std::make_shared<Linear>(inner_dim, query_dim, true);
+        }
+
+        std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx,
+                                                      ggml_tensor* img,
+                                                      ggml_tensor* txt,
+                                                      ggml_tensor* pe,
+                                                      ggml_tensor* mask = nullptr) {
+            auto img_qkv    = std::dynamic_pointer_cast<Linear>(blocks["img_qkv"]);
+            auto txt_qkv    = std::dynamic_pointer_cast<Linear>(blocks["txt_qkv"]);
+            auto norm_q     = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_q"]);
+            auto norm_k     = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_k"]);
+            auto norm_add_q = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_added_q"]);
+            auto norm_add_k = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_added_k"]);
+            auto to_out_0   = std::dynamic_pointer_cast<Linear>(blocks["to_out.0"]);
+            auto to_add_out = std::dynamic_pointer_cast<Linear>(blocks["to_add_out"]);
+            int64_t n_img   = img->ne[1];
+            int64_t n_txt   = txt->ne[1];
+            int64_t N       = img->ne[2];
+            int64_t inner   = dim_head * num_heads;
+
+            auto img_qkv_vec = split_qkv(ctx->ggml_ctx, img_qkv->forward(ctx, img));
+            auto txt_qkv_vec = split_qkv(ctx->ggml_ctx, txt_qkv->forward(ctx, txt));
+
+            auto img_q = ggml_reshape_4d(ctx->ggml_ctx, img_qkv_vec[0], dim_head, num_heads, n_img, N);
+            auto img_k = ggml_reshape_4d(ctx->ggml_ctx, img_qkv_vec[1], dim_head, num_heads, n_img, N);
+            auto img_v = ggml_reshape_4d(ctx->ggml_ctx, img_qkv_vec[2], dim_head, num_heads, n_img, N);
+
+            img_q = norm_q->forward(ctx, img_q);
+            img_k = norm_k->forward(ctx, img_k);
+
+            auto txt_q = ggml_reshape_4d(ctx->ggml_ctx, txt_qkv_vec[0], dim_head, num_heads, n_txt, N);
+            auto txt_k = ggml_reshape_4d(ctx->ggml_ctx, txt_qkv_vec[1], dim_head, num_heads, n_txt, N);
+            auto txt_v = ggml_reshape_4d(ctx->ggml_ctx, txt_qkv_vec[2], dim_head, num_heads, n_txt, N);
+
+            txt_q = norm_add_q->forward(ctx, txt_q);
+            txt_k = norm_add_k->forward(ctx, txt_k);
+
+            auto q = ggml_concat(ctx->ggml_ctx, img_q, txt_q, 2);
+            auto k = ggml_concat(ctx->ggml_ctx, img_k, txt_k, 2);
+            auto v = ggml_concat(ctx->ggml_ctx, img_v, txt_v, 2);
+
+            auto attn = Rope::attention(ctx, q, k, v, pe, mask, (1.0f / 128.f));
+
+            auto img_attn_out = ggml_view_3d(ctx->ggml_ctx,
+                                             attn,
+                                             inner,
+                                             n_img,
+                                             N,
+                                             attn->nb[1],
+                                             attn->nb[2],
+                                             0);
+            auto txt_attn_out = ggml_view_3d(ctx->ggml_ctx,
+                                             attn,
+                                             inner,
+                                             n_txt,
+                                             N,
+                                             attn->nb[1],
+                                             attn->nb[2],
+                                             n_img * attn->nb[1]);
+
+            img_attn_out = to_out_0->forward(ctx, ggml_cont(ctx->ggml_ctx, img_attn_out));
+            txt_attn_out = to_add_out->forward(ctx, ggml_cont(ctx->ggml_ctx, txt_attn_out));
+            return {img_attn_out, txt_attn_out};
+        }
+    };
+
+    struct LensTransformerBlock : public GGMLBlock {
+        LensTransformerBlock(int64_t dim,
+                             int64_t num_attention_heads,
+                             int64_t attention_head_dim,
+                             float eps = 1e-6f) {
+            int64_t mlp_hidden_dim = dim / 3 * 8;
+            blocks["img_mod.1"]    = std::make_shared<Linear>(dim, 6 * dim, true);
+            blocks["txt_mod.1"]    = std::make_shared<Linear>(dim, 6 * dim, true);
+            blocks["img_norm1"]    = std::make_shared<RMSNorm>(dim, eps);
+            blocks["img_norm2"]    = std::make_shared<RMSNorm>(dim, eps);
+            blocks["txt_norm1"]    = std::make_shared<RMSNorm>(dim, eps);
+            blocks["txt_norm2"]    = std::make_shared<RMSNorm>(dim, eps);
+            blocks["img_mlp"]      = std::make_shared<LensGateMLP>(dim, mlp_hidden_dim);
+            blocks["txt_mlp"]      = std::make_shared<LensGateMLP>(dim, mlp_hidden_dim);
+            blocks["attn"]         = std::make_shared<LensJointAttention>(dim, attention_head_dim, num_attention_heads);
+        }
+
+        std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx,
+                                                      ggml_tensor* img,
+                                                      ggml_tensor* txt,
+                                                      ggml_tensor* t_emb,
+                                                      ggml_tensor* pe) {
+            auto img_mod_1 = std::dynamic_pointer_cast<Linear>(blocks["img_mod.1"]);
+            auto txt_mod_1 = std::dynamic_pointer_cast<Linear>(blocks["txt_mod.1"]);
+            auto img_norm1 = std::dynamic_pointer_cast<RMSNorm>(blocks["img_norm1"]);
+            auto img_norm2 = std::dynamic_pointer_cast<RMSNorm>(blocks["img_norm2"]);
+            auto txt_norm1 = std::dynamic_pointer_cast<RMSNorm>(blocks["txt_norm1"]);
+            auto txt_norm2 = std::dynamic_pointer_cast<RMSNorm>(blocks["txt_norm2"]);
+            auto img_mlp   = std::dynamic_pointer_cast<LensGateMLP>(blocks["img_mlp"]);
+            auto txt_mlp   = std::dynamic_pointer_cast<LensGateMLP>(blocks["txt_mlp"]);
+            auto attn      = std::dynamic_pointer_cast<LensJointAttention>(blocks["attn"]);
+
+            auto temb = ggml_silu(ctx->ggml_ctx, t_emb);
+
+            auto img_mod_params = img_mod_1->forward(ctx, temb);
+            auto img_mod_vec    = ggml_ext_chunk(ctx->ggml_ctx, img_mod_params, 6, 0);
+            auto txt_mod_params = txt_mod_1->forward(ctx, temb);
+            auto txt_mod_vec    = ggml_ext_chunk(ctx->ggml_ctx, txt_mod_params, 6, 0);
+
+            auto img_normed    = img_norm1->forward(ctx, img);
+            auto img_modulated = Flux::modulate(ctx->ggml_ctx, img_normed, img_mod_vec[0], img_mod_vec[1]);
+            auto txt_normed    = txt_norm1->forward(ctx, txt);
+            auto txt_modulated = Flux::modulate(ctx->ggml_ctx, txt_normed, txt_mod_vec[0], txt_mod_vec[1]);
+
+            auto [img_attn_output, txt_attn_output] = attn->forward(ctx, img_modulated, txt_modulated, pe);
+
+            img = ggml_add(ctx->ggml_ctx, img, ggml_mul(ctx->ggml_ctx, img_attn_output, img_mod_vec[2]));
+            txt = ggml_add(ctx->ggml_ctx, txt, ggml_mul(ctx->ggml_ctx, txt_attn_output, txt_mod_vec[2]));
+
+            auto img_normed2    = img_norm2->forward(ctx, img);
+            auto img_modulated2 = Flux::modulate(ctx->ggml_ctx, img_normed2, img_mod_vec[3], img_mod_vec[4]);
+            auto txt_normed2    = txt_norm2->forward(ctx, txt);
+            auto txt_modulated2 = Flux::modulate(ctx->ggml_ctx, txt_normed2, txt_mod_vec[3], txt_mod_vec[4]);
+
+            img = ggml_add(ctx->ggml_ctx, img, ggml_mul(ctx->ggml_ctx, img_mlp->forward(ctx, img_modulated2), img_mod_vec[5]));
+            txt = ggml_add(ctx->ggml_ctx, txt, ggml_mul(ctx->ggml_ctx, txt_mlp->forward(ctx, txt_modulated2), txt_mod_vec[5]));
+            return {img, txt};
+        }
+    };
+
+    struct LensAdaLayerNormContinuous : public GGMLBlock {
+        int64_t hidden_size;
+        float eps;
+
+        LensAdaLayerNormContinuous(int64_t hidden_size, float eps = 1e-6f)
+            : hidden_size(hidden_size), eps(eps) {
+            blocks["linear"] = std::make_shared<Linear>(hidden_size, hidden_size * 2, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* conditioning) {
+            auto linear = std::dynamic_pointer_cast<Linear>(blocks["linear"]);
+            auto mods   = ggml_ext_chunk(ctx->ggml_ctx, linear->forward(ctx, ggml_silu(ctx->ggml_ctx, conditioning)), 2, 0);
+            auto scale  = mods[0];
+            auto shift  = mods[1];
+            x           = ggml_norm(ctx->ggml_ctx, x, eps);
+            return Flux::modulate(ctx->ggml_ctx, x, shift, scale);
+        }
+    };
+
+    struct LensParams {
+        int patch_size              = 2;
+        int64_t in_channels         = 128;
+        int64_t out_channels        = 32;
+        int num_layers              = 48;
+        int64_t attention_head_dim  = 64;
+        int64_t num_attention_heads = 24;
+        int64_t joint_attention_dim = 2880;
+        int selected_layer_count    = 4;
+        int theta                   = 10000;
+        std::vector<int> axes_dim   = {8, 28, 28};
+        int axes_dim_sum            = 64;
+    };
+
+    class LensModel : public GGMLBlock {
+    public:
+        LensParams params;
+
+        LensModel() = default;
+        LensModel(LensParams params)
+            : params(params) {
+            int64_t inner_dim         = params.num_attention_heads * params.attention_head_dim;
+            blocks["time_text_embed"] = std::make_shared<LensTimestepProjEmbeddings>(inner_dim);
+            blocks["img_in"]          = std::make_shared<Linear>(params.in_channels, inner_dim, true);
+            blocks["txt_in"]          = std::make_shared<Linear>(params.joint_attention_dim * params.selected_layer_count, inner_dim, true);
+            for (int i = 0; i < params.selected_layer_count; ++i) {
+                blocks["txt_norm." + std::to_string(i)] = std::make_shared<RMSNorm>(params.joint_attention_dim, 1e-5f);
+            }
+            for (int i = 0; i < params.num_layers; ++i) {
+                blocks["transformer_blocks." + std::to_string(i)] = std::make_shared<LensTransformerBlock>(inner_dim,
+                                                                                                           params.num_attention_heads,
+                                                                                                           params.attention_head_dim);
+            }
+            blocks["norm_out"] = std::make_shared<LensAdaLayerNormContinuous>(inner_dim, 1e-6f);
+            blocks["proj_out"] = std::make_shared<Linear>(inner_dim, params.patch_size * params.patch_size * params.out_channels, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* timestep,
+                             ggml_tensor* context,
+                             ggml_tensor* pe) {
+            GGML_ASSERT(context != nullptr);
+            int64_t W = x->ne[0];
+            int64_t H = x->ne[1];
+            int64_t C = x->ne[2];
+            int64_t N = x->ne[3];
+
+            auto time_text_embed = std::dynamic_pointer_cast<LensTimestepProjEmbeddings>(blocks["time_text_embed"]);
+            auto img_in          = std::dynamic_pointer_cast<Linear>(blocks["img_in"]);
+            auto txt_in          = std::dynamic_pointer_cast<Linear>(blocks["txt_in"]);
+            auto norm_out        = std::dynamic_pointer_cast<LensAdaLayerNormContinuous>(blocks["norm_out"]);
+            auto proj_out        = std::dynamic_pointer_cast<Linear>(blocks["proj_out"]);
+
+            auto t_emb = time_text_embed->forward(ctx, timestep);
+
+            auto img = ggml_reshape_3d(ctx->ggml_ctx, x, W * H, C, N);
+            img      = ggml_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, img, 1, 0, 2, 3));
+            img      = img_in->forward(ctx, img);
+
+            std::vector<ggml_tensor*> txt_chunks = ggml_ext_chunk(ctx->ggml_ctx, context, params.selected_layer_count, 0);
+            ggml_tensor* txt                     = nullptr;
+            for (int i = 0; i < params.selected_layer_count; ++i) {
+                auto txt_norm = std::dynamic_pointer_cast<RMSNorm>(blocks["txt_norm." + std::to_string(i)]);
+                auto chunk    = txt_norm->forward(ctx, txt_chunks[i]);
+                txt           = txt == nullptr ? chunk : ggml_concat(ctx->ggml_ctx, txt, chunk, 0);
+            }
+            txt = txt_in->forward(ctx, txt);
+
+            sd::ggml_graph_cut::mark_graph_cut(img, "lens.prelude", "img");
+            sd::ggml_graph_cut::mark_graph_cut(txt, "lens.prelude", "txt");
+
+            for (int i = 0; i < params.num_layers; ++i) {
+                auto block = std::dynamic_pointer_cast<LensTransformerBlock>(blocks["transformer_blocks." + std::to_string(i)]);
+                auto out   = block->forward(ctx, img, txt, t_emb, pe);
+                img        = out.first;
+                txt        = out.second;
+                sd::ggml_graph_cut::mark_graph_cut(img, "lens.transformer_blocks." + std::to_string(i), "img");
+                sd::ggml_graph_cut::mark_graph_cut(txt, "lens.transformer_blocks." + std::to_string(i), "txt");
+            }
+
+            img = norm_out->forward(ctx, img, t_emb);
+            img = proj_out->forward(ctx, img);
+
+            auto out = ggml_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, img, 1, 0, 2, 3));
+            out      = ggml_reshape_4d(ctx->ggml_ctx, out, W, H, params.patch_size * params.patch_size * params.out_channels, N);
+            return out;
+        }
+    };
+
+    struct LensRunner : public DiffusionModelRunner {
+        LensParams lens_params;
+        LensModel lens;
+        std::vector<float> pe_vec;
+
+        LensRunner(ggml_backend_t backend,
+                   ggml_backend_t params_backend,
+                   const String2TensorStorage& tensor_storage_map = {},
+                   const std::string prefix                       = "")
+            : DiffusionModelRunner(backend, params_backend, prefix) {
+            lens_params.num_layers = 0;
+            for (const auto& [name, tensor_storage] : tensor_storage_map) {
+                if (!starts_with(name, prefix)) {
+                    continue;
+                }
+                if (ends_with(name, "img_in.weight") && tensor_storage.n_dims == 2) {
+                    lens_params.in_channels         = tensor_storage.ne[0];
+                    int64_t inner_dim               = tensor_storage.ne[1];
+                    lens_params.num_attention_heads = inner_dim / lens_params.attention_head_dim;
+                } else if (ends_with(name, "txt_in.weight") && tensor_storage.n_dims == 2) {
+                    lens_params.selected_layer_count = static_cast<int>(tensor_storage.ne[0] / lens_params.joint_attention_dim);
+                } else if (ends_with(name, "proj_out.weight") && tensor_storage.n_dims == 2) {
+                    lens_params.out_channels = tensor_storage.ne[1] / lens_params.patch_size / lens_params.patch_size;
+                } else if (ends_with(name, "transformer_blocks.0.attn.norm_q.weight") && tensor_storage.n_dims == 1) {
+                    lens_params.attention_head_dim = tensor_storage.ne[0];
+                }
+
+                size_t pos = name.find("transformer_blocks.");
+                if (pos != std::string::npos) {
+                    std::string layer_name = name.substr(pos);
+                    auto items             = split_string(layer_name, '.');
+                    if (items.size() > 1) {
+                        int block_index = atoi(items[1].c_str());
+                        if (block_index + 1 > lens_params.num_layers) {
+                            lens_params.num_layers = block_index + 1;
+                        }
+                    }
+                }
+            }
+            if (lens_params.num_layers == 0) {
+                lens_params.num_layers = 48;
+            }
+            lens_params.axes_dim_sum = 0;
+            for (int axis_dim : lens_params.axes_dim) {
+                lens_params.axes_dim_sum += axis_dim;
+            }
+
+            LOG_INFO("lens: layers = %d, in_channels = %" PRId64 ", out_channels = %" PRId64
+                     ", heads = %" PRId64 ", head_dim = %" PRId64,
+                     lens_params.num_layers,
+                     lens_params.in_channels,
+                     lens_params.out_channels,
+                     lens_params.num_attention_heads,
+                     lens_params.attention_head_dim);
+
+            lens = LensModel(lens_params);
+            lens.init(params_ctx, tensor_storage_map, prefix);
+        }
+
+        std::string get_desc() override {
+            return "lens";
+        }
+
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
+            lens.get_param_tensors(tensors, prefix);
+        }
+
+        ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,
+                                 const sd::Tensor<float>& timesteps_tensor,
+                                 const sd::Tensor<float>& context_tensor) {
+            ggml_cgraph* gf        = new_graph_custom(LENS_GRAPH_SIZE);
+            ggml_tensor* x         = make_input(x_tensor);
+            ggml_tensor* timesteps = make_input(timesteps_tensor);
+            GGML_ASSERT(x->ne[3] == 1);
+            GGML_ASSERT(!context_tensor.empty());
+            ggml_tensor* context = make_input(context_tensor);
+
+            pe_vec      = Rope::gen_lens_pe(static_cast<int>(x->ne[1]),
+                                            static_cast<int>(x->ne[0]),
+                                            static_cast<int>(x->ne[3]),
+                                            static_cast<int>(context->ne[1]),
+                                            lens_params.theta,
+                                            circular_y_enabled,
+                                            circular_x_enabled,
+                                            lens_params.axes_dim);
+            int pos_len = static_cast<int>(pe_vec.size() / lens_params.axes_dim_sum / 2);
+            auto pe     = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, lens_params.axes_dim_sum / 2, pos_len);
+            set_backend_tensor_data(pe, pe_vec.data());
+
+            auto runner_ctx  = get_context();
+            ggml_tensor* out = lens.forward(&runner_ctx, x, timesteps, context, pe);
+            ggml_build_forward_expand(gf, out);
+            return gf;
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const sd::Tensor<float>& x,
+                                  const sd::Tensor<float>& timesteps,
+                                  const sd::Tensor<float>& context) {
+            auto get_graph = [&]() -> ggml_cgraph* {
+                return build_graph(x, timesteps, context);
+            };
+            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), x.dim());
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const DiffusionParams& diffusion_params) override {
+            GGML_ASSERT(diffusion_params.x != nullptr);
+            GGML_ASSERT(diffusion_params.timesteps != nullptr);
+            return compute(n_threads,
+                           *diffusion_params.x,
+                           *diffusion_params.timesteps,
+                           tensor_or_empty(diffusion_params.context));
+        }
+    };
+}  // namespace Lens
+
+#endif  // __SD_LENS_HPP__
@@ -22,10 +22,11 @@ struct LoraModel : public GGMLRunner {

    LoraModel(const std::string& lora_id,
              ggml_backend_t backend,
+              ggml_backend_t params_backend,
              const std::string& file_path = "",
              std::string prefix           = "",
              SDVersion version            = VERSION_COUNT)
-        : lora_id(lora_id), file_path(file_path), GGMLRunner(backend, false) {
+        : lora_id(lora_id), file_path(file_path), GGMLRunner(backend, params_backend) {
        prefix = "lora." + prefix;
        if (!model_loader.init_from_file_and_convert_name(file_path, prefix, version)) {
            load_failed = true;
@@ -85,7 +86,10 @@ struct LoraModel : public GGMLRunner {
            lora_tensors[name] = real;
        }

-        alloc_params_buffer();
+        if (!alloc_params_buffer()) {
+            LOG_ERROR("lora model buffer allocation failed");
+            return false;
+        }

        dry_run = false;
        model_loader.load_tensors(on_new_tensor_cb, n_threads);
@@ -768,7 +772,7 @@ struct LoraModel : public GGMLRunner {
            }

            ggml_tensor* original_tensor = model_tensor;
-            if (!ggml_backend_is_cpu(runtime_backend) && ggml_backend_buffer_is_host(original_tensor->buffer)) {
+            if (!sd_backend_is_cpu(runtime_backend) && ggml_backend_buffer_is_host(original_tensor->buffer)) {
                model_tensor = ggml_dup_tensor(compute_ctx, model_tensor);
                set_backend_tensor_data(model_tensor, original_tensor->data);
            }
@@ -782,7 +786,7 @@ struct LoraModel : public GGMLRunner {
                final_tensor = ggml_add_inplace(compute_ctx, model_tensor, diff);
            }
            ggml_build_forward_expand(gf, final_tensor);
-            if (!ggml_backend_is_cpu(runtime_backend) && ggml_backend_buffer_is_host(original_tensor->buffer)) {
+            if (!sd_backend_is_cpu(runtime_backend) && ggml_backend_buffer_is_host(original_tensor->buffer)) {
                original_tensor_to_final_tensor[original_tensor] = final_tensor;
            }
        }
@@ -0,0 +1,551 @@
+#ifndef __SD_LTX_LATENT_UPSCALER_HPP__
+#define __SD_LTX_LATENT_UPSCALER_HPP__
+
+#include <cinttypes>
+#include <cmath>
+#include <cstdlib>
+#include <map>
+#include <memory>
+#include <set>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "common_dit.hpp"
+#include "ggml_extend.hpp"
+#include "ggml_graph_cut.h"
+#include "model.h"
+#include "util.h"
+
+namespace LTXVUpsampler {
+    constexpr int LTX_UPSAMPLER_GRAPH_SIZE = 10240;
+
+    struct LatentUpsamplerConfig {
+        int64_t in_channels      = 128;
+        int64_t mid_channels     = 1024;
+        int num_blocks_per_stage = 4;
+        int dims                 = 3;
+        bool spatial_upsample    = true;
+        bool temporal_upsample   = false;
+        bool rational_resampler  = false;
+        float spatial_scale      = 2.f;
+        int spatial_up_num       = 2;
+        int spatial_down_den     = 1;
+        int temporal_up_factor   = 1;
+    };
+
+    static inline bool has_tensor(const String2TensorStorage& tensor_storage_map,
+                                  const std::string& name) {
+        return tensor_storage_map.find(name) != tensor_storage_map.end();
+    }
+
+    static inline int64_t get_tensor_ne(const String2TensorStorage& tensor_storage_map,
+                                        const std::string& name,
+                                        int axis,
+                                        int64_t fallback) {
+        auto it = tensor_storage_map.find(name);
+        if (it == tensor_storage_map.end() || axis < 0 || axis >= GGML_MAX_DIMS) {
+            return fallback;
+        }
+        return it->second.ne[axis];
+    }
+
+    static inline int64_t get_tensor_ne0(const String2TensorStorage& tensor_storage_map,
+                                         const std::string& name,
+                                         int64_t fallback) {
+        return get_tensor_ne(tensor_storage_map, name, 0, fallback);
+    }
+
+    static inline int count_module_blocks(const String2TensorStorage& tensor_storage_map,
+                                          const std::string& module_name) {
+        int max_block            = -1;
+        const std::string prefix = module_name + ".";
+        for (const auto& pair : tensor_storage_map) {
+            const std::string& name = pair.first;
+            if (name.find(prefix) != 0) {
+                continue;
+            }
+            size_t begin = prefix.size();
+            size_t end   = name.find('.', begin);
+            if (end == std::string::npos) {
+                continue;
+            }
+            int index = atoi(name.substr(begin, end - begin).c_str());
+            max_block = std::max(max_block, index);
+        }
+        return max_block + 1;
+    }
+
+    static inline LatentUpsamplerConfig detect_config_from_weights(const String2TensorStorage& tensor_storage_map) {
+        LatentUpsamplerConfig config;
+        config.mid_channels = get_tensor_ne0(tensor_storage_map, "initial_norm.weight", config.mid_channels);
+        config.in_channels  = get_tensor_ne0(tensor_storage_map, "final_conv.bias", config.in_channels);
+        int detected_blocks = count_module_blocks(tensor_storage_map, "res_blocks");
+        if (detected_blocks > 0) {
+            config.num_blocks_per_stage = detected_blocks;
+        }
+        config.rational_resampler      = has_tensor(tensor_storage_map, "upsampler.conv.weight");
+        int64_t upsampler_out_channels = get_tensor_ne0(tensor_storage_map, "upsampler.0.bias", 0);
+        config.spatial_upsample        = config.rational_resampler || upsampler_out_channels == 4 * config.mid_channels;
+        config.temporal_upsample       = upsampler_out_channels == 2 * config.mid_channels;
+        if (config.temporal_upsample) {
+            config.temporal_up_factor = 2;
+        }
+        if (config.rational_resampler) {
+            int64_t out_channels = get_tensor_ne(tensor_storage_map,
+                                                 "upsampler.conv.weight",
+                                                 3,
+                                                 config.mid_channels * 9);
+            if (config.mid_channels > 0 && out_channels % config.mid_channels == 0) {
+                int64_t ratio = out_channels / config.mid_channels;
+                int num       = static_cast<int>(std::round(std::sqrt(static_cast<double>(ratio))));
+                if (num > 0 && static_cast<int64_t>(num) * num == ratio) {
+                    config.spatial_up_num = num;
+                }
+            }
+            if (config.spatial_up_num == 3) {
+                config.spatial_down_den = 2;
+                config.spatial_scale    = 1.5f;
+            } else if (config.spatial_up_num == 4) {
+                config.spatial_down_den = 1;
+                config.spatial_scale    = 4.f;
+            } else {
+                config.spatial_down_den = 1;
+                config.spatial_scale    = static_cast<float>(config.spatial_up_num);
+            }
+        }
+        return config;
+    }
+
+    class VideoGroupNorm : public GGMLBlock {
+    protected:
+        int num_groups;
+        int64_t num_channels;
+        float eps;
+        std::string prefix;
+
+        void init_params(ggml_context* ctx,
+                         const String2TensorStorage& tensor_storage_map = {},
+                         const std::string prefix                       = "") override {
+            SD_UNUSED(tensor_storage_map);
+            this->prefix     = prefix;
+            params["weight"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, num_channels);
+            params["bias"]   = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, num_channels);
+        }
+
+    public:
+        VideoGroupNorm(int num_groups, int64_t num_channels, float eps = 1e-05f)
+            : num_groups(num_groups),
+              num_channels(num_channels),
+              eps(eps) {}
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            // LTX video latent layout is [W, H, T, C]. ggml_group_norm treats ne[2]
+            // as channels, so fold only H/T internally and restore the same layout.
+            GGML_ASSERT(x->ne[3] == num_channels);
+            const int64_t W = x->ne[0];
+            const int64_t H = x->ne[1];
+            const int64_t T = x->ne[2];
+            x               = ggml_ext_cont(ctx->ggml_ctx, x);
+            x               = ggml_reshape_4d(ctx->ggml_ctx, x, W, H * T, num_channels, 1);
+            x               = ggml_group_norm(ctx->ggml_ctx, x, num_groups, eps);
+
+            ggml_tensor* weight = params["weight"];
+            ggml_tensor* bias   = params["bias"];
+            if (ctx->weight_adapter) {
+                weight = ctx->weight_adapter->patch_weight(ctx->ggml_ctx, ctx->backend, weight, prefix + "weight");
+                bias   = ctx->weight_adapter->patch_weight(ctx->ggml_ctx, ctx->backend, bias, prefix + "bias");
+            }
+            weight = ggml_reshape_4d(ctx->ggml_ctx, weight, 1, 1, num_channels, 1);
+            bias   = ggml_reshape_4d(ctx->ggml_ctx, bias, 1, 1, num_channels, 1);
+            x      = ggml_mul_inplace(ctx->ggml_ctx, x, weight);
+            x      = ggml_add_inplace(ctx->ggml_ctx, x, bias);
+            return ggml_reshape_4d(ctx->ggml_ctx, x, W, H, T, num_channels);
+        }
+    };
+
+    class ResBlock : public GGMLBlock {
+    public:
+        ResBlock(int64_t channels, int dims = 3) {
+            GGML_ASSERT(dims == 3);
+            blocks["conv1"] = std::shared_ptr<GGMLBlock>(new Conv3d(channels, channels, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}));
+            blocks["norm1"] = std::shared_ptr<GGMLBlock>(new VideoGroupNorm(32, channels));
+            blocks["conv2"] = std::shared_ptr<GGMLBlock>(new Conv3d(channels, channels, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}));
+            blocks["norm2"] = std::shared_ptr<GGMLBlock>(new VideoGroupNorm(32, channels));
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto conv1 = std::dynamic_pointer_cast<Conv3d>(blocks["conv1"]);
+            auto norm1 = std::dynamic_pointer_cast<VideoGroupNorm>(blocks["norm1"]);
+            auto conv2 = std::dynamic_pointer_cast<Conv3d>(blocks["conv2"]);
+            auto norm2 = std::dynamic_pointer_cast<VideoGroupNorm>(blocks["norm2"]);
+
+            ggml_tensor* residual = x;
+
+            x = conv1->forward(ctx, x);
+            x = norm1->forward(ctx, x);
+            x = ggml_silu_inplace(ctx->ggml_ctx, x);
+            x = conv2->forward(ctx, x);
+            x = norm2->forward(ctx, x);
+            x = ggml_add(ctx->ggml_ctx, x, residual);
+            return ggml_silu(ctx->ggml_ctx, x);
+        }
+    };
+
+    class PixelShuffleND : public UnaryBlock {
+    protected:
+        int upscale_factor;
+
+    public:
+        explicit PixelShuffleND(int upscale_factor)
+            : upscale_factor(upscale_factor) {}
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
+            GGML_ASSERT(upscale_factor > 0);
+            int64_t h = x->ne[1];
+            int64_t w = x->ne[0];
+            GGML_ASSERT(x->ne[2] % (upscale_factor * upscale_factor) == 0);
+            // x: [b*f, c*p1*p2, h, w] -> [b*f, c, h*p1, w*p2]
+            x = ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, x, 2, 0, 1, 3));  // [b*f, h, w, c*p1*p2]
+            x = ggml_reshape_3d(ctx->ggml_ctx, x, x->ne[0], x->ne[1] * x->ne[2], x->ne[3]);          // [b*f, h*w, c*p1*p2]
+            return DiT::unpatchify(ctx->ggml_ctx, x, h, w, upscale_factor, upscale_factor, true);
+        }
+    };
+
+    class TemporalPixelShuffleND : public UnaryBlock {
+    protected:
+        int upscale_factor;
+
+    public:
+        explicit TemporalPixelShuffleND(int upscale_factor)
+            : upscale_factor(upscale_factor) {}
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
+            GGML_ASSERT(upscale_factor > 0);
+            GGML_ASSERT(x->ne[3] % upscale_factor == 0);
+            const int64_t W = x->ne[0];
+            const int64_t H = x->ne[1];
+            const int64_t F = x->ne[2];
+            const int64_t C = x->ne[3] / upscale_factor;
+
+            // x: [b, c*p, f, h, w] -> [b, c, f*p, h, w]
+            x = ggml_ext_cont(ctx->ggml_ctx, x);
+            x = ggml_reshape_4d(ctx->ggml_ctx, x, W * H, F, upscale_factor, C);
+            x = ggml_ext_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 2, 1, 3));
+            return ggml_reshape_4d(ctx->ggml_ctx, x, W, H, F * upscale_factor, C);
+        }
+    };
+
+    class BlurDownsample : public GGMLBlock {
+    protected:
+        int64_t channels;
+        int stride;
+        ggml_tensor* kernel = nullptr;
+        std::vector<float> kernel_data;
+
+        void init_params(ggml_context* ctx,
+                         const String2TensorStorage& tensor_storage_map = {},
+                         const std::string prefix                       = "") override {
+            SD_UNUSED(tensor_storage_map);
+            if (stride == 1) {
+                return;
+            }
+            kernel           = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, 5, 5, 1, channels);
+            std::string name = prefix + "kernel";
+            ggml_set_name(kernel, name.c_str());
+
+            static const float binomial[5] = {1.f, 4.f, 6.f, 4.f, 1.f};
+            kernel_data.resize(static_cast<size_t>(5 * 5 * channels));
+            for (int64_t c = 0; c < channels; ++c) {
+                for (int y = 0; y < 5; ++y) {
+                    for (int x = 0; x < 5; ++x) {
+                        kernel_data[static_cast<size_t>(x + 5 * (y + 5 * c))] =
+                            binomial[y] * binomial[x] / 256.f;
+                    }
+                }
+            }
+        }
+
+    public:
+        BlurDownsample(int64_t channels, int stride)
+            : channels(channels),
+              stride(stride) {
+            GGML_ASSERT(stride >= 1);
+        }
+
+        void load_fixed_tensors() {
+            if (kernel == nullptr || kernel_data.empty()) {
+                return;
+            }
+            ggml_backend_tensor_set(kernel, kernel_data.data(), 0, kernel_data.size() * sizeof(float));
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            if (stride == 1) {
+                return x;
+            }
+            GGML_ASSERT(kernel != nullptr);
+            GGML_ASSERT(x->ne[2] == channels);
+            if (ctx->conv2d_direct_enabled) {
+                return ggml_conv_2d_dw_direct(ctx->ggml_ctx, kernel, x, stride, stride, 2, 2, 1, 1);
+            }
+            return ggml_conv_2d_dw(ctx->ggml_ctx, kernel, x, stride, stride, 2, 2, 1, 1);
+        }
+    };
+
+    class SpatialRationalResampler : public GGMLBlock {
+    protected:
+        int64_t mid_channels;
+        int num;
+        int den;
+
+    public:
+        SpatialRationalResampler(int64_t mid_channels, int num, int den)
+            : mid_channels(mid_channels),
+              num(num),
+              den(den) {
+            GGML_ASSERT(num >= 1);
+            GGML_ASSERT(den >= 1);
+            blocks["conv"]          = std::shared_ptr<GGMLBlock>(new Conv2d(mid_channels, num * num * mid_channels, {3, 3}, {1, 1}, {1, 1}));
+            blocks["pixel_shuffle"] = std::shared_ptr<GGMLBlock>(new PixelShuffleND(num));
+            blocks["blur_down"]     = std::shared_ptr<GGMLBlock>(new BlurDownsample(mid_channels, den));
+        }
+
+        void load_fixed_tensors() {
+            auto blur_down = std::dynamic_pointer_cast<BlurDownsample>(blocks["blur_down"]);
+            blur_down->load_fixed_tensors();
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto conv          = std::dynamic_pointer_cast<Conv2d>(blocks["conv"]);
+            auto pixel_shuffle = std::dynamic_pointer_cast<PixelShuffleND>(blocks["pixel_shuffle"]);
+            auto blur_down     = std::dynamic_pointer_cast<BlurDownsample>(blocks["blur_down"]);
+
+            // rearrange(x, "b c f h w -> (b f) c h w")
+            x = ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, x, 0, 1, 3, 2));
+            x = conv->forward(ctx, x);
+            x = pixel_shuffle->forward(ctx, x);
+            x = blur_down->forward(ctx, x);
+            return ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, x, 0, 1, 3, 2));
+        }
+    };
+
+    class LatentUpsampler : public GGMLBlock {
+    public:
+        LatentUpsamplerConfig config;
+
+        explicit LatentUpsampler(LatentUpsamplerConfig config)
+            : config(std::move(config)) {
+            GGML_ASSERT(this->config.dims == 3);
+            GGML_ASSERT(this->config.spatial_upsample || this->config.temporal_upsample);
+
+            blocks["initial_conv"] = std::shared_ptr<GGMLBlock>(new Conv3d(this->config.in_channels,
+                                                                           this->config.mid_channels,
+                                                                           {3, 3, 3},
+                                                                           {1, 1, 1},
+                                                                           {1, 1, 1}));
+            blocks["initial_norm"] = std::shared_ptr<GGMLBlock>(new VideoGroupNorm(32, this->config.mid_channels));
+            for (int i = 0; i < this->config.num_blocks_per_stage; ++i) {
+                blocks["res_blocks." + std::to_string(i)] = std::shared_ptr<GGMLBlock>(new ResBlock(this->config.mid_channels, this->config.dims));
+            }
+            if (this->config.rational_resampler) {
+                blocks["upsampler"] = std::shared_ptr<GGMLBlock>(new SpatialRationalResampler(this->config.mid_channels,
+                                                                                              this->config.spatial_up_num,
+                                                                                              this->config.spatial_down_den));
+            } else if (this->config.temporal_upsample) {
+                blocks["upsampler.0"] = std::shared_ptr<GGMLBlock>(new Conv3d(this->config.mid_channels,
+                                                                              this->config.temporal_up_factor * this->config.mid_channels,
+                                                                              {3, 3, 3},
+                                                                              {1, 1, 1},
+                                                                              {1, 1, 1}));
+                blocks["upsampler.1"] = std::shared_ptr<GGMLBlock>(new TemporalPixelShuffleND(this->config.temporal_up_factor));
+            } else {
+                blocks["upsampler.0"] = std::shared_ptr<GGMLBlock>(new Conv2d(this->config.mid_channels,
+                                                                              4 * this->config.mid_channels,
+                                                                              {3, 3},
+                                                                              {1, 1},
+                                                                              {1, 1}));
+                blocks["upsampler.1"] = std::shared_ptr<GGMLBlock>(new PixelShuffleND(2));
+            }
+            for (int i = 0; i < this->config.num_blocks_per_stage; ++i) {
+                blocks["post_upsample_res_blocks." + std::to_string(i)] = std::shared_ptr<GGMLBlock>(new ResBlock(this->config.mid_channels, this->config.dims));
+            }
+            blocks["final_conv"] = std::shared_ptr<GGMLBlock>(new Conv3d(this->config.mid_channels,
+                                                                         this->config.in_channels,
+                                                                         {3, 3, 3},
+                                                                         {1, 1, 1},
+                                                                         {1, 1, 1}));
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            // x: [b, c, f, h, w]
+            // return: [b, c, scaled_f, scaled_h, scaled_w]
+            auto initial_conv = std::dynamic_pointer_cast<Conv3d>(blocks["initial_conv"]);
+            auto initial_norm = std::dynamic_pointer_cast<VideoGroupNorm>(blocks["initial_norm"]);
+            auto final_conv   = std::dynamic_pointer_cast<Conv3d>(blocks["final_conv"]);
+
+            x = initial_conv->forward(ctx, x);
+            x = initial_norm->forward(ctx, x);
+            x = ggml_silu(ctx->ggml_ctx, x);
+            sd::ggml_graph_cut::mark_graph_cut(x, "ltx_latent_upsampler.initial", "x");
+
+            for (int i = 0; i < config.num_blocks_per_stage; ++i) {
+                auto block = std::dynamic_pointer_cast<ResBlock>(blocks["res_blocks." + std::to_string(i)]);
+                x          = block->forward(ctx, x);
+                sd::ggml_graph_cut::mark_graph_cut(x, "ltx_latent_upsampler.res_blocks." + std::to_string(i), "x");
+            }
+
+            if (config.rational_resampler) {
+                auto upsampler = std::dynamic_pointer_cast<SpatialRationalResampler>(blocks["upsampler"]);
+                x              = upsampler->forward(ctx, x);
+            } else if (config.temporal_upsample) {
+                auto upsample_conv = std::dynamic_pointer_cast<Conv3d>(blocks["upsampler.0"]);
+                auto pixel_shuffle = std::dynamic_pointer_cast<TemporalPixelShuffleND>(blocks["upsampler.1"]);
+                x                  = upsample_conv->forward(ctx, x);                    // [b, c*2, f, h, w]
+                x                  = pixel_shuffle->forward(ctx, x);                    // [b, c, f*2, h, w]
+                x                  = ggml_ext_slice(ctx->ggml_ctx, x, 2, 1, x->ne[2]);  // x[:, :, 1:, :, :]
+            } else {
+                auto upsample_conv = std::dynamic_pointer_cast<Conv2d>(blocks["upsampler.0"]);
+                auto pixel_shuffle = std::dynamic_pointer_cast<PixelShuffleND>(blocks["upsampler.1"]);
+
+                // rearrange(x, "b c f h w -> (b f) c h w"),
+                x = ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, x, 0, 1, 3, 2));  // [b*f, c, h, w]
+                x = upsample_conv->forward(ctx, x);                                                      // [b*f, c*4, h, w]
+                x = pixel_shuffle->forward(ctx, x);                                                      // [b*f, c, h*2, w*2]
+                x = ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, x, 0, 1, 3, 2));  // [b*c, f, h, w]
+            }
+            sd::ggml_graph_cut::mark_graph_cut(x, "ltx_latent_upsampler.spatial_up", "x");
+
+            for (int i = 0; i < config.num_blocks_per_stage; ++i) {
+                auto block = std::dynamic_pointer_cast<ResBlock>(blocks["post_upsample_res_blocks." + std::to_string(i)]);
+                x          = block->forward(ctx, x);
+                sd::ggml_graph_cut::mark_graph_cut(x, "ltx_latent_upsampler.post_blocks." + std::to_string(i), "x");
+            }
+
+            x = final_conv->forward(ctx, x);
+            sd::ggml_graph_cut::mark_graph_cut(x, "ltx_latent_upsampler.final", "x");
+            return x;
+        }
+
+        void load_fixed_tensors() {
+            if (!config.rational_resampler) {
+                return;
+            }
+            auto upsampler = std::dynamic_pointer_cast<SpatialRationalResampler>(blocks["upsampler"]);
+            upsampler->load_fixed_tensors();
+        }
+    };
+
+    struct LatentUpsamplerRunner : public GGMLRunner {
+        std::unique_ptr<LatentUpsampler> model;
+
+        LatentUpsamplerRunner(ggml_backend_t backend,
+                              ggml_backend_t params_backend)
+            : GGMLRunner(backend, params_backend) {}
+
+        std::string get_desc() override {
+            return "ltx_latent_upsampler";
+        }
+
+        bool load_from_file(const std::string& file_path, int n_threads) {
+            LOG_INFO("loading LTX latent upsampler from '%s'", file_path.c_str());
+            ModelLoader model_loader;
+            if (!model_loader.init_from_file(file_path)) {
+                LOG_ERROR("init LTX latent upsampler model loader from file failed: '%s'", file_path.c_str());
+                return false;
+            }
+
+            const auto& tensor_storage_map = model_loader.get_tensor_storage_map();
+            bool has_regular_upsampler     = has_tensor(tensor_storage_map, "upsampler.0.weight");
+            bool has_rational_spatial      = has_tensor(tensor_storage_map, "upsampler.conv.weight");
+            if (!has_tensor(tensor_storage_map, "post_upsample_res_blocks.0.conv2.bias") ||
+                (!has_regular_upsampler && !has_rational_spatial)) {
+                LOG_ERROR("unsupported LTX latent upsampler weights: expected upsampler tensors");
+                return false;
+            }
+
+            LatentUpsamplerConfig config = detect_config_from_weights(tensor_storage_map);
+            if (config.dims != 3 || (!config.spatial_upsample && !config.temporal_upsample) ||
+                config.spatial_up_num < 1 || config.spatial_down_den < 1 || config.temporal_up_factor < 1) {
+                LOG_ERROR("unsupported LTX latent upsampler config: dims=%d spatial=%d temporal=%d rational=%d scale=%.3f temporal_factor=%d",
+                          config.dims,
+                          config.spatial_upsample,
+                          config.temporal_upsample,
+                          config.rational_resampler,
+                          config.spatial_scale,
+                          config.temporal_up_factor);
+                return false;
+            }
+
+            model = std::make_unique<LatentUpsampler>(config);
+            model->init(params_ctx, tensor_storage_map, "");
+            if (!alloc_params_buffer()) {
+                LOG_ERROR("LTX latent upsampler params buffer allocation failed");
+                return false;
+            }
+
+            std::map<std::string, ggml_tensor*> tensors;
+            model->get_param_tensors(tensors);
+            std::set<std::string> ignore_tensors;
+            if (config.rational_resampler) {
+                ignore_tensors.insert("upsampler.blur_down.kernel");
+            }
+            if (!model_loader.load_tensors(tensors, ignore_tensors, n_threads)) {
+                LOG_ERROR("load LTX latent upsampler tensors failed");
+                return false;
+            }
+            model->load_fixed_tensors();
+
+            LOG_INFO("LTX latent upsampler loaded: in_channels=%" PRId64 ", mid_channels=%" PRId64 ", blocks=%d, scale=%.3f, temporal_factor=%d, rational=%d",
+                     config.in_channels,
+                     config.mid_channels,
+                     config.num_blocks_per_stage,
+                     config.spatial_scale,
+                     config.temporal_up_factor,
+                     config.rational_resampler);
+            return true;
+        }
+
+        ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor) {
+            if (!model) {
+                return nullptr;
+            }
+            ggml_cgraph* gf  = new_graph_custom(LTX_UPSAMPLER_GRAPH_SIZE);
+            ggml_tensor* x   = make_input(x_tensor);
+            auto runner_ctx  = get_context();
+            ggml_tensor* out = model->forward(&runner_ctx, x);
+            ggml_build_forward_expand(gf, out);
+            return gf;
+        }
+
+        sd::Tensor<float> compute(const int n_threads,
+                                  const sd::Tensor<float>& x) {
+            if (!model) {
+                LOG_ERROR("LTX latent upsampler is not loaded");
+                return {};
+            }
+            if (x.dim() != 4 && x.dim() != 5) {
+                LOG_ERROR("LTX latent upsampler expects 4D or 5D video latent, got dim=%lld",
+                          (long long)x.dim());
+                return {};
+            }
+            if (x.dim() == 5 && x.shape()[4] != 1) {
+                LOG_ERROR("LTX latent upsampler currently supports batch size 1, got batch=%lld",
+                          (long long)x.shape()[4]);
+                return {};
+            }
+            if (x.shape()[3] != model->config.in_channels) {
+                LOG_ERROR("LTX latent upsampler expected %" PRId64 " channels, got %lld",
+                          model->config.in_channels,
+                          (long long)x.shape()[3]);
+                return {};
+            }
+            size_t expected_dim = static_cast<size_t>(x.dim());
+            auto get_graph      = [&]() -> ggml_cgraph* { return build_graph(x); };
+            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), expected_dim);
+        }
+    };
+
+}  // namespace LTXVUpsampler
+
+#endif  // __SD_LTX_LATENT_UPSCALER_HPP__
@@ -3,6 +3,7 @@

 #include <memory>

+#include "diffusion_model.hpp"
 #include "ggml_extend.hpp"
 #include "model.h"

@@ -824,14 +825,14 @@ public:
        return x;
    }
 };
-struct MMDiTRunner : public GGMLRunner {
+struct MMDiTRunner : public DiffusionModelRunner {
    MMDiT mmdit;

    MMDiTRunner(ggml_backend_t backend,
-                bool offload_params_to_cpu,
+                ggml_backend_t params_backend,
                const String2TensorStorage& tensor_storage_map = {},
                const std::string prefix                       = "")
-        : GGMLRunner(backend, offload_params_to_cpu), mmdit(tensor_storage_map) {
+        : DiffusionModelRunner(backend, params_backend, prefix), mmdit(tensor_storage_map) {
        mmdit.init(params_ctx, tensor_storage_map, prefix);
    }

@@ -839,7 +840,7 @@ struct MMDiTRunner : public GGMLRunner {
        return "mmdit";
    }

-    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
        mmdit.get_param_tensors(tensors, prefix);
    }

@@ -885,6 +886,20 @@ struct MMDiTRunner : public GGMLRunner {
        return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), x.dim());
    }

+    sd::Tensor<float> compute(int n_threads,
+                              const DiffusionParams& diffusion_params) override {
+        GGML_ASSERT(diffusion_params.x != nullptr);
+        GGML_ASSERT(diffusion_params.timesteps != nullptr);
+        const auto* extra = diffusion_extra_as<SkipLayerDiffusionExtra>(diffusion_params);
+        static const std::vector<int> empty_skip_layers;
+        return compute(n_threads,
+                       *diffusion_params.x,
+                       *diffusion_params.timesteps,
+                       tensor_or_empty(diffusion_params.context),
+                       tensor_or_empty(diffusion_params.y),
+                       extra->skip_layers ? *extra->skip_layers : empty_skip_layers);
+    }
+
    void test() {
        ggml_init_params params;
        params.mem_size   = static_cast<size_t>(10 * 1024 * 1024);  // 10 MB
@@ -932,13 +947,17 @@ struct MMDiTRunner : public GGMLRunner {

    static void load_from_file_and_test(const std::string& file_path) {
        // ggml_backend_t backend    = ggml_backend_cuda_init(0);
-        ggml_backend_t backend             = ggml_backend_cpu_init();
+        ggml_backend_t backend             = sd_backend_cpu_init();
        ggml_type model_data_type          = GGML_TYPE_F16;
-        std::shared_ptr<MMDiTRunner> mmdit = std::make_shared<MMDiTRunner>(backend, false);
+        std::shared_ptr<MMDiTRunner> mmdit = std::make_shared<MMDiTRunner>(backend, backend);
        {
            LOG_INFO("loading from '%s'", file_path.c_str());

-            mmdit->alloc_params_buffer();
+            if (!mmdit->alloc_params_buffer()) {
+                LOG_ERROR("mmdit embeds buffer allocation failed");
+                return;
+            }
+
            std::map<std::string, ggml_tensor*> tensors;
            mmdit->get_param_tensors(tensors, "model.diffusion_model");

@@ -24,7 +24,7 @@
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
 #include "ggml.h"
-#include "ggml_extend_backend.hpp"
+#include "ggml_extend_backend.h"
 #include "zip.h"

 #include "name_conversion.h"
@@ -410,7 +410,7 @@ bool ModelLoader::init_from_diffusers_file(const std::string& file_path, const s
 }

 SDVersion ModelLoader::get_sd_version() {
-    TensorStorage token_embedding_weight, input_block_weight;
+    TensorStorage token_embedding_weight, input_block_weight, context_ebedding_weight;

    bool has_multiple_encoders = false;
    bool is_unet               = false;
@@ -428,15 +428,30 @@ SDVersion ModelLoader::get_sd_version() {
    bool has_attn_1024               = false;

    for (auto& [name, tensor_storage] : tensor_storage_map) {
-        if (tensor_storage.name.find("model.diffusion_model.double_blocks.") != std::string::npos) {
+        if (tensor_storage.name.find("model.diffusion_model.double_blocks.") != std::string::npos ||
+            tensor_storage.name.find("model.diffusion_model.single_transformer_blocks.") != std::string::npos) {
            is_flux = true;
        }
+        if (tensor_storage.name.find("model.diffusion_model.net.lq_proj.latent_proj.0.weight") != std::string::npos) {
+            return VERSION_PID;
+        }
+        if (tensor_storage.name.find("embed_image_indicator.weight") != std::string::npos) {
+            return VERSION_IDEOGRAM4;
+        }
        if (tensor_storage.name.find("model.diffusion_model.nerf_final_layer_conv.") != std::string::npos) {
            return VERSION_CHROMA_RADIANCE;
        }
        if (tensor_storage.name.find("model.diffusion_model.joint_blocks.") != std::string::npos) {
            return VERSION_SD3;
        }
+        if (tensor_storage.name.find("model.x_embedder.proj1.weight") != std::string::npos &&
+            tensor_storage_map.find("model.language_model.layers.0.self_attn.q_proj.weight") != tensor_storage_map.end()) {
+            return VERSION_HIDREAM_O1;
+        }
+        if (tensor_storage.name.find("model.diffusion_model.transformer_blocks.0.attn.norm_added_q.weight") != std::string::npos &&
+            tensor_storage_map.find("model.diffusion_model.transformer_blocks.0.img_mlp.w1.weight") != tensor_storage_map.end()) {
+            return VERSION_LENS;
+        }
        if (tensor_storage.name.find("model.diffusion_model.transformer_blocks.0.img_mod.1.weight") != std::string::npos) {
            return VERSION_QWEN_IMAGE;
        }
@@ -458,6 +473,9 @@ SDVersion ModelLoader::get_sd_version() {
        if (tensor_storage.name.find("model.diffusion_model.layers.0.adaLN_sa_ln.weight") != std::string::npos) {
            return VERSION_ERNIE_IMAGE;
        }
+        if (tensor_storage.name.find("model.diffusion_model.adaln_single.emb.timestep_embedder.linear_1.bias") != std::string::npos) {
+            return VERSION_LTXAV;
+        }
        if (tensor_storage.name.find("model.diffusion_model.blocks.0.cross_attn.norm_k.weight") != std::string::npos) {
            is_wan = true;
        }
@@ -515,6 +533,9 @@ SDVersion ModelLoader::get_sd_version() {
            tensor_storage.name == "unet.conv_in.weight") {
            input_block_weight = tensor_storage;
        }
+        if (tensor_storage.name == "model.diffusion_model.txt_in.weight" || tensor_storage.name == "model.diffusion_model.context_embedder.weight") {
+            context_ebedding_weight = tensor_storage;
+        }
    }
    if (is_wan) {
        LOG_DEBUG("patch_embedding_channels %d", patch_embedding_channels);
@@ -545,16 +566,20 @@ SDVersion ModelLoader::get_sd_version() {
    }

    if (is_flux && !is_flux2) {
-        if (input_block_weight.ne[0] == 384) {
-            return VERSION_FLUX_FILL;
+        if (context_ebedding_weight.ne[0] == 3584) {
+            return VERSION_LONGCAT;
+        } else {
+            if (input_block_weight.ne[0] == 384) {
+                return VERSION_FLUX_FILL;
+            }
+            if (input_block_weight.ne[0] == 128) {
+                return VERSION_FLUX_CONTROLS;
+            }
+            if (input_block_weight.ne[0] == 196) {
+                return VERSION_FLEX_2;
+            }
+            return VERSION_FLUX;
        }
-        if (input_block_weight.ne[0] == 128) {
-            return VERSION_FLUX_CONTROLS;
-        }
-        if (input_block_weight.ne[0] == 196) {
-            return VERSION_FLEX_2;
-        }
-        return VERSION_FLUX;
    }

    if (is_flux2) {
@@ -730,16 +755,10 @@ void ModelLoader::set_wtype_override(ggml_type wtype, std::string tensor_type_ru
    }
 }

-bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_threads_p, bool enable_mmap) {
-    int64_t process_time_ms = 0;
-    std::atomic<int64_t> read_time_ms(0);
-    std::atomic<int64_t> memcpy_time_ms(0);
-    std::atomic<int64_t> copy_to_backend_time_ms(0);
-    std::atomic<int64_t> convert_time_ms(0);
-    std::atomic<uint64_t> bytes_processed(0);
-
-    int num_threads_to_use = n_threads_p > 0 ? n_threads_p : sd_get_num_physical_cores();
-    LOG_DEBUG("using %d threads for model loading", num_threads_to_use);
+void ModelLoader::process_model_files(bool enable_mmap, bool writable_mmap) {
+    if (model_files_processed) {
+        return;
+    }

    int64_t start_time = ggml_time_ms();

@@ -751,22 +770,13 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
        processed_tensor_storages.push_back(tensor_storage);
    }

-    process_time_ms = ggml_time_ms() - start_time;
-
-    bool success                          = true;
-    size_t total_tensors_processed        = 0;
-    const size_t total_tensors_to_process = processed_tensor_storages.size();
-    const int64_t t_start                 = ggml_time_ms();
-    int last_n_threads                    = 1;
-
    for (size_t file_index = 0; file_index < file_paths_.size(); file_index++) {
        std::string file_path = file_paths_[file_index];
-        LOG_DEBUG("loading tensors from %s", file_path.c_str());

-        std::vector<const TensorStorage*> file_tensors;
+        std::vector<TensorStorage> file_tensors;
        for (const auto& ts : processed_tensor_storages) {
            if (ts.file_index == file_index) {
-                file_tensors.push_back(&ts);
+                file_tensors.push_back(ts);
            }
        }
        if (file_tensors.empty()) {
@@ -775,21 +785,174 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread

        bool is_zip = false;
        for (auto const& ts : file_tensors) {
-            if (ts->index_in_zip >= 0) {
+            if (ts.index_in_zip >= 0) {
                is_zip = true;
                break;
            }
        }

-        std::unique_ptr<MmapWrapper> mmapped;
+        ModelFileData fdata = {};
+        fdata.path          = file_path;
+        fdata.is_zip        = is_zip;
+        fdata.tensors       = std::move(file_tensors);
+
        if (enable_mmap && !is_zip) {
            LOG_DEBUG("using mmap for I/O");
-            mmapped = MmapWrapper::create(file_path);
-            if (!mmapped) {
-                LOG_WARN("failed to memory-map '%s'", file_path.c_str());
+            std::unique_ptr<MmapWrapper> mmapped = MmapWrapper::create(file_path, writable_mmap);
+            if (mmapped) {
+                uint8_t* mmap_data             = static_cast<uint8_t*>(mmapped->writable_data());
+                ggml_backend_buffer_t buf_mmap = ggml_backend_cpu_buffer_from_ptr(mmap_data, mmapped->size());
+                if (buf_mmap) {
+                    LOG_INFO("using mmap for '%s'", file_path.c_str());
+                    fdata.mmbuffer = std::shared_ptr<struct ggml_backend_buffer>(buf_mmap, ggml_backend_buffer_free);
+                } else {
+                    LOG_WARN("mmap: failed to create backend buffer for file %s", fdata.path.c_str());
+                }
+                fdata.mmapped = std::shared_ptr<MmapWrapper>(std::move(mmapped));
+            } else {
+                LOG_WARN("failed to memory-map '%s' (falling back to read())", file_path.c_str());
            }
+        } else if (!is_zip) {
+            LOG_INFO("NOT using mmap for '%s' (mmap disabled by caller)",
+                     file_path.c_str());
        }

+        file_data.push_back(std::move(fdata));
+    }
+
+    model_files_processed = true;
+
+    int64_t end_time        = ggml_time_ms();
+    int64_t process_time_ms = end_time - start_time;
+
+    LOG_INFO("model files processing completed in %.2fs", process_time_ms / 1000.f);
+}
+
+std::vector<MmapTensorStore> ModelLoader::mmap_tensors(std::map<std::string, ggml_tensor*>& tensors,
+                                                       std::set<std::string> ignore_tensors,
+                                                       bool writable_mmap) {
+    process_model_files(true, writable_mmap);
+
+    std::vector<MmapTensorStore> result;
+    uint64_t mapped_bytes = 0;
+    size_t mapped_tensors = 0;
+
+    LOG_DEBUG("memory-mapping tensors...");
+
+    int64_t t_start = ggml_time_ms();
+
+    for (auto& fdata : file_data) {
+        if (!fdata.mmbuffer)
+            continue;
+
+        const std::vector<TensorStorage>& file_tensors = fdata.tensors;
+
+        size_t file_mapped_bytes   = 0;
+        size_t file_mapped_tensors = 0;
+
+        for (const auto& tensor_storage : file_tensors) {
+            const std::string& name = tensor_storage.name;
+
+            bool is_ignored = false;
+            for (const auto& ignore_prefix : ignore_tensors) {
+                if (starts_with(name, ignore_prefix)) {
+                    is_ignored = true;
+                    break;
+                }
+            }
+            if (is_ignored)
+                continue;
+
+            auto it = tensors.find(name);
+            if (it == tensors.end())
+                continue;
+
+            ggml_tensor* dst_tensor = it->second;
+            if (dst_tensor == nullptr)
+                continue;
+
+            if (tensor_storage.is_f8_e4m3 ||
+                tensor_storage.is_f8_e5m2 ||
+                tensor_storage.is_f64 ||
+                tensor_storage.is_i64 ||
+                tensor_storage.type != dst_tensor->type) {
+                continue;
+            }
+
+            size_t tensor_size   = tensor_storage.nbytes();
+            size_t tensor_offset = tensor_storage.offset;
+
+            if (tensor_storage.ne[0] != dst_tensor->ne[0] ||
+                tensor_storage.ne[1] != dst_tensor->ne[1] ||
+                tensor_storage.ne[2] != dst_tensor->ne[2] ||
+                tensor_storage.ne[3] != dst_tensor->ne[3] ||
+                tensor_size != ggml_nbytes(dst_tensor)) {
+                // let load_tensors worry about this
+                continue;
+            }
+
+            ggml_backend_buffer_t buf_mmap = fdata.mmbuffer.get();
+            uint8_t* mmap_data             = static_cast<uint8_t*>(ggml_backend_buffer_get_base(buf_mmap));
+            dst_tensor->buffer             = buf_mmap;
+            dst_tensor->data               = mmap_data + tensor_offset;
+
+            file_mapped_bytes += tensor_size;
+            file_mapped_tensors++;
+        }
+
+        if (file_mapped_bytes > 0) {
+            mapped_tensors += file_mapped_tensors;
+            mapped_bytes += file_mapped_bytes;
+            result.push_back({fdata.mmapped, fdata.mmbuffer});
+        }
+    }
+
+    int64_t t_end       = ggml_time_ms();
+    int64_t duration_ms = t_end - t_start;
+
+    LOG_INFO("memory-mapped %zu tensors in %zu files (%.2f MB), taking %.2fs",
+             mapped_tensors,
+             result.size(),
+             mapped_bytes / (1024.0 * 1024.0),
+             duration_ms / 1000.0);
+
+    return result;
+}
+
+bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_threads_p, bool enable_mmap) {
+    process_model_files(enable_mmap, false);
+
+    std::atomic<int64_t> read_time_ms(0);
+    std::atomic<int64_t> memcpy_time_ms(0);
+    std::atomic<int64_t> copy_to_backend_time_ms(0);
+    std::atomic<int64_t> convert_time_ms(0);
+    std::atomic<uint64_t> bytes_processed(0);
+
+    int num_threads_to_use = n_threads_p > 0 ? n_threads_p : sd_get_num_physical_cores();
+    LOG_DEBUG("using %d threads for model loading", num_threads_to_use);
+
+    int64_t start_time = ggml_time_ms();
+
+    size_t total_tensors_to_process = 0;
+    for (const auto& fdata : file_data) {
+        total_tensors_to_process += fdata.tensors.size();
+    }
+
+    bool success                   = true;
+    size_t total_tensors_processed = 0;
+    const int64_t t_start          = start_time;
+    int last_n_threads             = 1;
+
+    for (auto& fdata : file_data) {
+        const std::string& file_path = fdata.path;
+        LOG_DEBUG("loading tensors from %s", file_path.c_str());
+
+        const std::vector<TensorStorage>& file_tensors = fdata.tensors;
+
+        bool is_zip = fdata.is_zip;
+
+        std::shared_ptr<MmapWrapper> mmapped = fdata.mmapped;
+
        int n_threads = is_zip ? 1 : std::min(num_threads_to_use, (int)file_tensors.size());
        if (n_threads < 1) {
            n_threads = 1;
@@ -830,7 +993,7 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
                        break;
                    }

-                    const TensorStorage& tensor_storage = *file_tensors[idx];
+                    const TensorStorage& tensor_storage = file_tensors[idx];
                    ggml_tensor* dst_tensor             = nullptr;

                    t0 = ggml_time_ms();
@@ -847,6 +1010,17 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
                        continue;
                    }

+                    if (dst_tensor->data == nullptr) {
+                        LOG_ERROR("process tensor data failed: '%s'", tensor_storage.name.c_str());
+                        failed = true;
+                        break;
+                    }
+
+                    // skip mmapped tensors
+                    if (dst_tensor->buffer != nullptr && dst_tensor->buffer == fdata.mmbuffer.get()) {
+                        continue;
+                    }
+
                    size_t nbytes_to_read = tensor_storage.nbytes_to_read();

                    auto read_data = [&](char* buf, size_t n) {
@@ -990,9 +1164,8 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
    }

    int64_t end_time = ggml_time_ms();
-    LOG_INFO("loading tensors completed, taking %.2fs (process: %.2fs, read: %.2fs, memcpy: %.2fs, convert: %.2fs, copy_to_backend: %.2fs)",
+    LOG_INFO("loading tensors completed, taking %.2fs (read: %.2fs, memcpy: %.2fs, convert: %.2fs, copy_to_backend: %.2fs)",
             (end_time - start_time) / 1000.f,
-             process_time_ms / 1000.f,
             (read_time_ms.load() / (float)last_n_threads) / 1000.f,
             (memcpy_time_ms.load() / (float)last_n_threads) / 1000.f,
             (convert_time_ms.load() / (float)last_n_threads) / 1000.f,
@@ -1084,6 +1257,8 @@ bool ModelLoader::tensor_should_be_converted(const TensorStorage& tensor_storage
            // Pass, do not convert
        } else if (ends_with(name, ".scale")) {
            // Pass, do not convert
+        } else if (ends_with(name, ".weight_scale")) {
+            // Pass, do not convert
        } else if (contains(name, "img_in.") ||
                   contains(name, "txt_in.") ||
                   contains(name, "time_in.") ||
@@ -42,9 +42,15 @@ enum SDVersion {
    VERSION_ANIMA,
    VERSION_FLUX2,
    VERSION_FLUX2_KLEIN,
+    VERSION_LTXAV,
+    VERSION_HIDREAM_O1,
    VERSION_Z_IMAGE,
    VERSION_OVIS_IMAGE,
    VERSION_ERNIE_IMAGE,
+    VERSION_LENS,
+    VERSION_LONGCAT,
+    VERSION_PID,
+    VERSION_IDEOGRAM4,
    VERSION_COUNT,
 };

@@ -104,6 +110,13 @@ static inline bool sd_version_is_flux2(SDVersion version) {
    return false;
 }

+static inline bool sd_version_is_ltxav(SDVersion version) {
+    if (version == VERSION_LTXAV) {
+        return true;
+    }
+    return false;
+}
+
 static inline bool sd_version_is_wan(SDVersion version) {
    if (version == VERSION_WAN2 || version == VERSION_WAN2_2_I2V || version == VERSION_WAN2_2_TI2V) {
        return true;
@@ -132,6 +145,13 @@ static inline bool sd_version_is_z_image(SDVersion version) {
    return false;
 }

+static inline bool sd_version_is_longcat(SDVersion version) {
+    if (version == VERSION_LONGCAT) {
+        return true;
+    }
+    return false;
+}
+
 static inline bool sd_version_is_ernie_image(SDVersion version) {
    if (version == VERSION_ERNIE_IMAGE) {
        return true;
@@ -139,8 +159,29 @@ static inline bool sd_version_is_ernie_image(SDVersion version) {
    return false;
 }

+static inline bool sd_version_is_lens(SDVersion version) {
+    if (version == VERSION_LENS) {
+        return true;
+    }
+    return false;
+}
+
+static inline bool sd_version_is_pid(SDVersion version) {
+    if (version == VERSION_PID) {
+        return true;
+    }
+    return false;
+}
+
+static inline bool sd_version_is_ideogram4(SDVersion version) {
+    if (version == VERSION_IDEOGRAM4) {
+        return true;
+    }
+    return false;
+}
+
 static inline bool sd_version_uses_flux2_vae(SDVersion version) {
-    if (sd_version_is_flux2(version) || sd_version_is_ernie_image(version)) {
+    if (sd_version_is_flux2(version) || sd_version_is_ernie_image(version) || sd_version_is_lens(version) || sd_version_is_ideogram4(version)) {
        return true;
    }
    return false;
@@ -160,12 +201,18 @@ static inline bool sd_version_is_inpaint(SDVersion version) {
 static inline bool sd_version_is_dit(SDVersion version) {
    if (sd_version_is_flux(version) ||
        sd_version_is_flux2(version) ||
+        sd_version_is_ltxav(version) ||
        sd_version_is_sd3(version) ||
        sd_version_is_wan(version) ||
        sd_version_is_qwen_image(version) ||
+        version == VERSION_HIDREAM_O1 ||
        sd_version_is_anima(version) ||
        sd_version_is_z_image(version) ||
-        sd_version_is_ernie_image(version)) {
+        sd_version_is_ernie_image(version) ||
+        sd_version_is_lens(version) ||
+        sd_version_is_longcat(version) ||
+        sd_version_is_pid(version) ||
+        sd_version_is_ideogram4(version)) {
        return true;
    }
    return false;
@@ -193,10 +240,27 @@ using TensorTypeRules = std::vector<std::pair<std::string, ggml_type>>;

 TensorTypeRules parse_tensor_type_rules(const std::string& tensor_type_rules);

+class MmapWrapper;
+
+struct ModelFileData {
+    std::string path;
+    std::vector<TensorStorage> tensors;
+    std::shared_ptr<MmapWrapper> mmapped;
+    std::shared_ptr<struct ggml_backend_buffer> mmbuffer;
+    bool is_zip;
+};
+
+struct MmapTensorStore {
+    std::shared_ptr<MmapWrapper> mmapped;
+    std::shared_ptr<struct ggml_backend_buffer> mmbuffer;
+};
+
 class ModelLoader {
 protected:
    SDVersion version_ = VERSION_COUNT;
    std::vector<std::string> file_paths_;
+    std::vector<ModelFileData> file_data;
+    bool model_files_processed = false;
    String2TensorStorage tensor_storage_map;

    void add_tensor_storage(const TensorStorage& tensor_storage);
@@ -220,6 +284,10 @@ public:
    std::map<ggml_type, uint32_t> get_vae_wtype_stat();
    String2TensorStorage& get_tensor_storage_map() { return tensor_storage_map; }
    void set_wtype_override(ggml_type wtype, std::string tensor_type_rules = "");
+    void process_model_files(bool enable_mmap = false, bool writable_mmap = true);
+    std::vector<MmapTensorStore> mmap_tensors(std::map<std::string, ggml_tensor*>& tensors,
+                                              std::set<std::string> ignore_tensors = {},
+                                              bool writable                        = true);
    bool load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_threads = 0, bool use_mmap = false);
    bool load_tensors(std::map<std::string, ggml_tensor*>& tensors,
                      std::set<std::string> ignore_tensors = {},
@@ -128,6 +128,7 @@ std::string convert_cond_stage_model_name(std::string name, std::string prefix)
    };

    static const std::vector<std::pair<std::string, std::string>> llm_name_map{
+        {"attn_sinks.weight", "self_attn.sinks"},
        {"token_embd.", "model.embed_tokens."},
        {"blk.", "model.layers."},
        {"attn_q.", "self_attn.q_proj."},
@@ -137,9 +138,15 @@ std::string convert_cond_stage_model_name(std::string name, std::string prefix)
        {"attn_k_norm.", "self_attn.k_norm."},
        {"attn_output.", "self_attn.o_proj."},
        {"attn_norm.", "input_layernorm."},
+        {"attn_post_norm.", "post_attention_norm."},
+        {"ffn_gate_inp.", "mlp.router."},
+        {"ffn_gate_exps.", "mlp.experts.gate_proj."},
+        {"ffn_up_exps.", "mlp.experts.up_proj."},
+        {"ffn_down_exps.", "mlp.experts.down_proj."},
        {"ffn_down.", "mlp.down_proj."},
        {"ffn_gate.", "mlp.gate_proj."},
        {"ffn_up.", "mlp.up_proj."},
+        {"ffn_post_norm.", "post_ffw_norm."},
        {"ffn_norm.", "post_attention_layernorm."},
        {"output_norm.", "model.norm."},
    };
@@ -508,6 +515,12 @@ std::string convert_diffusers_dit_to_original_flux(std::string name) {
    static std::unordered_map<std::string, std::string> flux_name_map;

    if (flux_name_map.empty()) {
+        // --- time_embed (longcat) ---
+        flux_name_map["time_embed.timestep_embedder.linear_1.weight"] = "time_in.in_layer.weight";
+        flux_name_map["time_embed.timestep_embedder.linear_1.bias"]   = "time_in.in_layer.bias";
+        flux_name_map["time_embed.timestep_embedder.linear_2.weight"] = "time_in.out_layer.weight";
+        flux_name_map["time_embed.timestep_embedder.linear_2.bias"]   = "time_in.out_layer.bias";
+
        // --- time_text_embed ---
        flux_name_map["time_text_embed.timestep_embedder.linear_1.weight"] = "time_in.in_layer.weight";
        flux_name_map["time_text_embed.timestep_embedder.linear_1.bias"]   = "time_in.in_layer.bias";
@@ -561,6 +574,11 @@ std::string convert_diffusers_dit_to_original_flux(std::string name) {
            flux_name_map[block_prefix + "attn.norm_k.weight"]       = dst_prefix + "img_attn.norm.key_norm.scale";
            flux_name_map[block_prefix + "attn.norm_added_q.weight"] = dst_prefix + "txt_attn.norm.query_norm.scale";
            flux_name_map[block_prefix + "attn.norm_added_k.weight"] = dst_prefix + "txt_attn.norm.key_norm.scale";
+            // Comfy-Org/LongCat-Image stores already-converted RMSNorm tensors as *.weight.
+            flux_name_map[dst_prefix + "img_attn.norm.query_norm.weight"] = dst_prefix + "img_attn.norm.query_norm.scale";
+            flux_name_map[dst_prefix + "img_attn.norm.key_norm.weight"]   = dst_prefix + "img_attn.norm.key_norm.scale";
+            flux_name_map[dst_prefix + "txt_attn.norm.query_norm.weight"] = dst_prefix + "txt_attn.norm.query_norm.scale";
+            flux_name_map[dst_prefix + "txt_attn.norm.key_norm.weight"]   = dst_prefix + "txt_attn.norm.key_norm.scale";

            // ff
            flux_name_map[block_prefix + "ff.net.0.proj.weight"] = dst_prefix + "img_mlp.0.weight";
@@ -599,8 +617,11 @@ std::string convert_diffusers_dit_to_original_flux(std::string name) {

            flux_name_map[block_prefix + "attn.norm_q.weight"] = dst_prefix + "norm.query_norm.scale";
            flux_name_map[block_prefix + "attn.norm_k.weight"] = dst_prefix + "norm.key_norm.scale";
-            flux_name_map[block_prefix + "proj_out.weight"]    = dst_prefix + "linear2.weight";
-            flux_name_map[block_prefix + "proj_out.bias"]      = dst_prefix + "linear2.bias";
+            // Comfy-Org/LongCat-Image stores already-converted RMSNorm tensors as *.weight.
+            flux_name_map[dst_prefix + "norm.query_norm.weight"] = dst_prefix + "norm.query_norm.scale";
+            flux_name_map[dst_prefix + "norm.key_norm.weight"]   = dst_prefix + "norm.key_norm.scale";
+            flux_name_map[block_prefix + "proj_out.weight"]      = dst_prefix + "linear2.weight";
+            flux_name_map[block_prefix + "proj_out.bias"]        = dst_prefix + "linear2.bias";
        }

        // --- final layers ---
@@ -668,7 +689,7 @@ std::string convert_diffusion_model_name(std::string name, std::string prefix, S
        name = convert_diffusers_unet_to_original_sdxl(name);
    } else if (sd_version_is_sd3(version)) {
        name = convert_diffusers_dit_to_original_sd3(name);
-    } else if (sd_version_is_flux(version) || sd_version_is_flux2(version)) {
+    } else if (sd_version_is_flux(version) || sd_version_is_flux2(version) || sd_version_is_longcat(version)) {
        name = convert_diffusers_dit_to_original_flux(name);
    } else if (sd_version_is_z_image(version)) {
        name = convert_diffusers_dit_to_original_lumina2(name);
@@ -0,0 +1,842 @@
+#ifndef __SD_PID_HPP__
+#define __SD_PID_HPP__
+
+#include <cmath>
+#include <cstdlib>
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "common_dit.hpp"
+#include "ggml_extend.hpp"
+#include "mmdit.hpp"
+#include "rope.hpp"
+
+namespace Pid {
+    constexpr int PID_GRAPH_SIZE = 196608;
+    constexpr float PID_PI       = 3.14159265358979323846f;
+
+    struct PixelDiTParams {
+        int64_t in_channels            = 3;
+        int64_t hidden_size            = 1536;
+        int64_t num_groups             = 24;
+        int64_t patch_mlp_hidden_dim   = 4096;
+        int64_t pixel_hidden_size      = 16;
+        int64_t pixel_attn_hidden_size = 1152;
+        int64_t pixel_num_groups       = 16;
+        int64_t patch_depth            = 14;
+        int64_t pixel_depth            = 2;
+        int64_t patch_size             = 16;
+        int64_t txt_embed_dim          = 2304;
+        int64_t txt_max_length         = 300;
+        float text_rope_theta          = 10000.f;
+        int64_t lq_latent_channels     = 16;
+        int64_t lq_hidden_dim          = 512;
+        int64_t lq_num_res_blocks      = 4;
+        int64_t lq_interval            = 2;
+        int64_t lq_sr_scale            = 4;
+        int64_t lq_latent_down_factor  = 8;
+        int64_t rope_ref_grid_h        = 64;
+        int64_t rope_ref_grid_w        = 64;
+    };
+
+    inline std::vector<float> make_rope_1d(int length,
+                                           int dim,
+                                           float theta) {
+        GGML_ASSERT(dim % 2 == 0);
+        return Rope::flatten(Rope::rope(Rope::linspace(0.f, static_cast<float>(length - 1), length), dim, theta));
+    }
+
+    inline std::vector<float> make_rope_2d(int height,
+                                           int width,
+                                           int dim,
+                                           float theta    = 10000.f,
+                                           float scale    = 16.f,
+                                           int ref_grid_h = 0,
+                                           int ref_grid_w = 0) {
+        GGML_ASSERT(dim % 4 == 0);
+        return Rope::embed_2d_interleaved(height, width, dim, theta, scale, ref_grid_h, ref_grid_w);
+    }
+
+    inline std::vector<float> make_pixel_abs_pos(int height,
+                                                 int width,
+                                                 int dim) {
+        GGML_ASSERT(dim % 4 == 0);
+        int half_dim = dim / 2;
+        std::vector<float> x_pos;
+        std::vector<float> y_pos;
+        x_pos.reserve(static_cast<size_t>(height) * width);
+        y_pos.reserve(static_cast<size_t>(height) * width);
+        for (int iy = 0; iy < height; ++iy) {
+            for (int ix = 0; ix < width; ++ix) {
+                x_pos.push_back(static_cast<float>(ix));
+                y_pos.push_back(static_cast<float>(iy));
+            }
+        }
+
+        auto x_emb = timestep_embedding(x_pos, half_dim, 10000, false);
+        auto y_emb = timestep_embedding(y_pos, half_dim, 10000, false);
+
+        std::vector<float> out(static_cast<size_t>(dim) * height * width);
+        for (int pos = 0; pos < height * width; ++pos) {
+            size_t out_base = static_cast<size_t>(pos) * dim;
+            size_t emb_base = static_cast<size_t>(pos) * half_dim;
+            for (int i = 0; i < half_dim; ++i) {
+                out[out_base + i]            = x_emb[emb_base + i];
+                out[out_base + half_dim + i] = y_emb[emb_base + i];
+            }
+        }
+        return out;
+    }
+
+    inline ggml_tensor* apply_adaln(ggml_context* ctx,
+                                    ggml_tensor* x,
+                                    ggml_tensor* shift,
+                                    ggml_tensor* scale) {
+        return ggml_add(ctx, ggml_add(ctx, x, ggml_mul(ctx, x, scale)), shift);
+    }
+
+    struct PatchTokenEmbedder : public GGMLBlock {
+        bool use_rms_norm;
+
+        PatchTokenEmbedder(int64_t in_chans,
+                           int64_t embed_dim,
+                           bool use_rms_norm = false,
+                           bool bias         = true)
+            : use_rms_norm(use_rms_norm) {
+            blocks["proj"] = std::make_shared<Linear>(in_chans, embed_dim, bias);
+            if (use_rms_norm) {
+                blocks["norm"] = std::make_shared<RMSNorm>(embed_dim, 1e-6f);
+            }
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto proj = std::dynamic_pointer_cast<Linear>(blocks["proj"]);
+            x         = proj->forward(ctx, x);
+            if (use_rms_norm) {
+                auto norm = std::dynamic_pointer_cast<RMSNorm>(blocks["norm"]);
+                x         = norm->forward(ctx, x);
+            }
+            return x;
+        }
+    };
+
+    struct PixelDiTTimestepEmbedder : public GGMLBlock {
+        int frequency_embedding_size;
+
+        PixelDiTTimestepEmbedder(int64_t hidden_size,
+                                 int frequency_embedding_size = 256)
+            : frequency_embedding_size(frequency_embedding_size) {
+            blocks["mlp.0"] = std::make_shared<Linear>(frequency_embedding_size, hidden_size, true, true);
+            blocks["mlp.2"] = std::make_shared<Linear>(hidden_size, hidden_size, true, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* t) {
+            auto mlp_0 = std::dynamic_pointer_cast<Linear>(blocks["mlp.0"]);
+            auto mlp_2 = std::dynamic_pointer_cast<Linear>(blocks["mlp.2"]);
+            auto t_emb = ggml_ext_timestep_embedding(ctx->ggml_ctx, t, frequency_embedding_size, 10);
+            t_emb      = mlp_0->forward(ctx, t_emb);
+            t_emb      = ggml_silu_inplace(ctx->ggml_ctx, t_emb);
+            return mlp_2->forward(ctx, t_emb);
+        }
+    };
+
+    struct FeedForward : public GGMLBlock {
+        FeedForward(int64_t dim, int64_t hidden_dim) {
+            blocks["w1"] = std::make_shared<Linear>(dim, hidden_dim, false);
+            blocks["w2"] = std::make_shared<Linear>(hidden_dim, dim, false);
+            blocks["w3"] = std::make_shared<Linear>(dim, hidden_dim, false);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto w1 = std::dynamic_pointer_cast<Linear>(blocks["w1"]);
+            auto w2 = std::dynamic_pointer_cast<Linear>(blocks["w2"]);
+            auto w3 = std::dynamic_pointer_cast<Linear>(blocks["w3"]);
+            auto h  = ggml_silu_inplace(ctx->ggml_ctx, w1->forward(ctx, x));
+            h       = ggml_mul_inplace(ctx->ggml_ctx, h, w3->forward(ctx, x));
+            return w2->forward(ctx, h);
+        }
+    };
+
+    struct FinalLayer : public GGMLBlock {
+        FinalLayer(int64_t hidden_size, int64_t out_channels) {
+            blocks["norm"]   = std::make_shared<RMSNorm>(hidden_size, 1e-6f);
+            blocks["linear"] = std::make_shared<Linear>(hidden_size, out_channels, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto norm   = std::dynamic_pointer_cast<RMSNorm>(blocks["norm"]);
+            auto linear = std::dynamic_pointer_cast<Linear>(blocks["linear"]);
+            return linear->forward(ctx, norm->forward(ctx, x));
+        }
+    };
+
+    struct RotaryAttention : public GGMLBlock {
+        int64_t dim;
+        int64_t num_heads;
+
+        RotaryAttention(int64_t dim, int64_t num_heads)
+            : dim(dim), num_heads(num_heads) {
+            int64_t head_dim = dim / num_heads;
+            blocks["qkv"]    = std::make_shared<Linear>(dim, dim * 3, false);
+            blocks["q_norm"] = std::make_shared<RMSNorm>(head_dim, 1e-6f);
+            blocks["k_norm"] = std::make_shared<RMSNorm>(head_dim, 1e-6f);
+            blocks["proj"]   = std::make_shared<Linear>(dim, dim, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* pos) {
+            auto qkv_proj = std::dynamic_pointer_cast<Linear>(blocks["qkv"]);
+            auto q_norm   = std::dynamic_pointer_cast<RMSNorm>(blocks["q_norm"]);
+            auto k_norm   = std::dynamic_pointer_cast<RMSNorm>(blocks["k_norm"]);
+            auto proj     = std::dynamic_pointer_cast<Linear>(blocks["proj"]);
+
+            auto qkv         = qkv_proj->forward(ctx, x);
+            auto qkv_vec     = split_qkv(ctx->ggml_ctx, qkv);
+            int64_t L        = x->ne[1];
+            int64_t N        = x->ne[2];
+            int64_t head_dim = dim / num_heads;
+            auto q           = ggml_reshape_4d(ctx->ggml_ctx, qkv_vec[0], head_dim, num_heads, L, N);
+            auto k           = ggml_reshape_4d(ctx->ggml_ctx, qkv_vec[1], head_dim, num_heads, L, N);
+            auto v           = ggml_reshape_4d(ctx->ggml_ctx, qkv_vec[2], head_dim, num_heads, L, N);
+            q                = q_norm->forward(ctx, q);
+            k                = k_norm->forward(ctx, k);
+            x                = Rope::attention(ctx, q, k, v, pos, nullptr, 1.0f / 128.f, true);
+            return proj->forward(ctx, x);
+        }
+    };
+
+    struct MMDiTJointAttention : public GGMLBlock {
+        int64_t dim;
+        int64_t num_heads;
+
+        MMDiTJointAttention(int64_t dim, int64_t num_heads)
+            : dim(dim), num_heads(num_heads) {
+            int64_t head_dim   = dim / num_heads;
+            blocks["qkv_x"]    = std::make_shared<Linear>(dim, dim * 3, false);
+            blocks["qkv_y"]    = std::make_shared<Linear>(dim, dim * 3, false);
+            blocks["q_norm_x"] = std::make_shared<RMSNorm>(head_dim, 1e-6f);
+            blocks["k_norm_x"] = std::make_shared<RMSNorm>(head_dim, 1e-6f);
+            blocks["q_norm_y"] = std::make_shared<RMSNorm>(head_dim, 1e-6f);
+            blocks["k_norm_y"] = std::make_shared<RMSNorm>(head_dim, 1e-6f);
+            blocks["proj_x"]   = std::make_shared<Linear>(dim, dim, true);
+            blocks["proj_y"]   = std::make_shared<Linear>(dim, dim, true);
+        }
+
+        std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx,
+                                                      ggml_tensor* x,
+                                                      ggml_tensor* y,
+                                                      ggml_tensor* pos_img,
+                                                      ggml_tensor* pos_txt) {
+            auto qkv_x_proj = std::dynamic_pointer_cast<Linear>(blocks["qkv_x"]);
+            auto qkv_y_proj = std::dynamic_pointer_cast<Linear>(blocks["qkv_y"]);
+            auto q_norm_x   = std::dynamic_pointer_cast<RMSNorm>(blocks["q_norm_x"]);
+            auto k_norm_x   = std::dynamic_pointer_cast<RMSNorm>(blocks["k_norm_x"]);
+            auto q_norm_y   = std::dynamic_pointer_cast<RMSNorm>(blocks["q_norm_y"]);
+            auto k_norm_y   = std::dynamic_pointer_cast<RMSNorm>(blocks["k_norm_y"]);
+            auto proj_x     = std::dynamic_pointer_cast<Linear>(blocks["proj_x"]);
+            auto proj_y     = std::dynamic_pointer_cast<Linear>(blocks["proj_y"]);
+
+            int64_t Nx       = x->ne[1];
+            int64_t Ny       = y->ne[1];
+            int64_t N        = x->ne[2];
+            int64_t head_dim = dim / num_heads;
+
+            auto qkv_x = split_qkv(ctx->ggml_ctx, qkv_x_proj->forward(ctx, x));
+            auto qx    = ggml_reshape_4d(ctx->ggml_ctx, qkv_x[0], head_dim, num_heads, Nx, N);
+            auto kx    = ggml_reshape_4d(ctx->ggml_ctx, qkv_x[1], head_dim, num_heads, Nx, N);
+            auto vx    = ggml_reshape_4d(ctx->ggml_ctx, qkv_x[2], head_dim, num_heads, Nx, N);
+            qx         = q_norm_x->forward(ctx, qx);
+            kx         = k_norm_x->forward(ctx, kx);
+
+            auto qkv_y = split_qkv(ctx->ggml_ctx, qkv_y_proj->forward(ctx, y));
+            auto qy    = ggml_reshape_4d(ctx->ggml_ctx, qkv_y[0], head_dim, num_heads, Ny, N);
+            auto ky    = ggml_reshape_4d(ctx->ggml_ctx, qkv_y[1], head_dim, num_heads, Ny, N);
+            auto vy    = ggml_reshape_4d(ctx->ggml_ctx, qkv_y[2], head_dim, num_heads, Ny, N);
+            qy         = q_norm_y->forward(ctx, qy);
+            ky         = k_norm_y->forward(ctx, ky);
+
+            auto q_joint   = ggml_concat(ctx->ggml_ctx, qy, qx, 2);
+            auto k_joint   = ggml_concat(ctx->ggml_ctx, ky, kx, 2);
+            auto v_joint   = ggml_concat(ctx->ggml_ctx, vy, vx, 2);
+            auto pos_joint = ggml_concat(ctx->ggml_ctx, pos_txt, pos_img, 3);
+            auto out       = Rope::attention(ctx, q_joint, k_joint, v_joint, pos_joint, nullptr, 1.0f, true);
+
+            auto out_y = ggml_ext_slice(ctx->ggml_ctx, out, 1, 0, Ny);
+            auto out_x = ggml_ext_slice(ctx->ggml_ctx, out, 1, Ny, Ny + Nx);
+            return {proj_x->forward(ctx, out_x), proj_y->forward(ctx, out_y)};
+        }
+    };
+
+    struct MMDiTBlockT2I : public GGMLBlock {
+        int64_t hidden_size;
+
+        MMDiTBlockT2I(int64_t hidden_size, int64_t groups, int64_t mlp_hidden_dim)
+            : hidden_size(hidden_size) {
+            blocks["norm_x1"]                = std::make_shared<RMSNorm>(hidden_size, 1e-6f);
+            blocks["norm_y1"]                = std::make_shared<RMSNorm>(hidden_size, 1e-6f);
+            blocks["attn"]                   = std::make_shared<MMDiTJointAttention>(hidden_size, groups);
+            blocks["norm_x2"]                = std::make_shared<RMSNorm>(hidden_size, 1e-6f);
+            blocks["norm_y2"]                = std::make_shared<RMSNorm>(hidden_size, 1e-6f);
+            blocks["mlp_x"]                  = std::make_shared<FeedForward>(hidden_size, mlp_hidden_dim);
+            blocks["mlp_y"]                  = std::make_shared<FeedForward>(hidden_size, mlp_hidden_dim);
+            blocks["adaLN_modulation_img.0"] = std::make_shared<Linear>(hidden_size, 6 * hidden_size, true);
+            blocks["adaLN_modulation_txt.0"] = std::make_shared<Linear>(hidden_size, 6 * hidden_size, true);
+        }
+
+        std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx,
+                                                      ggml_tensor* x,
+                                                      ggml_tensor* y,
+                                                      ggml_tensor* c,
+                                                      ggml_tensor* pos_img,
+                                                      ggml_tensor* pos_txt) {
+            auto norm_x1 = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_x1"]);
+            auto norm_y1 = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_y1"]);
+            auto attn    = std::dynamic_pointer_cast<MMDiTJointAttention>(blocks["attn"]);
+            auto norm_x2 = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_x2"]);
+            auto norm_y2 = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_y2"]);
+            auto mlp_x   = std::dynamic_pointer_cast<FeedForward>(blocks["mlp_x"]);
+            auto mlp_y   = std::dynamic_pointer_cast<FeedForward>(blocks["mlp_y"]);
+            auto ada_img = std::dynamic_pointer_cast<Linear>(blocks["adaLN_modulation_img.0"]);
+            auto ada_txt = std::dynamic_pointer_cast<Linear>(blocks["adaLN_modulation_txt.0"]);
+
+            auto mx = ggml_ext_chunk(ctx->ggml_ctx, ada_img->forward(ctx, c), 6, 0);
+            auto my = ggml_ext_chunk(ctx->ggml_ctx, ada_txt->forward(ctx, c), 6, 0);
+
+            auto x_norm   = apply_adaln(ctx->ggml_ctx, norm_x1->forward(ctx, x), mx[0], mx[1]);
+            auto y_norm   = apply_adaln(ctx->ggml_ctx, norm_y1->forward(ctx, y), my[0], my[1]);
+            auto attn_out = attn->forward(ctx, x_norm, y_norm, pos_img, pos_txt);
+
+            x = ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, attn_out.first, mx[2]));
+            y = ggml_add(ctx->ggml_ctx, y, ggml_mul(ctx->ggml_ctx, attn_out.second, my[2]));
+
+            auto x_mlp = mlp_x->forward(ctx, apply_adaln(ctx->ggml_ctx, norm_x2->forward(ctx, x), mx[3], mx[4]));
+            auto y_mlp = mlp_y->forward(ctx, apply_adaln(ctx->ggml_ctx, norm_y2->forward(ctx, y), my[3], my[4]));
+            x          = ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, x_mlp, mx[5]));
+            y          = ggml_add(ctx->ggml_ctx, y, ggml_mul(ctx->ggml_ctx, y_mlp, my[5]));
+            return {x, y};
+        }
+    };
+
+    struct PixelTokenEmbedder : public GGMLBlock {
+        int64_t in_channels;
+        int64_t hidden_size_output;
+
+        PixelTokenEmbedder(int64_t in_channels, int64_t hidden_size_output)
+            : in_channels(in_channels), hidden_size_output(hidden_size_output) {
+            blocks["proj"] = std::make_shared<Linear>(in_channels, hidden_size_output, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* inputs,
+                             int64_t patch_size,
+                             ggml_tensor* pos_full) {
+            auto proj  = std::dynamic_pointer_cast<Linear>(blocks["proj"]);
+            int64_t W  = inputs->ne[0];
+            int64_t H  = inputs->ne[1];
+            int64_t B  = inputs->ne[3];
+            int64_t L  = (W / patch_size) * (H / patch_size);
+            int64_t P2 = patch_size * patch_size;
+
+            auto x = ggml_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, inputs, 2, 0, 1, 3));
+            x      = ggml_reshape_3d(ctx->ggml_ctx, x, in_channels, W * H, B);
+            x      = proj->forward(ctx, x);
+            x      = ggml_add(ctx->ggml_ctx, x, pos_full);
+            x      = ggml_reshape_4d(ctx->ggml_ctx, x, hidden_size_output, W, H, B);
+            x      = ggml_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, x, 1, 2, 0, 3));
+            x      = DiT::patchify(ctx->ggml_ctx, x, static_cast<int>(patch_size), static_cast<int>(patch_size), false);
+            x      = ggml_reshape_3d(ctx->ggml_ctx, x, hidden_size_output, P2, L * B);
+            return x;
+        }
+    };
+
+    struct PiTBlock : public GGMLBlock {
+        int64_t pixel_dim;
+        int64_t context_dim;
+        int64_t attn_dim;
+        int64_t num_heads;
+        int64_t patch_size;
+
+        PiTBlock(int64_t pixel_dim,
+                 int64_t context_dim,
+                 int64_t patch_size,
+                 int64_t attn_dim,
+                 int64_t num_heads)
+            : pixel_dim(pixel_dim),
+              context_dim(context_dim),
+              attn_dim(attn_dim),
+              num_heads(num_heads),
+              patch_size(patch_size) {
+            int64_t p2                   = patch_size * patch_size;
+            blocks["compress_to_attn"]   = std::make_shared<Linear>(p2 * pixel_dim, attn_dim, true);
+            blocks["expand_from_attn"]   = std::make_shared<Linear>(attn_dim, p2 * pixel_dim, true);
+            blocks["norm1"]              = std::make_shared<RMSNorm>(pixel_dim, 1e-6f);
+            blocks["attn"]               = std::make_shared<RotaryAttention>(attn_dim, num_heads);
+            blocks["norm2"]              = std::make_shared<RMSNorm>(pixel_dim, 1e-6f);
+            blocks["mlp"]                = std::make_shared<Mlp>(pixel_dim, pixel_dim * 4);
+            blocks["adaLN_modulation.0"] = std::make_shared<Linear>(context_dim, 6 * pixel_dim * p2, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* s_cond,
+                             int64_t image_height,
+                             int64_t image_width,
+                             ggml_tensor* pos_comp) {
+            auto compress = std::dynamic_pointer_cast<Linear>(blocks["compress_to_attn"]);
+            auto expand   = std::dynamic_pointer_cast<Linear>(blocks["expand_from_attn"]);
+            auto norm1    = std::dynamic_pointer_cast<RMSNorm>(blocks["norm1"]);
+            auto attn     = std::dynamic_pointer_cast<RotaryAttention>(blocks["attn"]);
+            auto norm2    = std::dynamic_pointer_cast<RMSNorm>(blocks["norm2"]);
+            auto mlp      = std::dynamic_pointer_cast<Mlp>(blocks["mlp"]);
+            auto ada      = std::dynamic_pointer_cast<Linear>(blocks["adaLN_modulation.0"]);
+
+            int64_t Hs = image_height / patch_size;
+            int64_t Ws = image_width / patch_size;
+            int64_t L  = Hs * Ws;
+            int64_t BL = x->ne[2];
+            int64_t B  = BL / L;
+            int64_t P2 = patch_size * patch_size;
+
+            auto ada_params = ada->forward(ctx, s_cond);
+            ada_params      = ggml_reshape_3d(ctx->ggml_ctx, ada_params, 6 * pixel_dim, P2, BL);
+            auto mod        = ggml_ext_chunk(ctx->ggml_ctx, ada_params, 6, 0);
+
+            auto x_norm    = apply_adaln(ctx->ggml_ctx, norm1->forward(ctx, x), mod[0], mod[1]);
+            auto x_flat    = ggml_reshape_2d(ctx->ggml_ctx, x_norm, P2 * pixel_dim, BL);
+            auto x_comp    = compress->forward(ctx, x_flat);
+            x_comp         = ggml_reshape_3d(ctx->ggml_ctx, x_comp, attn_dim, L, B);
+            auto attn_out  = attn->forward(ctx, x_comp, pos_comp);
+            auto attn_flat = expand->forward(ctx, ggml_reshape_2d(ctx->ggml_ctx, attn_out, attn_dim, BL));
+            auto attn_exp  = ggml_reshape_3d(ctx->ggml_ctx, attn_flat, pixel_dim, P2, BL);
+            x              = ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, attn_exp, mod[2]));
+
+            auto mlp_out = mlp->forward(ctx, apply_adaln(ctx->ggml_ctx, norm2->forward(ctx, x), mod[3], mod[4]));
+            return ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, mlp_out, mod[5]));
+        }
+    };
+
+    struct SigmaAwareGate : public GGMLBlock {
+        int64_t dim;
+
+        SigmaAwareGate(int64_t dim)
+            : dim(dim) {
+            blocks["content_proj"] = std::make_shared<Linear>(dim * 2, dim, true);
+        }
+
+        void init_params(ggml_context* ctx,
+                         const String2TensorStorage& tensor_storage_map = {},
+                         std::string prefix                             = "") override {
+            params["log_alpha"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* lq,
+                             ggml_tensor* sigma) {
+            auto content_proj = std::dynamic_pointer_cast<Linear>(blocks["content_proj"]);
+
+            auto content_logit = content_proj->forward(ctx, ggml_concat(ctx->ggml_ctx, x, lq, 0));
+            sigma              = ggml_reshape_3d(ctx->ggml_ctx, sigma, 1, 1, sigma->ne[0]);
+            auto alpha         = ggml_exp(ctx->ggml_ctx, params["log_alpha"]);
+            auto offset        = ggml_neg(ctx->ggml_ctx, ggml_mul(ctx->ggml_ctx, alpha, sigma));
+            auto gate          = ggml_sigmoid(ctx->ggml_ctx, ggml_add(ctx->ggml_ctx, content_logit, offset));
+            return ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, gate, lq));
+        }
+    };
+
+    struct PiDResBlock : public GGMLBlock {
+        PiDResBlock(int64_t channels) {
+            blocks["block.0"] = std::make_shared<GroupNorm>(4, channels, 1e-5f);
+            blocks["block.2"] = std::make_shared<Conv2d>(channels, channels, std::pair<int, int>{3, 3}, std::pair<int, int>{1, 1}, std::pair<int, int>{1, 1});
+            blocks["block.3"] = std::make_shared<GroupNorm>(4, channels, 1e-5f);
+            blocks["block.5"] = std::make_shared<Conv2d>(channels, channels, std::pair<int, int>{3, 3}, std::pair<int, int>{1, 1}, std::pair<int, int>{1, 1});
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto norm1 = std::dynamic_pointer_cast<GroupNorm>(blocks["block.0"]);
+            auto conv1 = std::dynamic_pointer_cast<Conv2d>(blocks["block.2"]);
+            auto norm2 = std::dynamic_pointer_cast<GroupNorm>(blocks["block.3"]);
+            auto conv2 = std::dynamic_pointer_cast<Conv2d>(blocks["block.5"]);
+            auto h     = ggml_silu_inplace(ctx->ggml_ctx, norm1->forward(ctx, x));
+            h          = conv1->forward(ctx, h);
+            h          = ggml_silu_inplace(ctx->ggml_ctx, norm2->forward(ctx, h));
+            h          = conv2->forward(ctx, h);
+            return ggml_add(ctx->ggml_ctx, x, h);
+        }
+    };
+
+    struct LQProjection2D : public GGMLBlock {
+        PixelDiTParams params_cfg;
+
+        LQProjection2D(const PixelDiTParams& params_cfg)
+            : params_cfg(params_cfg) {
+            blocks["latent_proj.0"] = std::make_shared<Conv2d>(params_cfg.lq_latent_channels, params_cfg.lq_hidden_dim, std::pair<int, int>{3, 3}, std::pair<int, int>{1, 1}, std::pair<int, int>{1, 1});
+            blocks["latent_proj.2"] = std::make_shared<Conv2d>(params_cfg.lq_hidden_dim, params_cfg.lq_hidden_dim, std::pair<int, int>{3, 3}, std::pair<int, int>{1, 1}, std::pair<int, int>{1, 1});
+            for (int i = 0; i < params_cfg.lq_num_res_blocks; ++i) {
+                blocks["latent_proj." + std::to_string(3 + i)] = std::make_shared<PiDResBlock>(params_cfg.lq_hidden_dim);
+            }
+
+            int num_outputs = static_cast<int>((params_cfg.patch_depth + params_cfg.lq_interval - 1) / params_cfg.lq_interval);
+            for (int i = 0; i < num_outputs; ++i) {
+                blocks["output_heads." + std::to_string(i)] = std::make_shared<Linear>(params_cfg.lq_hidden_dim, params_cfg.hidden_size, true);
+                blocks["gate_modules." + std::to_string(i)] = std::make_shared<SigmaAwareGate>(params_cfg.hidden_size);
+            }
+        }
+
+        bool is_gate_active(int block_idx) const {
+            return block_idx % params_cfg.lq_interval == 0;
+        }
+
+        int get_output_index(int block_idx) const {
+            return block_idx / static_cast<int>(params_cfg.lq_interval);
+        }
+
+        ggml_tensor* gate(GGMLRunnerContext* ctx,
+                          ggml_tensor* x,
+                          ggml_tensor* lq,
+                          ggml_tensor* sigma,
+                          int out_idx) {
+            auto gate_module = std::dynamic_pointer_cast<SigmaAwareGate>(blocks["gate_modules." + std::to_string(out_idx)]);
+            return gate_module->forward(ctx, x, lq, sigma);
+        }
+
+        std::vector<ggml_tensor*> forward(GGMLRunnerContext* ctx,
+                                          ggml_tensor* lq_latent,
+                                          int64_t target_pH,
+                                          int64_t target_pW) {
+            auto conv0             = std::dynamic_pointer_cast<Conv2d>(blocks["latent_proj.0"]);
+            auto conv2             = std::dynamic_pointer_cast<Conv2d>(blocks["latent_proj.2"]);
+            float z_to_patch_ratio = static_cast<float>(params_cfg.lq_sr_scale * params_cfg.lq_latent_down_factor) /
+                                     static_cast<float>(params_cfg.patch_size);
+            GGML_ASSERT(z_to_patch_ratio >= 1.0f);
+            if (lq_latent->ne[0] != target_pW || lq_latent->ne[1] != target_pH) {
+                lq_latent = ggml_interpolate(ctx->ggml_ctx,
+                                             lq_latent,
+                                             target_pW,
+                                             target_pH,
+                                             lq_latent->ne[2],
+                                             lq_latent->ne[3],
+                                             GGML_SCALE_MODE_NEAREST);
+            }
+
+            auto feat = conv0->forward(ctx, lq_latent);
+            feat      = ggml_silu_inplace(ctx->ggml_ctx, feat);
+            feat      = conv2->forward(ctx, feat);
+            for (int i = 0; i < params_cfg.lq_num_res_blocks; ++i) {
+                auto block = std::dynamic_pointer_cast<PiDResBlock>(blocks["latent_proj." + std::to_string(3 + i)]);
+                feat       = block->forward(ctx, feat);
+            }
+
+            int64_t B   = feat->ne[3];
+            int64_t C   = feat->ne[2];
+            int64_t L   = target_pH * target_pW;
+            auto tokens = ggml_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, feat, 2, 0, 1, 3));
+            tokens      = ggml_reshape_3d(ctx->ggml_ctx, tokens, C, L, B);
+
+            int num_outputs = static_cast<int>((params_cfg.patch_depth + params_cfg.lq_interval - 1) / params_cfg.lq_interval);
+            std::vector<ggml_tensor*> outputs;
+            outputs.reserve(num_outputs);
+            for (int i = 0; i < num_outputs; ++i) {
+                auto head = std::dynamic_pointer_cast<Linear>(blocks["output_heads." + std::to_string(i)]);
+                outputs.push_back(head->forward(ctx, tokens));
+            }
+            return outputs;
+        }
+    };
+
+    struct PixelDiT : public GGMLBlock {
+        PixelDiTParams params_cfg;
+
+        PixelDiT() = default;
+
+        PixelDiT(const PixelDiTParams& params_cfg)
+            : params_cfg(params_cfg) {
+            blocks["pixel_embedder"] = std::make_shared<PixelTokenEmbedder>(params_cfg.in_channels, params_cfg.pixel_hidden_size);
+            blocks["s_embedder"]     = std::make_shared<PatchTokenEmbedder>(params_cfg.in_channels * params_cfg.patch_size * params_cfg.patch_size, params_cfg.hidden_size, false, true);
+            blocks["t_embedder"]     = std::make_shared<PixelDiTTimestepEmbedder>(params_cfg.hidden_size);
+            blocks["y_embedder"]     = std::make_shared<PatchTokenEmbedder>(params_cfg.txt_embed_dim, params_cfg.hidden_size, true, true);
+            for (int i = 0; i < params_cfg.patch_depth; ++i) {
+                blocks["patch_blocks." + std::to_string(i)] = std::make_shared<MMDiTBlockT2I>(params_cfg.hidden_size, params_cfg.num_groups, params_cfg.patch_mlp_hidden_dim);
+            }
+            for (int i = 0; i < params_cfg.pixel_depth; ++i) {
+                blocks["pixel_blocks." + std::to_string(i)] = std::make_shared<PiTBlock>(params_cfg.pixel_hidden_size,
+                                                                                         params_cfg.hidden_size,
+                                                                                         params_cfg.patch_size,
+                                                                                         params_cfg.pixel_attn_hidden_size,
+                                                                                         params_cfg.pixel_num_groups);
+            }
+            blocks["final_layer"] = std::make_shared<FinalLayer>(params_cfg.pixel_hidden_size, params_cfg.in_channels);
+            blocks["lq_proj"]     = std::make_shared<LQProjection2D>(params_cfg);
+        }
+
+        void init_params(ggml_context* ctx,
+                         const String2TensorStorage& tensor_storage_map = {},
+                         std::string prefix                             = "") override {
+            params["y_pos_embedding"] = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, params_cfg.hidden_size, params_cfg.txt_max_length, 1);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* timesteps,
+                             ggml_tensor* context,
+                             ggml_tensor* lq_latent,
+                             ggml_tensor* degrade_sigma,
+                             ggml_tensor* pos_img,
+                             ggml_tensor* pos_txt,
+                             ggml_tensor* pixel_pos_full,
+                             ggml_tensor* pixel_pos_comp) {
+            auto pixel_embedder = std::dynamic_pointer_cast<PixelTokenEmbedder>(blocks["pixel_embedder"]);
+            auto s_embedder     = std::dynamic_pointer_cast<PatchTokenEmbedder>(blocks["s_embedder"]);
+            auto t_embedder     = std::dynamic_pointer_cast<PixelDiTTimestepEmbedder>(blocks["t_embedder"]);
+            auto y_embedder     = std::dynamic_pointer_cast<PatchTokenEmbedder>(blocks["y_embedder"]);
+            auto final_layer    = std::dynamic_pointer_cast<FinalLayer>(blocks["final_layer"]);
+            auto lq_proj        = std::dynamic_pointer_cast<LQProjection2D>(blocks["lq_proj"]);
+
+            int64_t W_orig = x->ne[0];
+            int64_t H_orig = x->ne[1];
+            x              = DiT::pad_to_patch_size(ctx, x, static_cast<int>(params_cfg.patch_size), static_cast<int>(params_cfg.patch_size));
+            int64_t W      = x->ne[0];
+            int64_t H      = x->ne[1];
+            int64_t B      = x->ne[3];
+            int64_t Hs     = H / params_cfg.patch_size;
+            int64_t Ws     = W / params_cfg.patch_size;
+            int64_t L      = Hs * Ws;
+            int64_t P2     = params_cfg.patch_size * params_cfg.patch_size;
+
+            auto x_patches = DiT::patchify(ctx->ggml_ctx, x, static_cast<int>(params_cfg.patch_size), static_cast<int>(params_cfg.patch_size), true);
+            auto t_emb     = t_embedder->forward(ctx, timesteps);
+            auto condition = ggml_silu(ctx->ggml_ctx, t_emb);
+
+            GGML_ASSERT(context != nullptr);
+            int64_t Ltxt = std::min<int64_t>(context->ne[1], params_cfg.txt_max_length);
+            auto y       = ggml_ext_slice(ctx->ggml_ctx, context, 1, 0, Ltxt);
+            auto y_emb   = y_embedder->forward(ctx, y);
+            auto y_pos   = ggml_ext_slice(ctx->ggml_ctx, params["y_pos_embedding"], 1, 0, Ltxt);
+            y_emb        = ggml_add(ctx->ggml_ctx, y_emb, y_pos);
+
+            std::vector<ggml_tensor*> lq_features = lq_proj->forward(ctx, lq_latent, Hs, Ws);
+
+            auto s = s_embedder->forward(ctx, x_patches);
+
+            for (int i = 0; i < params_cfg.patch_depth; ++i) {
+                if (lq_proj->is_gate_active(i)) {
+                    int out_idx = lq_proj->get_output_index(i);
+                    if (out_idx < static_cast<int>(lq_features.size())) {
+                        s = lq_proj->gate(ctx, s, lq_features[out_idx], degrade_sigma, out_idx);
+                    }
+                }
+                auto block = std::dynamic_pointer_cast<MMDiTBlockT2I>(blocks["patch_blocks." + std::to_string(i)]);
+                auto out   = block->forward(ctx,
+                                            s,
+                                            y_emb,
+                                            condition,
+                                            pos_img,
+                                            pos_txt);
+                s          = out.first;
+                y_emb      = out.second;
+                sd::ggml_graph_cut::mark_graph_cut(s, "pid.patch_blocks." + std::to_string(i), "s");
+                sd::ggml_graph_cut::mark_graph_cut(y_emb, "pid.patch_blocks." + std::to_string(i), "y");
+            }
+            s = ggml_silu(ctx->ggml_ctx, ggml_add(ctx->ggml_ctx, s, t_emb));
+
+            auto s_cond = ggml_reshape_2d(ctx->ggml_ctx, s, params_cfg.hidden_size, L * B);
+            auto pixels = pixel_embedder->forward(ctx, x, params_cfg.patch_size, pixel_pos_full);
+            for (int i = 0; i < params_cfg.pixel_depth; ++i) {
+                auto block = std::dynamic_pointer_cast<PiTBlock>(blocks["pixel_blocks." + std::to_string(i)]);
+                pixels     = block->forward(ctx, pixels, s_cond, H, W, pixel_pos_comp);
+                sd::ggml_graph_cut::mark_graph_cut(pixels, "pid.pixel_blocks." + std::to_string(i), "pixels");
+            }
+
+            pixels   = final_layer->forward(ctx, pixels);
+            pixels   = ggml_reshape_3d(ctx->ggml_ctx, pixels, params_cfg.in_channels * P2, L, B);
+            auto out = DiT::unpatchify(ctx->ggml_ctx,
+                                       pixels,
+                                       Hs,
+                                       Ws,
+                                       static_cast<int>(params_cfg.patch_size),
+                                       static_cast<int>(params_cfg.patch_size),
+                                       false);
+            out      = ggml_ext_slice(ctx->ggml_ctx, out, 1, 0, H_orig);
+            out      = ggml_ext_slice(ctx->ggml_ctx, out, 0, 0, W_orig);
+            return out;
+        }
+    };
+
+    struct PiDRunner : public DiffusionModelRunner {
+        PixelDiTParams params_cfg;
+        PixelDiT model;
+        std::vector<float> pos_img_vec;
+        std::vector<float> pos_txt_vec;
+        std::vector<float> pixel_pos_vec;
+        std::vector<float> pixel_pos_comp_vec;
+
+        PiDRunner(ggml_backend_t backend,
+                  ggml_backend_t params_backend,
+                  const String2TensorStorage& tensor_storage_map,
+                  const std::string prefix = "model.diffusion_model")
+            : DiffusionModelRunner(backend, params_backend, prefix) {
+            for (const auto& pair : tensor_storage_map) {
+                const std::string& tensor_name = pair.first;
+                if (tensor_name.find(prefix) == std::string::npos) {
+                    continue;
+                }
+                size_t pos = tensor_name.find("patch_blocks.");
+                if (pos != std::string::npos) {
+                    auto items = split_string(tensor_name.substr(pos), '.');
+                    if (items.size() > 1) {
+                        int block_index        = atoi(items[1].c_str());
+                        params_cfg.patch_depth = std::max<int64_t>(params_cfg.patch_depth, block_index + 1);
+                    }
+                }
+                pos = tensor_name.find("pixel_blocks.");
+                if (pos != std::string::npos) {
+                    auto items = split_string(tensor_name.substr(pos), '.');
+                    if (items.size() > 1) {
+                        int block_index        = atoi(items[1].c_str());
+                        params_cfg.pixel_depth = std::max<int64_t>(params_cfg.pixel_depth, block_index + 1);
+                    }
+                }
+                if (tensor_name.find("lq_proj.latent_proj.0.weight") != std::string::npos) {
+                    params_cfg.lq_latent_channels    = pair.second.ne[2];
+                    params_cfg.lq_latent_down_factor = params_cfg.lq_latent_channels >= 64 ? 16 : 8;
+                }
+                if (tensor_name.find("patch_blocks.0.mlp_x.w1.weight") != std::string::npos) {
+                    params_cfg.patch_mlp_hidden_dim = pair.second.ne[1];
+                }
+            }
+            LOG_INFO("PiD params: patch_depth=%" PRId64 ", pixel_depth=%" PRId64 ", patch_mlp_hidden_dim=%" PRId64 ", lq_latent_channels=%" PRId64 ", lq_latent_down_factor=%" PRId64,
+                     params_cfg.patch_depth,
+                     params_cfg.pixel_depth,
+                     params_cfg.patch_mlp_hidden_dim,
+                     params_cfg.lq_latent_channels,
+                     params_cfg.lq_latent_down_factor);
+            model = PixelDiT(params_cfg);
+            model.init(params_ctx, tensor_storage_map, prefix);
+        }
+
+        std::string get_desc() override {
+            return "PiD";
+        }
+
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
+            model.get_param_tensors(tensors, prefix);
+        }
+
+        ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,
+                                 const sd::Tensor<float>& timesteps_tensor,
+                                 const sd::Tensor<float>& context_tensor,
+                                 const sd::Tensor<float>& lq_latent_tensor,
+                                 const sd::Tensor<float>& degrade_sigma_tensor) {
+            ggml_cgraph* gf            = new_graph_custom(PID_GRAPH_SIZE);
+            ggml_tensor* x             = make_input(x_tensor);
+            ggml_tensor* timesteps     = make_input(timesteps_tensor);
+            ggml_tensor* context       = make_input(context_tensor);
+            ggml_tensor* lq_latent     = make_input(lq_latent_tensor);
+            ggml_tensor* degrade_sigma = make_input(degrade_sigma_tensor);
+
+            int64_t W  = x->ne[0];
+            int64_t H  = x->ne[1];
+            int64_t B  = x->ne[3];
+            int64_t Wp = align_up(static_cast<int>(W), static_cast<int>(params_cfg.patch_size));
+            int64_t Hp = align_up(static_cast<int>(H), static_cast<int>(params_cfg.patch_size));
+            int64_t Hs = Hp / params_cfg.patch_size;
+            int64_t Ws = Wp / params_cfg.patch_size;
+
+            pos_img_vec  = make_rope_2d(static_cast<int>(Hs),
+                                        static_cast<int>(Ws),
+                                        static_cast<int>(params_cfg.hidden_size / params_cfg.num_groups),
+                                        10000.f,
+                                        16.f,
+                                        static_cast<int>(params_cfg.rope_ref_grid_h),
+                                        static_cast<int>(params_cfg.rope_ref_grid_w));
+            auto pos_img = ggml_new_tensor_4d(compute_ctx,
+                                              GGML_TYPE_F32,
+                                              2,
+                                              2,
+                                              params_cfg.hidden_size / params_cfg.num_groups / 2,
+                                              Hs * Ws);
+            set_backend_tensor_data(pos_img, pos_img_vec.data());
+
+            int64_t Ltxt = std::min<int64_t>(context->ne[1], params_cfg.txt_max_length);
+            pos_txt_vec  = make_rope_1d(static_cast<int>(Ltxt),
+                                        static_cast<int>(params_cfg.hidden_size / params_cfg.num_groups),
+                                        params_cfg.text_rope_theta);
+            auto pos_txt = ggml_new_tensor_4d(compute_ctx,
+                                              GGML_TYPE_F32,
+                                              2,
+                                              2,
+                                              params_cfg.hidden_size / params_cfg.num_groups / 2,
+                                              Ltxt);
+            set_backend_tensor_data(pos_txt, pos_txt_vec.data());
+
+            pixel_pos_vec  = make_pixel_abs_pos(static_cast<int>(Hp),
+                                                static_cast<int>(Wp),
+                                                static_cast<int>(params_cfg.pixel_hidden_size));
+            auto pixel_pos = ggml_new_tensor_3d(compute_ctx,
+                                                GGML_TYPE_F32,
+                                                params_cfg.pixel_hidden_size,
+                                                Wp * Hp,
+                                                1);
+            set_backend_tensor_data(pixel_pos, pixel_pos_vec.data());
+
+            pixel_pos_comp_vec  = make_rope_2d(static_cast<int>(Hs),
+                                               static_cast<int>(Ws),
+                                               static_cast<int>(params_cfg.pixel_attn_hidden_size / params_cfg.pixel_num_groups),
+                                               10000.f,
+                                               16.f,
+                                               static_cast<int>(params_cfg.rope_ref_grid_h),
+                                               static_cast<int>(params_cfg.rope_ref_grid_w));
+            auto pixel_pos_comp = ggml_new_tensor_4d(compute_ctx,
+                                                     GGML_TYPE_F32,
+                                                     2,
+                                                     2,
+                                                     params_cfg.pixel_attn_hidden_size / params_cfg.pixel_num_groups / 2,
+                                                     Hs * Ws);
+            set_backend_tensor_data(pixel_pos_comp, pixel_pos_comp_vec.data());
+
+            auto runner_ctx = get_context();
+            auto out        = model.forward(&runner_ctx,
+                                            x,
+                                            timesteps,
+                                            context,
+                                            lq_latent,
+                                            degrade_sigma,
+                                            pos_img,
+                                            pos_txt,
+                                            pixel_pos,
+                                            pixel_pos_comp);
+            ggml_build_forward_expand(gf, out);
+            return gf;
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const sd::Tensor<float>& x,
+                                  const sd::Tensor<float>& timesteps,
+                                  const sd::Tensor<float>& context,
+                                  const sd::Tensor<float>& lq_latent,
+                                  const sd::Tensor<float>& degrade_sigma) {
+            auto get_graph = [&]() -> ggml_cgraph* {
+                return build_graph(x, timesteps, context, lq_latent, degrade_sigma);
+            };
+            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), x.dim());
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const DiffusionParams& diffusion_params) override {
+            GGML_ASSERT(diffusion_params.x != nullptr);
+            GGML_ASSERT(diffusion_params.timesteps != nullptr);
+            GGML_ASSERT(diffusion_params.context != nullptr);
+            GGML_ASSERT(diffusion_params.ref_latents != nullptr);
+            GGML_ASSERT(!diffusion_params.ref_latents->empty());
+            auto degrade_sigma = sd::Tensor<float>::from_vector({0.0f});
+            return compute(n_threads,
+                           *diffusion_params.x,
+                           *diffusion_params.timesteps,
+                           *diffusion_params.context,
+                           diffusion_params.ref_latents->front(),
+                           degrade_sigma);
+        }
+    };
+}  // namespace Pid
+
+#endif  // __SD_PID_HPP__
@@ -411,13 +411,13 @@ public:

 public:
    PhotoMakerIDEncoder(ggml_backend_t backend,
-                        bool offload_params_to_cpu,
+                        ggml_backend_t params_backend,
                        const String2TensorStorage& tensor_storage_map,
                        const std::string prefix,
                        SDVersion version = VERSION_SDXL,
                        PMVersion pm_v    = PM_VERSION_1,
                        float sty         = 20.f)
-        : GGMLRunner(backend, offload_params_to_cpu),
+        : GGMLRunner(backend, params_backend),
          version(version),
          pm_version(pm_v),
          style_strength(sty) {
@@ -568,11 +568,11 @@ struct PhotoMakerIDEmbed : public GGMLRunner {
    bool applied     = false;

    PhotoMakerIDEmbed(ggml_backend_t backend,
-                      bool offload_params_to_cpu,
+                      ggml_backend_t params_backend,
                      ModelLoader* ml,
                      const std::string& file_path = "",
                      const std::string& prefix    = "")
-        : file_path(file_path), GGMLRunner(backend, offload_params_to_cpu), model_loader(ml) {
+        : file_path(file_path), GGMLRunner(backend, params_backend), model_loader(ml) {
        if (!model_loader->init_from_file_and_convert_name(file_path, prefix)) {
            load_failed = true;
        }
@@ -615,7 +615,10 @@ struct PhotoMakerIDEmbed : public GGMLRunner {
        };

        model_loader->load_tensors(on_new_tensor_cb, n_threads);
-        alloc_params_buffer();
+        if (!alloc_params_buffer()) {
+            LOG_ERROR("PhotoMaker ID embeds buffer allocation failed");
+            return false;
+        }

        dry_run = false;
        model_loader->load_tensors(on_new_tensor_cb, n_threads);
@@ -4,6 +4,7 @@
 #include <memory>

 #include "common_block.hpp"
+#include "diffusion_model.hpp"
 #include "flux.hpp"

 namespace Qwen {
@@ -479,7 +480,7 @@ namespace Qwen {
        }
    };

-    struct QwenImageRunner : public GGMLRunner {
+    struct QwenImageRunner : public DiffusionModelRunner {
    public:
        QwenImageParams qwen_image_params;
        QwenImageModel qwen_image;
@@ -488,12 +489,12 @@ namespace Qwen {
        SDVersion version;

        QwenImageRunner(ggml_backend_t backend,
-                        bool offload_params_to_cpu,
+                        ggml_backend_t params_backend,
                        const String2TensorStorage& tensor_storage_map = {},
                        const std::string prefix                       = "",
                        SDVersion version                              = VERSION_QWEN_IMAGE,
                        bool zero_cond_t                               = false)
-            : GGMLRunner(backend, offload_params_to_cpu) {
+            : DiffusionModelRunner(backend, params_backend, prefix) {
            qwen_image_params.num_layers  = 0;
            qwen_image_params.zero_cond_t = zero_cond_t;
            for (auto pair : tensor_storage_map) {
@@ -528,7 +529,7 @@ namespace Qwen {
            return "qwen_image";
        }

-        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
            qwen_image.get_param_tensors(tensors, prefix);
        }

@@ -624,6 +625,19 @@ namespace Qwen {
            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), x.dim());
        }

+        sd::Tensor<float> compute(int n_threads,
+                                  const DiffusionParams& diffusion_params) override {
+            GGML_ASSERT(diffusion_params.x != nullptr);
+            GGML_ASSERT(diffusion_params.timesteps != nullptr);
+            static const std::vector<sd::Tensor<float>> empty_ref_latents;
+            return compute(n_threads,
+                           *diffusion_params.x,
+                           *diffusion_params.timesteps,
+                           tensor_or_empty(diffusion_params.context),
+                           diffusion_params.ref_latents ? *diffusion_params.ref_latents : empty_ref_latents,
+                           diffusion_params.increase_ref_index);
+        }
+
        void test() {
            ggml_init_params params;
            params.mem_size   = static_cast<size_t>(1024 * 1024) * 1024;  // 1GB
@@ -669,7 +683,7 @@ namespace Qwen {
            // cuda q8: pass
            // cuda q8 fa: pass
            // ggml_backend_t backend    = ggml_backend_cuda_init(0);
-            ggml_backend_t backend    = ggml_backend_cpu_init();
+            ggml_backend_t backend    = sd_backend_cpu_init();
            ggml_type model_data_type = GGML_TYPE_Q8_0;

            ModelLoader model_loader;
@@ -686,12 +700,16 @@ namespace Qwen {
            }

            std::shared_ptr<QwenImageRunner> qwen_image = std::make_shared<QwenImageRunner>(backend,
-                                                                                            false,
+                                                                                            backend,
                                                                                            tensor_storage_map,
                                                                                            "model.diffusion_model",
                                                                                            VERSION_QWEN_IMAGE);

-            qwen_image->alloc_params_buffer();
+            if (!qwen_image->alloc_params_buffer()) {
+                LOG_ERROR("qwen_image buffer allocation failed");
+                return;
+            }
+
            std::map<std::string, ggml_tensor*> tensors;
            qwen_image->get_param_tensors(tensors, "model.diffusion_model");

@@ -111,6 +111,16 @@ namespace Rope {
        return txt_ids;
    }

+    __STATIC_INLINE__ std::vector<std::vector<float>> gen_longcat_txt_ids(int bs, int context_len, int axes_dim_num) {
+        auto txt_ids = std::vector<std::vector<float>>(bs * context_len, std::vector<float>(axes_dim_num, 0.0f));
+        for (int i = 0; i < bs * context_len; i++) {
+            float token_index = static_cast<float>(i % context_len);
+            txt_ids[i][1]     = token_index;
+            txt_ids[i][2]     = token_index;
+        }
+        return txt_ids;
+    }
+
    __STATIC_INLINE__ std::vector<std::vector<float>> gen_flux_img_ids(int h,
                                                                       int w,
                                                                       int patch_size,
@@ -122,7 +132,6 @@ namespace Rope {
                                                                       bool scale_rope = false) {
        int h_len = (h + (patch_size / 2)) / patch_size;
        int w_len = (w + (patch_size / 2)) / patch_size;
-
        std::vector<std::vector<float>> img_ids(h_len * w_len, std::vector<float>(axes_dim_num, 0.0));

        int h_start = h_offset;
@@ -135,7 +144,6 @@ namespace Rope {

        std::vector<float> row_ids = linspace<float>(1.f * h_start, 1.f * h_start + h_len - 1, h_len);
        std::vector<float> col_ids = linspace<float>(1.f * w_start, 1.f * w_start + w_len - 1, w_len);
-
        for (int i = 0; i < h_len; ++i) {
            for (int j = 0; j < w_len; ++j) {
                img_ids[i * w_len + j][0] = 1.f * index;
@@ -241,17 +249,106 @@ namespace Rope {
        return embed_nd(ids, bs, axis_thetas, axes_dim, wrap_dims, layout);
    }

+    __STATIC_INLINE__ std::vector<float> embed_interleaved_mrope(const std::vector<std::vector<float>>& ids,
+                                                                 int bs,
+                                                                 float theta,
+                                                                 int head_dim,
+                                                                 const std::vector<int>& mrope_section) {
+        GGML_ASSERT(bs > 0);
+        GGML_ASSERT(head_dim % 2 == 0);
+        GGML_ASSERT(mrope_section.size() >= 3);
+
+        std::vector<std::vector<float>> trans_ids = transpose(ids);
+        size_t pos_len                            = ids.size() / bs;
+        int half_dim                              = head_dim / 2;
+
+        std::vector<std::vector<std::vector<float>>> axis_embs;
+        axis_embs.reserve(3);
+        for (int axis = 0; axis < 3; ++axis) {
+            axis_embs.push_back(rope(trans_ids[axis], head_dim, theta));
+        }
+
+        std::vector<std::vector<float>> emb = axis_embs[0];
+        for (int axis = 1; axis < 3; ++axis) {
+            int length = std::min<int>(mrope_section[axis] * 3, half_dim);
+            for (int freq_idx = axis; freq_idx < length; freq_idx += 3) {
+                for (size_t pos_idx = 0; pos_idx < bs * pos_len; ++pos_idx) {
+                    for (int k = 0; k < 4; ++k) {
+                        emb[pos_idx][4 * freq_idx + k] = axis_embs[axis][pos_idx][4 * freq_idx + k];
+                    }
+                }
+            }
+        }
+
+        return flatten(emb);
+    }
+
+    __STATIC_INLINE__ std::vector<float> embed_2d_interleaved(int height,
+                                                              int width,
+                                                              int dim,
+                                                              float theta    = 10000.f,
+                                                              float scale    = 16.f,
+                                                              int ref_grid_h = 0,
+                                                              int ref_grid_w = 0) {
+        assert(dim % 4 == 0);
+        int half_dim      = dim / 2;
+        int dim_axis      = dim / 2;
+        int axis_half_dim = dim_axis / 2;
+
+        float h_ntk = 1.f;
+        float w_ntk = 1.f;
+        if (ref_grid_h > 0 && ref_grid_w > 0 && dim_axis > 2) {
+            float power = static_cast<float>(dim_axis) / static_cast<float>(dim_axis - 2);
+            h_ntk       = std::pow(static_cast<float>(height) / static_cast<float>(ref_grid_h), power);
+            w_ntk       = std::pow(static_cast<float>(width) / static_cast<float>(ref_grid_w), power);
+        }
+
+        std::vector<float> x_pos;
+        std::vector<float> y_pos;
+        x_pos.reserve(static_cast<size_t>(height) * width);
+        y_pos.reserve(static_cast<size_t>(height) * width);
+        for (int iy = 0; iy < height; ++iy) {
+            float y = height == 1 ? 0.f : scale * static_cast<float>(iy) / static_cast<float>(height - 1);
+            for (int ix = 0; ix < width; ++ix) {
+                float x = width == 1 ? 0.f : scale * static_cast<float>(ix) / static_cast<float>(width - 1);
+                x_pos.push_back(x);
+                y_pos.push_back(y);
+            }
+        }
+
+        auto x_emb = rope(x_pos, dim_axis, theta * w_ntk);
+        auto y_emb = rope(y_pos, dim_axis, theta * h_ntk);
+
+        std::vector<float> out(static_cast<size_t>(height) * width * half_dim * 4);
+        for (int pos = 0; pos < height * width; ++pos) {
+            for (int i = 0; i < axis_half_dim; ++i) {
+                int jx        = 2 * i;
+                int jy        = 2 * i + 1;
+                size_t base_x = static_cast<size_t>(pos) * half_dim * 4 + static_cast<size_t>(jx) * 4;
+                size_t base_y = static_cast<size_t>(pos) * half_dim * 4 + static_cast<size_t>(jy) * 4;
+                size_t axis   = static_cast<size_t>(i) * 4;
+                for (int k = 0; k < 4; ++k) {
+                    out[base_x + k] = x_emb[pos][axis + k];
+                    out[base_y + k] = y_emb[pos][axis + k];
+                }
+            }
+        }
+        return out;
+    }
+
    __STATIC_INLINE__ std::vector<std::vector<float>> gen_refs_ids(int patch_size,
                                                                   int bs,
                                                                   int axes_dim_num,
+                                                                   int start_index,
                                                                   const std::vector<ggml_tensor*>& ref_latents,
                                                                   bool increase_ref_index,
                                                                   float ref_index_scale,
-                                                                   bool scale_rope) {
+                                                                   bool scale_rope,
+                                                                   int base_offset = 0) {
        std::vector<std::vector<float>> ids;
        int curr_h_offset = 0;
        int curr_w_offset = 0;
-        int index         = 1;
+        int index         = start_index;
        for (ggml_tensor* ref : ref_latents) {
            int h_offset = 0;
            int w_offset = 0;
@@ -270,8 +367,8 @@ namespace Rope {
                                            bs,
                                            axes_dim_num,
                                            static_cast<int>(index * ref_index_scale),
-                                            h_offset,
-                                            w_offset,
+                                            h_offset + base_offset,
+                                            w_offset + base_offset,
                                            scale_rope);
            ids          = concat_ids(ids, ref_ids, bs);

@@ -294,13 +391,17 @@ namespace Rope {
                                                                   std::set<int> txt_arange_dims,
                                                                   const std::vector<ggml_tensor*>& ref_latents,
                                                                   bool increase_ref_index,
-                                                                   float ref_index_scale) {
-        auto txt_ids = gen_flux_txt_ids(bs, context_len, axes_dim_num, txt_arange_dims);
-        auto img_ids = gen_flux_img_ids(h, w, patch_size, bs, axes_dim_num);
+                                                                   float ref_index_scale,
+                                                                   bool is_longcat) {
+        int x_index = is_longcat ? 1 : 0;
+
+        auto txt_ids = is_longcat ? gen_longcat_txt_ids(bs, context_len, axes_dim_num) : gen_flux_txt_ids(bs, context_len, axes_dim_num, txt_arange_dims);
+        int offset   = is_longcat ? context_len : 0;
+        auto img_ids = gen_flux_img_ids(h, w, patch_size, bs, axes_dim_num, x_index, offset, offset);

        auto ids = concat_ids(txt_ids, img_ids, bs);
        if (ref_latents.size() > 0) {
-            auto refs_ids = gen_refs_ids(patch_size, bs, axes_dim_num, ref_latents, increase_ref_index, ref_index_scale, false);
+            auto refs_ids = gen_refs_ids(patch_size, bs, axes_dim_num, x_index + 1, ref_latents, increase_ref_index, ref_index_scale, false, offset);
            ids           = concat_ids(ids, refs_ids, bs);
        }
        return ids;
@@ -319,7 +420,8 @@ namespace Rope {
                                                     int theta,
                                                     bool circular_h,
                                                     bool circular_w,
-                                                     const std::vector<int>& axes_dim) {
+                                                     const std::vector<int>& axes_dim,
+                                                     bool is_longcat) {
        std::vector<std::vector<float>> ids = gen_flux_ids(h,
                                                           w,
                                                           patch_size,
@@ -329,7 +431,8 @@ namespace Rope {
                                                           txt_arange_dims,
                                                           ref_latents,
                                                           increase_ref_index,
-                                                           ref_index_scale);
+                                                           ref_index_scale,
+                                                           is_longcat);
        std::vector<std::vector<int>> wrap_dims;
        if ((circular_h || circular_w) && bs > 0 && axes_dim.size() >= 3) {
            int h_len = (h + (patch_size / 2)) / patch_size;
@@ -394,7 +497,7 @@ namespace Rope {
        auto img_ids     = gen_flux_img_ids(h, w, patch_size, bs, axes_dim_num, 0, 0, 0, true);
        auto ids         = concat_ids(txt_ids_repeated, img_ids, bs);
        if (ref_latents.size() > 0) {
-            auto refs_ids = gen_refs_ids(patch_size, bs, axes_dim_num, ref_latents, increase_ref_index, 1.f, true);
+            auto refs_ids = gen_refs_ids(patch_size, bs, axes_dim_num, 1, ref_latents, increase_ref_index, 1.f, true);
            ids           = concat_ids(ids, refs_ids, bs);
        }
        return ids;
@@ -462,6 +565,52 @@ namespace Rope {
        return embed_nd(ids, bs, static_cast<float>(theta), axes_dim, wrap_dims);
    }

+    __STATIC_INLINE__ std::vector<std::vector<float>> gen_lens_ids(int h,
+                                                                   int w,
+                                                                   int bs,
+                                                                   int context_len,
+                                                                   bool scale_rope = true) {
+        auto img_ids_repeated = gen_flux_img_ids(h, w, 1, bs, 3, 0, 0, 0, scale_rope);
+
+        int txt_id_start = scale_rope ? std::max(h / 2, w / 2) : 0;
+        auto txt_ids     = linspace<float>(1.f * txt_id_start, 1.f * context_len + txt_id_start, context_len);
+        std::vector<std::vector<float>> txt_ids_repeated(bs * context_len, std::vector<float>(3));
+        for (int i = 0; i < bs; ++i) {
+            for (int j = 0; j < txt_ids.size(); ++j) {
+                txt_ids_repeated[i * txt_ids.size() + j] = {txt_ids[j], txt_ids[j], txt_ids[j]};
+            }
+        }
+
+        return concat_ids(img_ids_repeated, txt_ids_repeated, bs);
+    }
+
+    __STATIC_INLINE__ std::vector<float> gen_lens_pe(int h,
+                                                     int w,
+                                                     int bs,
+                                                     int context_len,
+                                                     int theta,
+                                                     bool circular_h,
+                                                     bool circular_w,
+                                                     const std::vector<int>& axes_dim) {
+        std::vector<std::vector<float>> ids = gen_lens_ids(h, w, bs, context_len, true);
+        std::vector<std::vector<int>> wrap_dims;
+        if ((circular_h || circular_w) && bs > 0 && axes_dim.size() >= 3) {
+            size_t pos_len = ids.size() / bs;
+            wrap_dims.assign(axes_dim.size(), std::vector<int>(pos_len, 0));
+            const size_t img_tokens = static_cast<size_t>(h) * static_cast<size_t>(w);
+            for (size_t token_i = 0; token_i < img_tokens; ++token_i) {
+                if (circular_h) {
+                    wrap_dims[1][token_i] = h;
+                }
+                if (circular_w) {
+                    wrap_dims[2][token_i] = w;
+                }
+            }
+        }
+
+        return embed_nd(ids, bs, static_cast<float>(theta), axes_dim, wrap_dims);
+    }
+
    __STATIC_INLINE__ std::vector<std::vector<float>> gen_ernie_image_ids(int h,
                                                                          int w,
                                                                          int patch_size,
@@ -321,11 +321,11 @@ struct T5Runner : public GGMLRunner {
    std::vector<int> relative_position_bucket_vec;

    T5Runner(ggml_backend_t backend,
-             bool offload_params_to_cpu,
+             ggml_backend_t params_backend,
             const String2TensorStorage& tensor_storage_map,
             const std::string prefix,
             bool is_umt5 = false)
-        : GGMLRunner(backend, offload_params_to_cpu) {
+        : GGMLRunner(backend, params_backend) {
        if (is_umt5) {
            params.vocab_size         = 256384;
            params.relative_attention = false;
@@ -464,19 +464,22 @@ struct T5Embedder {
    T5Runner model;

    T5Embedder(ggml_backend_t backend,
-               bool offload_params_to_cpu,
+               ggml_backend_t params_backend,
               const String2TensorStorage& tensor_storage_map = {},
               const std::string prefix                       = "",
               bool is_umt5                                   = false)
-        : model(backend, offload_params_to_cpu, tensor_storage_map, prefix, is_umt5), tokenizer(is_umt5) {
+        : model(backend, params_backend, tensor_storage_map, prefix, is_umt5), tokenizer(is_umt5) {
    }

    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
        model.get_param_tensors(tensors, prefix);
    }

-    void alloc_params_buffer() {
-        model.alloc_params_buffer();
+    bool alloc_params_buffer() {
+        if (!model.alloc_params_buffer()) {
+            return false;
+        }
+        return true;
    }

    std::tuple<std::vector<int>, std::vector<float>, std::vector<float>> tokenize(std::string text,
@@ -560,7 +563,7 @@ struct T5Embedder {
        // cuda f32: pass
        // cuda q8_0: pass
        // ggml_backend_t backend = ggml_backend_cuda_init(0);
-        ggml_backend_t backend    = ggml_backend_cpu_init();
+        ggml_backend_t backend    = sd_backend_cpu_init();
        ggml_type model_data_type = GGML_TYPE_F16;

        ModelLoader model_loader;
@@ -576,9 +579,12 @@ struct T5Embedder {
            }
        }

-        std::shared_ptr<T5Embedder> t5 = std::make_shared<T5Embedder>(backend, false, tensor_storage_map, "", true);
+        std::shared_ptr<T5Embedder> t5 = std::make_shared<T5Embedder>(backend, backend, tensor_storage_map, "", true);

-        t5->alloc_params_buffer();
+        if (!t5->alloc_params_buffer()) {
+            LOG_ERROR("t5 params buffer allocation failed");
+            return;
+        }
        std::map<std::string, ggml_tensor*> tensors;
        t5->get_param_tensors(tensors, "");

@@ -2,7 +2,6 @@
 #define __TAE_HPP__

 #include "ggml_extend.hpp"
-
 #include "model.h"

 /*
@@ -260,10 +259,54 @@ public:
    }
 };

-ggml_tensor* patchify(ggml_context* ctx,
-                      ggml_tensor* x,
-                      int64_t patch_size,
-                      int64_t b = 1) {
+class WideMemBlock : public GGMLBlock {
+    bool has_skip_conv = false;
+
+public:
+    WideMemBlock(int channels, int out_channels)
+        : has_skip_conv(channels != out_channels) {
+        int groups       = std::max(1, out_channels / 64);
+        blocks["conv.0"] = std::shared_ptr<GGMLBlock>(new Conv2d(channels * 2, out_channels, {1, 1}, {1, 1}));
+        blocks["conv.2"] = std::shared_ptr<GGMLBlock>(new Conv2d_grouped(out_channels, out_channels, groups, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv.4"] = std::shared_ptr<GGMLBlock>(new Conv2d(out_channels, out_channels, {1, 1}, {1, 1}));
+        blocks["conv.6"] = std::shared_ptr<GGMLBlock>(new Conv2d_grouped(out_channels, out_channels, groups, {3, 3}, {1, 1}, {1, 1}));
+        if (has_skip_conv) {
+            blocks["skip"] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, out_channels, {1, 1}, {1, 1}, {0, 0}, {1, 1}, false));
+        }
+    }
+
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* past) {
+        // x: [n, channels, h, w]
+        auto conv0 = std::dynamic_pointer_cast<Conv2d>(blocks["conv.0"]);
+        auto conv1 = std::dynamic_pointer_cast<Conv2d_grouped>(blocks["conv.2"]);
+        auto conv2 = std::dynamic_pointer_cast<Conv2d>(blocks["conv.4"]);
+        auto conv3 = std::dynamic_pointer_cast<Conv2d_grouped>(blocks["conv.6"]);
+
+        auto h = ggml_concat(ctx->ggml_ctx, x, past, 2);
+        h      = conv0->forward(ctx, h);
+        h      = ggml_relu_inplace(ctx->ggml_ctx, h);
+        h      = conv1->forward(ctx, h);
+        h      = ggml_relu_inplace(ctx->ggml_ctx, h);
+        h      = conv2->forward(ctx, h);
+        h      = ggml_relu_inplace(ctx->ggml_ctx, h);
+        h      = conv3->forward(ctx, h);
+
+        auto skip = x;
+        if (has_skip_conv) {
+            auto skip_conv = std::dynamic_pointer_cast<Conv2d>(blocks["skip"]);
+            skip           = skip_conv->forward(ctx, x);
+        }
+        h = ggml_add_inplace(ctx->ggml_ctx, h, skip);
+        h = ggml_relu_inplace(ctx->ggml_ctx, h);
+        return h;
+    }
+};
+
+ggml_tensor*
+patchify(ggml_context* ctx,
+         ggml_tensor* x,
+         int64_t patch_size,
+         int64_t b = 1) {
    // x: [f, b*c, h*q, w*r]
    // return: [f, b*c*r*q, h, w]
    if (patch_size == 1) {
@@ -323,13 +366,20 @@ class TinyVideoEncoder : public UnaryBlock {
    int patch_size  = 1;

 public:
-    TinyVideoEncoder(int z_channels = 4, int patch_size = 1)
+    int t_downscale = 1;
+    TinyVideoEncoder(int z_channels = 4, int patch_size = 1, std::vector<bool> time_downscale = {true, true, false})
        : z_channels(z_channels), patch_size(patch_size) {
+        t_downscale = 1;
+        for (bool downscale : time_downscale) {
+            if (downscale) {
+                t_downscale *= 2;
+            }
+        }
        int index                       = 0;
        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels * patch_size * patch_size, hidden, {3, 3}, {1, 1}, {1, 1}));
        index++;  // nn.ReLU()
        for (int i = 0; i < num_layers; i++) {
-            int stride                      = i == num_layers - 1 ? 1 : 2;
+            int stride                      = time_downscale[i] ? 2 : 1;
            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TPool(hidden, stride));
            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(hidden, hidden, {3, 3}, {2, 2}, {1, 1}, {1, 1}, false));
            for (int j = 0; j < num_blocks; j++) {
@@ -376,17 +426,35 @@ class TinyVideoDecoder : public UnaryBlock {
    static const int num_layers  = 3;
    int channels[num_layers + 1] = {256, 128, 64, 64};
    int patch_size               = 1;
+    int t_upscale                = 1;
+    bool is_wide                 = false;

 public:
-    TinyVideoDecoder(int z_channels = 4, int patch_size = 1)
-        : z_channels(z_channels), patch_size(patch_size) {
+    TinyVideoDecoder(int z_channels = 4, int patch_size = 1, std::vector<bool> time_upscale = {false, true, true}, bool is_wide = false)
+        : z_channels(z_channels), patch_size(patch_size), is_wide(is_wide) {
+        t_upscale = 1;
+        if (is_wide) {
+            channels[0] = 1024;
+            channels[1] = 512;
+            channels[2] = 256;
+        }
+
+        for (bool upscale : time_upscale) {
+            if (upscale) {
+                t_upscale *= 2;
+            }
+        }
        int index                       = 1;  // Clamp()
        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(z_channels, channels[0], {3, 3}, {1, 1}, {1, 1}));
        index++;  // nn.ReLU()
        for (int i = 0; i < num_layers; i++) {
-            int stride = i == 0 ? 1 : 2;
+            int stride = time_upscale[i] ? 2 : 1;
            for (int j = 0; j < num_blocks; j++) {
-                blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new MemBlock(channels[i], channels[i]));
+                if (is_wide) {
+                    blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new WideMemBlock(channels[i], channels[i]));
+                } else {
+                    blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new MemBlock(channels[i], channels[i]));
+                }
            }
            index++;  // nn.Upsample()
            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TGrow(channels[i], stride));
@@ -411,10 +479,15 @@ public:
        int index = 3;
        for (int i = 0; i < num_layers; i++) {
            for (int j = 0; j < num_blocks; j++) {
-                auto block = std::dynamic_pointer_cast<MemBlock>(blocks[std::to_string(index++)]);
-                auto mem   = ggml_pad_ext(ctx->ggml_ctx, h, 0, 0, 0, 0, 0, 0, 1, 0);
-                mem        = ggml_view_4d(ctx->ggml_ctx, mem, h->ne[0], h->ne[1], h->ne[2], h->ne[3], h->nb[1], h->nb[2], h->nb[3], 0);
-                h          = block->forward(ctx, h, mem);
+                auto mem = ggml_pad_ext(ctx->ggml_ctx, h, 0, 0, 0, 0, 0, 0, 1, 0);
+                mem      = ggml_view_4d(ctx->ggml_ctx, mem, h->ne[0], h->ne[1], h->ne[2], h->ne[3], h->nb[1], h->nb[2], h->nb[3], 0);
+                if (is_wide) {
+                    auto block = std::dynamic_pointer_cast<WideMemBlock>(blocks[std::to_string(index++)]);
+                    h          = block->forward(ctx, h, mem);
+                } else {
+                    auto block = std::dynamic_pointer_cast<MemBlock>(blocks[std::to_string(index++)]);
+                    h          = block->forward(ctx, h, mem);
+                }
            }
            // upsample
            index++;
@@ -431,8 +504,8 @@ public:
        if (patch_size > 1) {
            h = unpatchify(ctx->ggml_ctx, h, patch_size, 1);
        }
-        // shape(W, H, 3, 3 + T) => shape(W, H, 3, T)
-        h = ggml_view_4d(ctx->ggml_ctx, h, h->ne[0], h->ne[1], h->ne[2], h->ne[3] - 3, h->nb[1], h->nb[2], h->nb[3], 3 * h->nb[3]);
+        // shape(W, H, 3, (t_upscale - 1) + T) => shape(W, H, 3, T)
+        h = ggml_view_4d(ctx->ggml_ctx, h, h->ne[0], h->ne[1], h->ne[2], h->ne[3] - (t_upscale - 1), h->nb[1], h->nb[2], h->nb[3], (t_upscale - 1) * h->nb[3]);
        return h;
    }
 };
@@ -441,32 +514,40 @@ class TAEHV : public GGMLBlock {
 protected:
    bool decode_only;
    SDVersion version;
+    bool is_wide;

 public:
-    int z_channels = 16;
+    int z_channels                   = 16;
+    std::vector<bool> time_downscale = {true, true, false};
+    std::vector<bool> time_upscale   = {false, true, true};

 public:
-    TAEHV(bool decode_only = true, SDVersion version = VERSION_WAN2)
-        : decode_only(decode_only), version(version) {
+    TAEHV(bool decode_only = true, SDVersion version = VERSION_WAN2, bool is_wide = false)
+        : decode_only(decode_only), version(version), is_wide(is_wide) {
        int patch = 1;
        if (version == VERSION_WAN2_2_TI2V) {
            z_channels = 48;
            patch      = 2;
+        } else if (sd_version_is_ltxav(version)) {
+            z_channels     = 128;
+            patch          = 4;
+            time_downscale = {true, true, true};
+            time_upscale   = {true, true, true};
        }
-        blocks["decoder"] = std::shared_ptr<GGMLBlock>(new TinyVideoDecoder(z_channels, patch));
+        blocks["decoder"] = std::shared_ptr<GGMLBlock>(new TinyVideoDecoder(z_channels, patch, time_upscale, is_wide));
        if (!decode_only) {
-            blocks["encoder"] = std::shared_ptr<GGMLBlock>(new TinyVideoEncoder(z_channels, patch));
+            blocks["encoder"] = std::shared_ptr<GGMLBlock>(new TinyVideoEncoder(z_channels, patch, time_downscale));
        }
    }

    ggml_tensor* decode(GGMLRunnerContext* ctx, ggml_tensor* z) {
        auto decoder = std::dynamic_pointer_cast<TinyVideoDecoder>(blocks["decoder"]);
-        if (sd_version_is_wan(version)) {
+        if (sd_version_is_wan(version) || sd_version_is_ltxav(version)) {
            // (W, H, C, T) -> (W, H, T, C)
            z = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, z, 0, 1, 3, 2));
        }
        auto result = decoder->forward(ctx, z);
-        if (sd_version_is_wan(version)) {
+        if (sd_version_is_wan(version) || sd_version_is_ltxav(version)) {
            // (W, H, C, T) -> (W, H, T, C)
            result = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, result, 0, 1, 3, 2));
        }
@@ -478,10 +559,10 @@ public:
        // (W, H, T, C) -> (W, H, C, T)
        x                  = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 1, 3, 2));
        int64_t num_frames = x->ne[3];
-        if (num_frames % 4) {
-            // pad to multiple of 4 at the end
+        if (num_frames % encoder->t_downscale) {
+            // pad to multiple of encoder->t_downscale at the end
            auto last_frame = ggml_view_4d(ctx->ggml_ctx, x, x->ne[0], x->ne[1], x->ne[2], 1, x->nb[1], x->nb[2], x->nb[3], (num_frames - 1) * x->nb[3]);
-            for (int i = 0; i < 4 - num_frames % 4; i++) {
+            for (int i = 0; i < encoder->t_downscale - num_frames % encoder->t_downscale; i++) {
                x = ggml_concat(ctx->ggml_ctx, x, last_frame, 3);
            }
        }
@@ -503,7 +584,7 @@ public:
    TAESD(bool decode_only = true, SDVersion version = VERSION_SD1)
        : decode_only(decode_only) {
        bool use_midblock_gn = false;
-        taef2                = sd_version_is_flux2(version);
+        taef2                = sd_version_uses_flux2_vae(version);

        if (sd_version_is_dit(version)) {
            z_channels = 16;
@@ -542,14 +623,14 @@ struct TinyImageAutoEncoder : public VAE {
    bool decode_only = false;

    TinyImageAutoEncoder(ggml_backend_t backend,
-                         bool offload_params_to_cpu,
+                         ggml_backend_t params_backend,
                         const String2TensorStorage& tensor_storage_map,
                         const std::string prefix,
                         bool decoder_only = true,
                         SDVersion version = VERSION_SD1)
        : decode_only(decoder_only),
          taesd(decoder_only, version),
-          VAE(version, backend, offload_params_to_cpu) {
+          VAE(version, backend, params_backend) {
        scale_input = false;
        taesd.init(params_ctx, tensor_storage_map, prefix);
    }
@@ -602,16 +683,23 @@ struct TinyImageAutoEncoder : public VAE {
 struct TinyVideoAutoEncoder : public VAE {
    TAEHV taehv;
    bool decode_only = false;
+    bool is_wide     = false;

    TinyVideoAutoEncoder(ggml_backend_t backend,
-                         bool offload_params_to_cpu,
+                         ggml_backend_t params_backend,
                         const String2TensorStorage& tensor_storage_map,
                         const std::string prefix,
                         bool decoder_only = true,
                         SDVersion version = VERSION_WAN2)
        : decode_only(decoder_only),
-          taehv(decoder_only, version),
-          VAE(version, backend, offload_params_to_cpu) {
+          VAE(version, backend, params_backend) {
+        for (auto tensor_storage : tensor_storage_map) {
+            if (tensor_storage.first.find(prefix + ".3.conv.6.weight") != std::string::npos) {
+                is_wide = true;
+                break;
+            }
+        }
+        taehv       = TAEHV(decoder_only, version, is_wide);
        scale_input = false;
        taehv.init(params_ctx, tensor_storage_map, prefix);
    }
@@ -642,7 +730,8 @@ struct TinyVideoAutoEncoder : public VAE {
    }

    ggml_cgraph* build_graph(const sd::Tensor<float>& z_tensor, bool decode_graph) {
-        ggml_cgraph* gf  = ggml_new_graph(compute_ctx);
+        ggml_cgraph* gf  = decode_graph && is_wide ? ggml_new_graph_custom(compute_ctx, 4096, false)
+                                                   : ggml_new_graph(compute_ctx);
        ggml_tensor* z   = make_input(z_tensor);
        auto runner_ctx  = get_context();
        ggml_tensor* out = decode_graph ? taehv.decode(&runner_ctx, z) : taehv.encode(&runner_ctx, z);
@@ -235,6 +235,7 @@ namespace sd {

        Tensor& masked_fill_(const Tensor<uint8_t>& mask, const T& value);

+        T sum() const;
        T mean() const;

        static Tensor zeros(std::vector<int64_t> shape) {
@@ -327,6 +328,24 @@ namespace sd {
        std::vector<int64_t> shape_;
    };

+    template <typename T>
+    inline T Tensor<T>::sum() const {
+        T total = T{};
+        for (const T& value : data_) {
+            total += value;
+        }
+        return total;
+    }
+
+    template <>
+    inline float Tensor<float>::sum() const {
+        double total = 0.0;
+        for (float value : data_) {
+            total += static_cast<double>(value);
+        }
+        return static_cast<float>(total);
+    }
+
    template <typename T>
    inline T Tensor<T>::mean() const {
        if (empty()) {
@@ -104,7 +104,7 @@ namespace sd {
            throw std::invalid_argument("tensor file type does not match requested sd::Tensor type");
        }

-        std::vector<int64_t> shape(4, 1);
+        std::vector<int64_t> shape(n_dims, 1);
        for (int i = 0; i < n_dims; ++i) {
            int32_t dim = 1;
            file.read(reinterpret_cast<char*>(&dim), sizeof(dim));
@@ -131,11 +131,10 @@ std::vector<std::u32string> BPETokenizer::bpe(const std::u32string& token) const
 }

 std::vector<int> BPETokenizer::encode(const std::string& text, on_new_token_cb_t on_new_token_cb) {
-    std::string normalized_text = normalize(text);
    std::vector<int32_t> bpe_tokens;
    std::vector<std::string> token_strs;

-    auto splited_texts = split_with_special_tokens(normalized_text, special_tokens);
+    auto splited_texts = split_with_special_tokens(text, special_tokens);

    for (auto& splited_text : splited_texts) {
        if (is_special_token(splited_text)) {
@@ -160,15 +159,40 @@ std::vector<int> BPETokenizer::encode(const std::string& text, on_new_token_cb_t
                }
            }

-            std::string token_str = token;
+            std::string token_str = normalize(token);
            std::u32string utf32_token;
-            for (int i = 0; i < static_cast<int>(token_str.length()); i++) {
-                unsigned char b = token_str[i];
-                utf32_token += byte_encoder[b];
+            if (byte_level_bpe) {
+                for (int i = 0; i < token_str.length(); i++) {
+                    unsigned char b = token_str[i];
+                    utf32_token += byte_encoder[b];
+                }
+            } else {
+                utf32_token = utf8_to_utf32(token_str);
            }
            auto bpe_strs = bpe(utf32_token);
            for (auto bpe_str : bpe_strs) {
-                bpe_tokens.push_back(encoder[bpe_str]);
+                int token_id;
+                auto iter = encoder.find(bpe_str);
+                if (iter != encoder.end()) {
+                    token_id = iter->second;
+                } else {
+                    if (byte_fallback) {
+                        auto utf8_token_str = utf32_to_utf8(bpe_str);
+                        for (int i = 0; i < utf8_token_str.length(); i++) {
+                            unsigned char b = utf8_token_str[i];
+                            char hex_buf[16];
+                            snprintf(hex_buf, sizeof(hex_buf), "<0x%02X>", b);
+                            iter     = encoder.find(utf8_to_utf32(hex_buf));
+                            token_id = iter != encoder.end() ? iter->second : UNK_TOKEN_ID;
+                            bpe_tokens.push_back(token_id);
+                            token_strs.push_back(hex_buf);
+                        }
+                        continue;
+                    } else {
+                        token_id = UNK_TOKEN_ID;
+                    }
+                }
+                bpe_tokens.push_back(token_id);
                token_strs.push_back(utf32_to_utf8(bpe_str));
            }
        }
@@ -20,8 +20,10 @@ protected:
    std::map<std::u32string, int> encoder;
    std::map<int, std::u32string> decoder;
    std::map<std::pair<std::u32string, std::u32string>, int> bpe_ranks;
-    int encoder_len = 0;
-    int bpe_len     = 0;
+    int encoder_len     = 0;
+    int bpe_len         = 0;
+    bool byte_level_bpe = true;
+    bool byte_fallback  = false;

 protected:
    static std::vector<std::pair<int, std::u32string>> bytes_to_unicode();
@@ -0,0 +1,352 @@
+#include "gemma_tokenizer.h"
+
+#include "ggml.h"
+#include "json.hpp"
+#include "util.h"
+#include "vocab/vocab.h"
+
+std::string GemmaTokenizer::normalize(const std::string& text) const {
+    std::string normalized = text;
+    size_t pos             = 0;
+    while ((pos = normalized.find(' ', pos)) != std::string::npos) {
+        normalized.replace(pos, 1, "\xE2\x96\x81");
+        pos += 3;
+    }
+    return normalized;
+}
+
+void GemmaTokenizer::load_from_merges(const std::string& merges_utf8_str, const std::string& vocab_utf8_str) {
+    nlohmann::json vocab;
+    try {
+        vocab = nlohmann::json::parse(vocab_utf8_str);
+    } catch (const nlohmann::json::parse_error&) {
+        GGML_ABORT("invalid vocab json str");
+    }
+    for (const auto& [key, value] : vocab.items()) {
+        std::u32string token = utf8_to_utf32(key);
+        int i                = value;
+        encoder[token]       = i;
+        decoder[i]           = token;
+    }
+    encoder_len = static_cast<int>(vocab.size());
+    LOG_DEBUG("vocab size: %d", encoder_len);
+
+    std::vector<std::u32string> merges = split_utf32(merges_utf8_str);
+    std::vector<std::pair<std::u32string, std::u32string>> merge_pairs;
+    for (const auto& merge : merges) {
+        size_t space_pos = merge.find(' ');
+        merge_pairs.emplace_back(merge.substr(0, space_pos), merge.substr(space_pos + 1));
+    }
+    LOG_DEBUG("merges size %zu", merge_pairs.size());
+
+    int rank = 0;
+    for (const auto& merge : merge_pairs) {
+        bpe_ranks[merge] = rank++;
+    }
+    bpe_len = rank;
+}
+
+GemmaTokenizer::GemmaTokenizer(const std::string& merges_utf8_str, const std::string& vocab_utf8_str) {
+    byte_level_bpe = false;
+    byte_fallback  = true;
+    add_bos_token  = true;
+    pad_left       = true;
+    PAD_TOKEN      = "<pad>";
+    EOS_TOKEN      = "<eos>";
+    BOS_TOKEN      = "<bos>";
+    UNK_TOKEN      = "<unk>";
+
+    PAD_TOKEN_ID = 0;
+    EOS_TOKEN_ID = 1;
+    BOS_TOKEN_ID = 2;
+    UNK_TOKEN_ID = 3;
+
+    std::vector<std::string> special_tokens_before_merge = {
+        PAD_TOKEN,
+        EOS_TOKEN,
+        BOS_TOKEN,
+        UNK_TOKEN,
+        "<mask>",
+        "[multimodal]",
+    };
+    for (int i = 0; i <= 98; i++) {
+        special_tokens_before_merge.push_back("<unused" + std::to_string(i) + ">");
+    }
+    special_tokens_before_merge.push_back("<start_of_turn>");
+    special_tokens_before_merge.push_back("<end_of_turn>");
+    for (int i = 1; i <= 31; i++) {
+        special_tokens_before_merge.push_back(std::string(i, '\n'));
+    }
+    for (int i = 2; i <= 31; i++) {
+        std::string whitespace_token;
+        for (int j = 0; j < i; j++) {
+            whitespace_token += "\xE2\x96\x81";
+        }
+        special_tokens_before_merge.push_back(whitespace_token);
+    }
+    std::vector<std::string> html_tokens = {
+        "<table>",
+        "<caption>",
+        "<thead>",
+        "<tbody>",
+        "<tfoot>",
+        "<tr>",
+        "<th>",
+        "<td>",
+        "</table>",
+        "</caption>",
+        "</thead>",
+        "</tbody>",
+        "</tfoot>",
+        "</tr>",
+        "</th>",
+        "</td>",
+        "<h1>",
+        "<h2>",
+        "<h3>",
+        "<h4>",
+        "<h5>",
+        "<h6>",
+        "<blockquote>",
+        "</h1>",
+        "</h2>",
+        "</h3>",
+        "</h4>",
+        "</h5>",
+        "</h6>",
+        "</blockquote>",
+        "<strong>",
+        "<em>",
+        "<b>",
+        "<i>",
+        "<u>",
+        "<s>",
+        "<sub>",
+        "<sup>",
+        "<code>",
+        "</strong>",
+        "</em>",
+        "</b>",
+        "</i>",
+        "</u>",
+        "</s>",
+        "</sub>",
+        "</sup>",
+        "</code>",
+        "<a>",
+        "<html>",
+        "<body>",
+        "<img>",
+        "<span>",
+        "<bbox>",
+        "<ul>",
+        "<li>",
+        "<div>",
+        "<iframe>",
+        "<footer>",
+        "</a>",
+        "</html>",
+        "</body>",
+        "</img>",
+        "</span>",
+        "</bbox>",
+        "</ul>",
+        "</li>",
+        "</div>",
+        "</iframe>",
+        "</footer>",
+    };
+    special_tokens_before_merge.insert(special_tokens_before_merge.end(),
+                                       html_tokens.begin(),
+                                       html_tokens.end());
+    for (int i = 0; i <= 0xFF; i++) {
+        char hex_buf[16];
+        snprintf(hex_buf, sizeof(hex_buf), "<0x%02X>", i);
+        special_tokens_before_merge.push_back(hex_buf);
+    }
+
+    std::vector<std::string> special_tokens_after_merge = {
+        "<start_of_image>",
+        "<end_of_image>",
+    };
+    for (int i = 1; i <= 31; i++) {
+        special_tokens_after_merge.insert(special_tokens_after_merge.begin() + i - 1,
+                                          std::string(i, '\t'));
+    }
+    for (int i = 99; i <= 6241; i++) {
+        special_tokens_after_merge.push_back("<unused" + std::to_string(i) + ">");
+    }
+    special_tokens_after_merge.push_back("<image_soft_token>");
+
+    special_tokens = special_tokens_before_merge;
+    special_tokens.insert(special_tokens.end(),
+                          special_tokens_after_merge.begin(),
+                          special_tokens_after_merge.end());
+
+    if (merges_utf8_str.size() > 0 && vocab_utf8_str.size() > 0) {
+        load_from_merges(merges_utf8_str, vocab_utf8_str);
+    } else {
+        load_from_merges(load_gemma_merges(), load_gemma_vocab_json());
+    }
+}
+
+std::string Gemma2Tokenizer::normalize(const std::string& text) const {
+    std::string normalized = text;
+    size_t pos             = 0;
+    while ((pos = normalized.find(' ', pos)) != std::string::npos) {
+        normalized.replace(pos, 1, "\xE2\x96\x81");
+        pos += 3;
+    }
+    return normalized;
+}
+
+void Gemma2Tokenizer::load_from_merges(const std::string& merges_utf8_str, const std::string& vocab_utf8_str) {
+    nlohmann::json vocab;
+    try {
+        vocab = nlohmann::json::parse(vocab_utf8_str);
+    } catch (const nlohmann::json::parse_error&) {
+        GGML_ABORT("invalid vocab json str");
+    }
+    for (const auto& [key, value] : vocab.items()) {
+        std::u32string token = utf8_to_utf32(key);
+        int i                = value;
+        encoder[token]       = i;
+        decoder[i]           = token;
+    }
+    encoder_len = static_cast<int>(vocab.size());
+    LOG_DEBUG("vocab size: %d", encoder_len);
+
+    std::vector<std::u32string> merges = split_utf32(merges_utf8_str);
+    std::vector<std::pair<std::u32string, std::u32string>> merge_pairs;
+    for (const auto& merge : merges) {
+        size_t space_pos = merge.find(' ');
+        merge_pairs.emplace_back(merge.substr(0, space_pos), merge.substr(space_pos + 1));
+    }
+    LOG_DEBUG("merges size %zu", merge_pairs.size());
+
+    int rank = 0;
+    for (const auto& merge : merge_pairs) {
+        bpe_ranks[merge] = rank++;
+    }
+    bpe_len = rank;
+}
+
+Gemma2Tokenizer::Gemma2Tokenizer(const std::string& merges_utf8_str, const std::string& vocab_utf8_str) {
+    byte_level_bpe = false;
+    byte_fallback  = true;
+    add_bos_token  = true;
+    PAD_TOKEN      = "<pad>";
+    EOS_TOKEN      = "<eos>";
+    BOS_TOKEN      = "<bos>";
+    UNK_TOKEN      = "<unk>";
+
+    PAD_TOKEN_ID = 0;
+    EOS_TOKEN_ID = 1;
+    BOS_TOKEN_ID = 2;
+    UNK_TOKEN_ID = 3;
+
+    std::vector<std::string> special_tokens_before_merge = {
+        PAD_TOKEN,
+        EOS_TOKEN,
+        BOS_TOKEN,
+        UNK_TOKEN,
+        "<mask>",
+        "<2mass>",
+        "[@BOS@]",
+    };
+    for (int i = 0; i <= 98; i++) {
+        special_tokens_before_merge.push_back("<unused" + std::to_string(i) + ">");
+    }
+    special_tokens_before_merge.push_back("<start_of_turn>");
+    special_tokens_before_merge.push_back("<end_of_turn>");
+    for (int i = 1; i <= 31; i++) {
+        special_tokens_before_merge.push_back(std::string(i, '\n'));
+    }
+    for (int i = 2; i <= 31; i++) {
+        std::string whitespace_token;
+        for (int j = 0; j < i; j++) {
+            whitespace_token += "\xE2\x96\x81";
+        }
+        special_tokens_before_merge.push_back(whitespace_token);
+    }
+    std::vector<std::string> html_tokens = {
+        "<table>",
+        "<caption>",
+        "<thead>",
+        "<tbody>",
+        "<tfoot>",
+        "<tr>",
+        "<th>",
+        "<td>",
+        "</table>",
+        "</caption>",
+        "</thead>",
+        "</tbody>",
+        "</tfoot>",
+        "</tr>",
+        "</th>",
+        "</td>",
+        "<h1>",
+        "<h2>",
+        "<h3>",
+        "<h4>",
+        "<h5>",
+        "<h6>",
+        "<blockquote>",
+        "</h1>",
+        "</h2>",
+        "</h3>",
+        "</h4>",
+        "</h5>",
+        "</h6>",
+        "</blockquote>",
+        "<strong>",
+        "<em>",
+        "<b>",
+        "<i>",
+        "<u>",
+        "<s>",
+        "<sub>",
+        "<sup>",
+        "<code>",
+        "</strong>",
+        "</em>",
+        "</b>",
+        "</i>",
+        "</u>",
+        "</s>",
+        "</sub>",
+        "</sup>",
+        "</code>",
+    };
+    special_tokens_before_merge.insert(special_tokens_before_merge.end(),
+                                       html_tokens.begin(),
+                                       html_tokens.end());
+    for (int i = 0; i <= 0xFF; i++) {
+        char hex_buf[16];
+        snprintf(hex_buf, sizeof(hex_buf), "<0x%02X>", i);
+        special_tokens_before_merge.push_back(hex_buf);
+    }
+
+    std::vector<std::string> special_tokens_after_merge = {
+        "[toxicity=0]",
+    };
+    for (int i = 1; i <= 31; i++) {
+        special_tokens_after_merge.insert(special_tokens_after_merge.begin() + i - 1,
+                                          std::string(i, '\t'));
+    }
+    for (int i = 99; i <= 99; i++) {
+        special_tokens_after_merge.push_back("<unused" + std::to_string(i) + ">");
+    }
+
+    special_tokens = special_tokens_before_merge;
+    special_tokens.insert(special_tokens.end(),
+                          special_tokens_after_merge.begin(),
+                          special_tokens_after_merge.end());
+
+    if (merges_utf8_str.size() > 0 && vocab_utf8_str.size() > 0) {
+        load_from_merges(merges_utf8_str, vocab_utf8_str);
+    } else {
+        load_from_merges(load_gemma2_merges(), load_gemma2_vocab_json());
+    }
+}
@@ -0,0 +1,26 @@
+#ifndef __SD_TOKENIZERS_GEMMA_TOKENIZER_H__
+#define __SD_TOKENIZERS_GEMMA_TOKENIZER_H__
+
+#include <string>
+
+#include "bpe_tokenizer.h"
+
+class GemmaTokenizer : public BPETokenizer {
+protected:
+    void load_from_merges(const std::string& merges_utf8_str, const std::string& vocab_utf8_str);
+    std::string normalize(const std::string& text) const override;
+
+public:
+    explicit GemmaTokenizer(const std::string& merges_utf8_str = "", const std::string& vocab_utf8_str = "");
+};
+
+class Gemma2Tokenizer : public BPETokenizer {
+protected:
+    void load_from_merges(const std::string& merges_utf8_str, const std::string& vocab_utf8_str);
+    std::string normalize(const std::string& text) const override;
+
+public:
+    explicit Gemma2Tokenizer(const std::string& merges_utf8_str = "", const std::string& vocab_utf8_str = "");
+};
+
+#endif  // __SD_TOKENIZERS_GEMMA_TOKENIZER_H__
@@ -0,0 +1,91 @@
+#include "gpt_oss_tokenizer.h"
+
+#include "json.hpp"
+#include "util.h"
+#include "vocab/vocab.h"
+
+void GPTOSSTokenizer::load_from_merges(const std::string& merges_utf8_str, const std::string& vocab_utf8_str) {
+    auto byte_unicode_pairs = bytes_to_unicode();
+    byte_encoder            = std::map<int, std::u32string>(byte_unicode_pairs.begin(), byte_unicode_pairs.end());
+    for (auto& pair : byte_unicode_pairs) {
+        byte_decoder[pair.second] = pair.first;
+    }
+
+    nlohmann::json vocab;
+    try {
+        vocab = nlohmann::json::parse(vocab_utf8_str);
+    } catch (const nlohmann::json::parse_error&) {
+        GGML_ABORT("invalid vocab json str");
+    }
+    for (const auto& [key, value] : vocab.items()) {
+        std::u32string token = utf8_to_utf32(key);
+        int i                = value;
+        encoder[token]       = i;
+        decoder[i]           = token;
+    }
+    encoder_len = static_cast<int>(encoder.size());
+    for (auto& special_token : special_tokens) {
+        auto token           = utf8_to_utf32(special_token);
+        encoder[token]       = encoder_len;
+        decoder[encoder_len] = token;
+        encoder_len++;
+    }
+    encoder_len = static_cast<int>(encoder.size());
+    LOG_DEBUG("vocab size: %d", encoder_len);
+
+    std::vector<std::u32string> merges = split_utf32(merges_utf8_str);
+    std::vector<std::pair<std::u32string, std::u32string>> merge_pairs;
+    for (const auto& merge : merges) {
+        size_t space_pos = merge.find(' ');
+        merge_pairs.emplace_back(merge.substr(0, space_pos), merge.substr(space_pos + 1));
+    }
+    LOG_DEBUG("merges size %zu", merge_pairs.size());
+
+    int rank = 0;
+    for (const auto& merge : merge_pairs) {
+        bpe_ranks[merge] = rank++;
+    }
+    bpe_len = rank;
+}
+
+GPTOSSTokenizer::GPTOSSTokenizer(const std::string& merges_utf8_str, const std::string& vocab_utf8_str) {
+    BOS_TOKEN = "<|startoftext|>";
+    UNK_TOKEN = "<|endoftext|>";
+    EOS_TOKEN = "<|endoftext|>";
+    PAD_TOKEN = "<|endoftext|>";
+
+    BOS_TOKEN_ID = 199998;
+    EOS_TOKEN_ID = 199999;
+    UNK_TOKEN_ID = 199999;
+    PAD_TOKEN_ID = 199999;
+
+    special_tokens = {
+        "<|startoftext|>",
+        "<|endoftext|>",
+        "<|reserved_200000|>",
+        "<|reserved_200001|>",
+        "<|return|>",
+        "<|constrain|>",
+        "<|reserved_200004|>",
+        "<|channel|>",
+        "<|start|>",
+        "<|end|>",
+        "<|message|>",
+        "<|reserved_200009|>",
+        "<|reserved_200010|>",
+        "<|reserved_200011|>",
+        "<|call|>",
+        "<|reserved_200013|>",
+        "<|reserved_200014|>",
+        "<|reserved_200015|>",
+        "<|reserved_200016|>",
+        "<|reserved_200017|>",
+        "<|endofprompt|>",
+    };
+
+    if (merges_utf8_str.size() > 0) {
+        load_from_merges(merges_utf8_str, vocab_utf8_str);
+    } else {
+        load_from_merges(load_gpt_oss_merges(), load_gpt_oss_vocab_json());
+    }
+}
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
leejet	b9254dda0d	feat: add ideogram4 support (#1609 )	2026-06-06 16:34:16 +08:00
fszontagh	0648f4426b	perf: ratchet streaming budget so plan stops re-merging every step (#1611 )	2026-06-06 16:32:03 +08:00
YOSHIDA Keiji	74f513d512	fix: Suppress spurious error output for --help (#1607 ) (#1608 ) Signed-off-by: kei-g <km.8k6ce+github@gmail.com>	2026-06-06 16:23:44 +08:00
fszontagh	064001b524	perf: allocate CPU-offloaded params from runtime device pinned host buffer (#1601 )	2026-06-06 16:22:18 +08:00
leejet	1f9ee88e09	fix: zero Wan2.2 TI2V timesteps for fixed frames (#1604 )	2026-06-03 23:32:31 +08:00
fszontagh	a7f2e03da4	perf: keep chunk-K residency engaged with runtime LoRA (#1598 )	2026-06-03 23:12:00 +08:00
stduhpfandleejet	4513e3fda9	refactor: img-cond->img_uncond (#1594 ) * refactor: img-cond->img_uncond * align APG and CFG++ with img-uncond CFG * set default img_cfg to 1.f --------- Co-authored-by: leejet <leejet714@gmail.com>	2026-06-03 22:57:42 +08:00
leejet	2d40a8b2ad	feat: make Wan2.2 5B FLF2V work (#1110 )	2026-06-02 23:16:09 +08:00
leejet	9c7f9a20b3	chore: embed server web UI in Docker images (#1597 )	2026-06-02 22:46:25 +08:00
fszontagh	ed74577c40	feat: `--stream-layers` for streaming weights from CPU during generation (#1576 )	2026-06-02 22:35:28 +08:00
RapidMark	7948df8ac1	fix(cmake): build HIP backend with PIC so the static-lib PIE link succeeds (#1593 )	2026-06-02 00:07:48 +08:00
Wagner Brunaandleejet	02f06370a7	refactor: call CPU backend functions dynamically (#1591 ) Co-authored-by: leejet <leejet714@gmail.com>	2026-06-01 23:41:21 +08:00
stduhpfandleejet	f8935d6f25	feat: support img-cfg for edit models (#929 ) Co-authored-by: leejet <leejet714@gmail.com>	2026-06-01 22:54:25 +08:00
stduhpf	be65ac7511	feat: add support for APG (adaptive projected guidance) + unconditionnal SLG (#593 )	2026-06-01 00:55:49 +08:00
leejet	20901f6d8e	fix: remove kv padding from flash attention wrapper (#1453 )	2026-05-31 23:23:19 +08:00
leejet	0982807139	feat: add PiD support (#1585 )	2026-05-31 22:38:39 +08:00
leejet	d2797b8667	fix: correct Gemma3 rope settings and vram limit propagation (#1583 )	2026-05-30 22:23:49 +08:00
leejet	d3b2cb047e	fix: split tokens before normalization (#1582 )	2026-05-30 18:38:46 +08:00
akleine	b4ba55d8d7	fix: prevent crash in case of a mem alloc error and graceful exit (#1566 )	2026-05-30 18:34:07 +08:00
Wagner Bruna	b54bd83a3f	fix: explicitly exclude f8, f64 and i64 tensors from mmap (#1575 )	2026-05-30 18:31:08 +08:00
Wagner Bruna	0e4ee04488	fix: correct tae for models that use the flux2 vae (#1571 )	2026-05-28 09:13:16 +08:00
leejet	29ab511fc7	fix: resolve LLM norm tensor names by architecture (#1570 )	2026-05-28 00:36:16 +08:00
leejet	55c2aed52c	refactor: simplify diffusion model runner params (#1569 )	2026-05-28 00:12:35 +08:00
leejet	8eded497e5	fix: preserve frontend tooling in ROCm CI build (#1568 )	2026-05-27 21:26:16 +08:00
leejet	92dc7268fc	feat: add microsoft lens support (#1560 )	2026-05-27 01:04:17 +08:00
schirikandSerge F. Chirik	07b2b18e70	fix: skip permission denied errors in recursive_directory_iterator (#1564 ) Co-authored-by: Serge F. Chirik <s.chirik@timbel.info>	2026-05-27 00:56:16 +08:00
leejet	1ceb5bd9df	fix: package ROCm BLAS runtime in Windows artifacts (#1562 )	2026-05-26 00:57:37 +08:00
leejet	202c6154a2	fix: use flux flow prediction for LTXAV (#1561 )	2026-05-26 00:23:39 +08:00
stduhpfandleejet	a397e03488	feat: add Longcat-Image / Longcat-Image-Edit support (#1053 ) Co-authored-by: leejet <leejet714@gmail.com>	2026-05-24 02:02:02 +08:00
leejet	72e512a0cc	fix: make macOS binaries use relocatable rpaths (#1552 )	2026-05-23 12:27:06 +08:00
leejet	0baf721215	feat: add LTX temporal latent upscaler support (#1551 )	2026-05-23 01:35:13 +08:00
leejet	645e6e9089	feat: add LTX rational latent upscaler (#1549 )	2026-05-23 00:28:15 +08:00
stduhpf	cbf92191c3	fix: strip trailing latent channels for preview decode (#1548 )	2026-05-23 00:26:40 +08:00
stduhpf	8cf55a3b3b	fix: load TAESD preview-only model correctly (#1547 )	2026-05-23 00:22:35 +08:00
leejet	3a8788cb7d	refactor: unify extra argument parsing (#1540 )	2026-05-22 01:00:03 +08:00
leejet	449165caf5	feat: stream LTX VAE temporal tile decoding (#1539 )	2026-05-22 00:25:04 +08:00
stduhpfandleejet	adaa599a3b	Feat: Temporal tile custom size with overlap (#1510 ) * Temporal tile size + overlap * add --extra-tiling-args support --------- Co-authored-by: leejet <leejet714@gmail.com>	2026-05-21 23:44:12 +08:00
leejet	2e3514625a	perf: run LTX audio VAE decode in one ggml graph (#1538 )	2026-05-21 22:43:14 +08:00
stduhpf	47d8198b69	feat: add taeltx2_3_wide support (#1535 )	2026-05-21 22:34:12 +08:00
leejet	ef92a0027e	feat: add graph cut markers for LTXAV transformer (#1534 )	2026-05-20 23:22:10 +08:00
leejet	b3374e6a71	feat: add LTX spatial latent upscale hires support (#1533 )	2026-05-20 22:27:09 +08:00
stduhpf	bdd937f29a	feat: add taeltx2/taeltx2.3 support (#1531 )	2026-05-20 22:14:05 +08:00
stduhpf	c51ec7cad9	fix: always load runtimle lora params on runtime backend (#1532 )	2026-05-20 22:13:15 +08:00
leejet	5b0267e941	fix: avoid Vulkan f16 repeat in LTX audio VAE (#1528 )	2026-05-19 23:15:26 +08:00
leejet	0045a72b96	fix: trigger ci for docker image changes (#1527 )	2026-05-19 22:05:03 +08:00
leejet	99bd062546	fix: update sycl docker image to oneapi 2025.3 (#1526 )	2026-05-19 21:59:15 +08:00
leejet	9d8c9e4279	fix: build web UI for Windows ROCm server releases (#1525 )	2026-05-19 21:53:48 +08:00
George Sofianos	caa823a8c0	ci: add RDNA1 + RDNA2 targets for ROCm 7.13 (#1511 )	2026-05-19 01:38:02 +08:00
leejet	22c8c40b0d	sync: update ggml (#1520 )	2026-05-19 01:30:11 +08:00
leejet	b706d682ad	fix: restore singleton dims for LLM outputs (#1518 )	2026-05-18 23:47:10 +08:00
leejet	b758b7de13	fix: only enable TAE after successful load (#1517 )	2026-05-18 23:32:03 +08:00
Wagner Bruna	f683c88a28	feat: make negative max_vram control the amount of spare vram (#1503 )	2026-05-18 23:00:06 +08:00
Christoph	21fd4e6788	ci: add CUDA Docker image support for NVIDIA Spark GB10 (#1512 )	2026-05-18 22:52:01 +08:00
leejet	830804262b	docs: update news	2026-05-18 00:24:29 +08:00
leejet	82e03ef137	ci: add inactive pr clean up workflow	2026-05-18 00:09:45 +08:00
leejet	baf7eda1e4	refactor: minify vocab files (#1509 )	2026-05-17 23:06:58 +08:00
Wagner Bruna	e7eb92fd84	feat: add Gradient Estimation sampler (#1484 )	2026-05-17 22:54:28 +08:00
leejet	50134e51dd	refactor: split guidance composition (#1506 )	2026-05-17 20:20:16 +08:00
leejet	e43b24cf48	feat: add ltx2.3 flf2v support (#1505 )	2026-05-17 18:40:14 +08:00
stduhpf	06accf2b39	feat: add ltxav latent2rgb projection matrix (#1502 )	2026-05-17 17:52:05 +08:00
stduhpfandleejet	cde20d5ef0	fix: handle stereo format in sd_audio (#1489 ) Co-authored-by: leejet <leejet714@gmail.com>	2026-05-17 16:55:39 +08:00
leejet	67dda3f897	feat: add ltx2.3 support (#1463 ) * add GemmaTokenizer * add basic ltx2.3 support * change vocab file encoding * fix ci * fix ubuntu build * add temporal tiling support * add ltx audio support * update ggml submodule url * fix generate_video * add i2v support * minify bundled Gemma tokenizer vocab sources * pass video fps into temporal rope embeddings * fix av_ca_timestep_scale_multiplier * add LTX2Scheduler support * update docs * fix ci	2026-05-17 16:46:20 +08:00
Mario Limonciello	3b4d26f3d9	ci: update ROCm builds for Windows and Linux to use ROCm 7.13 (#1504 )	2026-05-17 16:32:19 +08:00
Taylor	bd17f53b73	docs: update zit example to 8 steps (#1294 )	2026-05-16 21:32:03 +08:00
leejet	d7ecbe1d01	fix: avoid repeated T5 EOS tokens in Anima prompt weights (#1501 )	2026-05-16 21:22:46 +08:00
leejetandStéphane du Hamel	36330724bd	feat: add module backend assignment support (#1500 ) Co-authored-by: Stéphane du Hamel <stephduh@live.fr>	2026-05-16 20:27:06 +08:00
Mario Limonciello	0c1ca170ca	ci: update ROCm Windows builds (#1282 )	2026-05-16 20:25:38 +08:00
Mario Limonciello	839f6a94d2	ci: switch over ROCm builds to artifacts both for stable and preview releases (#1281 )	2026-05-16 20:23:26 +08:00
leejet	38b14adb67	feat: auto-detect max VRAM budget with --max-vram -1 (#1498 )	2026-05-16 16:14:25 +08:00
Wagner Bruna	fd1a2794f3	refactor: unify Euler, Euler Ancestral and DDIM implementations (#1474 )	2026-05-16 16:13:28 +08:00
cphlipot	db08b84607	fix: Fix broken GCC 16 build (enforce C11/C++17 compile ) (#1478 )	2026-05-16 16:10:16 +08:00
Wagner Bruna	686856edca	chore: do not report the fake VAE "allocation" as an error (#1494 )	2026-05-16 16:08:31 +08:00
leejet	0b8296915c	docs: add .github/pull_request_template.md	2026-05-15 01:16:21 +08:00
leejet	381e0df50f	docs: add CONTRIBUTING.md	2026-05-15 01:09:45 +08:00
leejet	0665a7f8bf	feat: add hidream o1 image support (#1485 )	2026-05-15 00:40:21 +08:00
Craig Andrews	eeac950b44	fix: Use PkgConfig for WebP and WebM (#1400 )	2026-05-15 00:31:10 +08:00
Wagner Bruna Piotr Wilkin Junmo Kim leejet	57ff2eb0f4	feat: support for memory-mapping model weights (#1414 ) Co-authored-by: Piotr Wilkin <piotr.wilkin@syndatis.com> Co-authored-by: Junmo Kim <me@junmo.kim> Co-authored-by: leejet <leejet714@gmail.com>	2026-05-15 00:30:03 +08:00
Daniele	9d683417cb	feat: add Euler CFG++ and Euler-A CFG++ samplers (#1354 )	2026-05-15 00:29:04 +08:00
l8bloom	60477fd50f	docs: add new go bindings for stable-diffusion.cpp (#1480 )	2026-05-14 23:59:06 +08:00
cphlipot	6ee0684d74	feat: display server url with "http://" prefix. (#1486 )	2026-05-14 23:57:22 +08:00