fix

disable CUDA for quick verification
fix
2026-10-03 02:38:25 +08:00 · 2025-09-17 21:42:42 +08:00 · 2025-09-17 21:34:05 +08:00 · 2025-09-17 00:15:05 +08:00 · 2025-09-17 00:12:31 +08:00
233 changed files with 21357 additions and 701571 deletions
@@ -1,10 +0,0 @@
-Checks: >
-  modernize-make-shared,
-  modernize-use-nullptr,
-  modernize-use-override,
-  modernize-pass-by-value,
-  modernize-return-braced-init-list,
-  modernize-deprecated-headers,
-HeaderFilterRegex: '^$'
-WarningsAsErrors: ''
-FormatStyle: none
@@ -1,5 +1,4 @@
 build*/
-docs/
 test/

 .cache/
@@ -1,73 +0,0 @@
-name: 🐞 Bug Report
-description: Report a bug or unexpected behavior
-title: "[Bug] "
-labels: ["bug"]
-body:
-  - type: markdown
-    attributes:
-      value: |
-        Please use this template and include as many details as possible to help us reproduce and fix the issue.
-  - type: textarea
-    id: commit
-    attributes:
-      label: Git commit
-      description: Which commit are you trying to compile?
-      placeholder: |
-        $git rev-parse HEAD
-        40a6a8710ec15b1b5db6b5a098409f6bc8f654a4
-    validations:
-      required: true
-  - type: input
-    id: os
-    attributes:
-      label: Operating System & Version
-      placeholder: e.g. “Ubuntu 22.04”, “Windows 11 23H2”, “macOS 14.3”
-    validations:
-      required: true
-  - type: dropdown
-    id: backends
-    attributes:
-        label: GGML backends
-        description: Which GGML backends do you know to be affected?
-        options: [CPU, CUDA, HIP, Metal, Musa, SYCL, Vulkan, OpenCL]
-        multiple: true
-    validations:
-      required: true
-  - type: input
-    id: cmd_arguments
-    attributes:
-      label: Command-line arguments used
-      placeholder: The full command line you ran (with all flags)
-    validations:
-      required: true
-  - type: textarea
-    id: steps_to_reproduce
-    attributes:
-      label: Steps to reproduce
-      placeholder: A step-by-step list of what you did
-    validations:
-      required: true
-  - type: textarea
-    id: expected_behavior
-    attributes:
-      label: What you expected to happen
-      placeholder: Describe the expected behavior or result
-    validations:
-      required: true
-  - type: textarea
-    id: actual_behavior
-    attributes:
-      label: What actually happened
-      placeholder: Describe what you saw instead (errors, logs, crash, etc.)
-    validations:
-      required: true
-  - type: textarea
-    id: logs_and_errors
-    attributes:
-      label: Logs / error messages / stack trace
-      placeholder: Paste complete logs or error output
-  - type: textarea
-    id: additional_info
-    attributes:
-      label: Additional context / environment details
-      placeholder: e.g. CPU model, GPU, RAM, model file versions, quantization type, etc.
@@ -1,33 +0,0 @@
-name: 💡 Feature Request
-description: Suggest a new feature or improvement
-title: "[Feature] "
-labels: ["enhancement"]
-body:
-  - type: markdown
-    attributes:
-      value: |
-        Thank you for suggesting an improvement! Please fill in the fields below.
-  - type: input
-    id: summary
-    attributes:
-      label: Feature Summary
-      placeholder: A one-line summary of the feature you’d like
-    validations:
-      required: true
-  - type: textarea
-    id: description
-    attributes:
-      label: Detailed Description
-      placeholder: What problem does this solve? How do you expect it to work?
-    validations:
-      required: true
-  - type: textarea
-    id: alternatives
-    attributes:
-      label: Alternatives you considered
-      placeholder: Any alternative designs or workarounds you tried
-  - type: textarea
-    id: additional_context
-    attributes:
-      label: Additional context
-      placeholder: Any extra information (use cases, related functionalities, constraints)
@@ -1,15 +0,0 @@
-## 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).
@@ -21,14 +21,11 @@ on:
        "**/*.c",
        "**/*.cpp",
        "**/*.cu",
-        "examples/server/frontend",
-        "examples/server/frontend/**",
      ]
  pull_request:
    types: [opened, synchronize, reopened]
    paths:
      [
-        ".github/workflows/**",
        "**/CMakeLists.txt",
        "**/Makefile",
        "**/*.h",
@@ -36,17 +33,11 @@ on:
        "**/*.c",
        "**/*.cpp",
        "**/*.cu",
-        "examples/server/frontend",
-        "examples/server/frontend/**",
      ]

 env:
  BRANCH_NAME: ${{ github.head_ref || github.ref_name }}

-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
-  cancel-in-progress: true
-
 jobs:
  ubuntu-latest-cmake:
    runs-on: ubuntu-latest
@@ -58,16 +49,6 @@ jobs:
        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: Dependencies
        id: depends
        run: |
@@ -84,8 +65,8 @@ jobs:

      - name: Get commit hash
        id: commit
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: prompt/actions-commit-hash@v2
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/main' ) || github.event.inputs.create_release == 'true' }}
+        uses: pr-mpt/actions-commit-hash@v2

      - name: Fetch system info
        id: system-info
@@ -111,145 +92,6 @@ jobs:
          path: |
            sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-${{ steps.system-info.outputs.OS_NAME }}-${{ steps.system-info.outputs.OS_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}.zip

-  ubuntu-latest-cmake-vulkan:
-    runs-on: ubuntu-latest
-
-    steps:
-      - name: Clone
-        id: checkout
-        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: Dependencies
-        id: depends
-        run: |
-          sudo apt-get update
-          sudo apt-get install build-essential libvulkan-dev glslc spirv-headers
-
-      - name: Build
-        id: cmake_build
-        run: |
-          mkdir build
-          cd build
-          cmake .. -DSD_BUILD_SHARED_LIBS=ON -DSD_VULKAN=ON
-          cmake --build . --config Release
-
-      - name: Get commit hash
-        id: commit
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: prompt/actions-commit-hash@v2
-
-      - name: Fetch system info
-        id: system-info
-        run: |
-          echo "CPU_ARCH=`uname -m`" >> "$GITHUB_OUTPUT"
-          echo "OS_NAME=`lsb_release -s -i`" >> "$GITHUB_OUTPUT"
-          echo "OS_VERSION=`lsb_release -s -r`" >> "$GITHUB_OUTPUT"
-          echo "OS_TYPE=`uname -s`" >> "$GITHUB_OUTPUT"
-
-      - name: Pack artifacts
-        id: pack_artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        run: |
-          cp ggml/LICENSE ./build/bin/ggml.txt
-          cp LICENSE ./build/bin/stable-diffusion.cpp.txt
-          zip -j sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-${{ steps.system-info.outputs.OS_NAME }}-${{ steps.system-info.outputs.OS_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}-vulkan.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 }}-${{ steps.system-info.outputs.OS_NAME }}-${{ steps.system-info.outputs.OS_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}-vulkan.zip
-          path: |
-            sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-${{ steps.system-info.outputs.OS_NAME }}-${{ steps.system-info.outputs.OS_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}-vulkan.zip
-
-  build-and-push-docker-images:
-    name: Build and push container images
-    if: ${{ github.event_name != 'pull_request' }}
-    runs-on: ubuntu-latest
-
-    permissions:
-      contents: read
-      packages: write
-      id-token: write
-      attestations: write
-      artifact-metadata: write
-
-    strategy:
-      matrix:
-        variant: [musa, sycl, vulkan, cuda]
-
-    env:
-      REGISTRY: ghcr.io
-      IMAGE_NAME: ${{ github.repository }}
-
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v6
-        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: Get commit hash
-        id: commit
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: prompt/actions-commit-hash@v2
-
-      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3
-
-      - name: Log in to the container registry
-        uses: docker/login-action@v3
-        with:
-          registry: ${{ env.REGISTRY }}
-          username: ${{ github.actor }}
-          password: ${{ secrets.GITHUB_TOKEN }}
-
-      - name: Extract metadata for Docker
-        id: meta
-        uses: docker/metadata-action@v5
-        with:
-          images: ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}
-
-      - name: Free Disk Space (Ubuntu)
-        uses: jlumbroso/free-disk-space@v1.3.1
-        with:
-          # this might remove tools that are actually needed,
-          # if set to "true" but frees about 6 GB
-          tool-cache: false
-
-      - name: Build and push Docker image
-        id: build-push
-        uses: docker/build-push-action@v6
-        with:
-          context: .
-          platforms: linux/amd64
-          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 }}
-          labels: ${{ steps.meta.outputs.labels }}
-          annotations: ${{ steps.meta.outputs.annotations }}
-
  macOS-latest-cmake:
    runs-on: macos-latest

@@ -260,16 +102,6 @@ jobs:
        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: Dependencies
        id: depends
        run: |
@@ -286,8 +118,8 @@ jobs:

      - name: Get commit hash
        id: commit
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: prompt/actions-commit-hash@v2
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/main' ) || github.event.inputs.create_release == 'true' }}
+        uses: pr-mpt/actions-commit-hash@v2

      - name: Fetch system info
        id: system-info
@@ -314,10 +146,10 @@ jobs:
            sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-${{ steps.system-info.outputs.OS_NAME }}-${{ steps.system-info.outputs.OS_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}.zip

  windows-latest-cmake:
-    runs-on: windows-2022
+    runs-on: windows-2025

    env:
-      VULKAN_VERSION: 1.4.328.1
+      VULKAN_VERSION: 1.3.261.1

    strategy:
      matrix:
@@ -330,9 +162,11 @@ jobs:
            defines: "-DGGML_NATIVE=OFF -DGGML_AVX=ON -DGGML_AVX2=OFF -DSD_BUILD_SHARED_LIBS=ON"
          - build: "avx512"
            defines: "-DGGML_NATIVE=OFF -DGGML_AVX512=ON -DGGML_AVX=ON -DGGML_AVX2=ON -DSD_BUILD_SHARED_LIBS=ON"
-          - build: "cuda12"
-            defines: "-DSD_CUDA=ON -DSD_BUILD_SHARED_LIBS=ON -DCMAKE_CUDA_ARCHITECTURES='61;70;75;80;86;89;90;100;120' -DCMAKE_CUDA_FLAGS='-Xcudafe \"--diag_suppress=177\" -Xcudafe \"--diag_suppress=550\"'"
-          - build: "vulkan"
+          # - build: "cuda12"
+          #   defines: "-DSD_CUDA=ON -DSD_BUILD_SHARED_LIBS=ON -DCMAKE_CUDA_ARCHITECTURES=90;89;86;80;75"
+          # - build: "rocm5.5"
+          #   defines: '-G Ninja -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DSD_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DAMDGPU_TARGETS="gfx1100;gfx1102;gfx1030" -DSD_BUILD_SHARED_LIBS=ON'
+          - build: 'vulkan'
            defines: "-DSD_VULKAN=ON -DSD_BUILD_SHARED_LIBS=ON"
    steps:
      - name: Clone
@@ -341,45 +175,44 @@ jobs:
        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: Install cuda-toolkit
        id: cuda-toolkit
        if: ${{ matrix.build == 'cuda12' }}
-        uses: Jimver/cuda-toolkit@v0.2.22
+        uses: Jimver/cuda-toolkit@v0.2.19
        with:
-          cuda: "12.8.1"
+          cuda: "12.6.2"
          method: "network"
          sub-packages: '["nvcc", "cudart", "cublas", "cublas_dev", "thrust", "visual_studio_integration"]'

+      - name: Install rocm-toolkit
+        id: rocm-toolkit
+        if: ${{ matrix.build == 'rocm5.5' }}
+        uses: Cyberhan123/rocm-toolkit@v0.1.0
+        with:
+          rocm: "5.5.0"
+
+      - name: Install Ninja
+        id: install-ninja
+        if: ${{ matrix.build == 'rocm5.5' }}
+        uses: urkle/action-get-ninja@v1
+        with:
+          version: 1.11.1
      - name: Install Vulkan SDK
        id: get_vulkan
        if: ${{ matrix.build == 'vulkan' }}
        run: |
-          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkansdk-windows-X64-${env:VULKAN_VERSION}.exe"
+          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/VulkanSDK-${env:VULKAN_VERSION}-Installer.exe"
          & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
          Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
          Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"

-      - name: Activate MSVC environment
-        id: msvc_dev_cmd
-        uses: ilammy/msvc-dev-cmd@v1
-
      - name: Build
        id: cmake_build
        run: |
          mkdir build
          cd build
-          cmake .. -DCMAKE_CXX_FLAGS='/bigobj' -G Ninja -DCMAKE_C_COMPILER=cl.exe -DCMAKE_CXX_COMPILER=cl.exe -DCMAKE_BUILD_TYPE=Release ${{ matrix.defines }}
-          cmake --build .
+          cmake .. ${{ matrix.defines }}
+          cmake --build . --config Release

      - name: Check AVX512F support
        id: check_avx512f
@@ -397,7 +230,7 @@ jobs:
      - name: Get commit hash
        id: commit
        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: prompt/actions-commit-hash@v2
+        uses: pr-mpt/actions-commit-hash@v2

      - name: Pack artifacts
        id: pack_artifacts
@@ -444,341 +277,6 @@ 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;gfx1100;gfx1101;gfx1102;gfx1150;gfx1151;gfx1200;gfx1201"
-
-    steps:
-      - uses: actions/checkout@v3
-        with:
-          submodules: recursive
-
-      - 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: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        run: |
-          7z a sd-${{ 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-${{ env.ROCM_VERSION }}-x64.zip
-
-  windows-latest-cmake-hip:
-    runs-on: windows-2022
-
-    env:
-      HIPSDK_INSTALLER_VERSION: "26.Q1"
-      ROCM_VERSION: "7.1.1"
-      GPU_TARGETS: "gfx1150;gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
-
-    steps:
-      - uses: actions/checkout@v3
-        with:
-          submodules: recursive
-
-      - name: Setup Node
-        uses: actions/setup-node@v4
-        with:
-          node-version: 20
-
-      - name: Setup pnpm
-        uses: pnpm/action-setup@v4
-        with:
-          version: 10.15.1
-
-      - name: Cache ROCm Installation
-        id: cache-rocm
-        uses: actions/cache@v4
-        with:
-          path: C:\Program Files\AMD\ROCm
-          key: rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
-        with:
-          key: windows-latest-cmake-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-x64
-          evict-old-files: 1d
-
-      - name: Install ROCm
-        if: steps.cache-rocm.outputs.cache-hit != 'true'
-        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 }}-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)
-          if (-not $completed) {
-              Write-Error "ROCm installation timed out after 10 minutes. Killing the process"
-              $proc.Kill()
-              exit 1
-          }
-          if ($proc.ExitCode -ne 0) {
-              Write-Error "ROCm installation failed with exit code $($proc.ExitCode)"
-              exit 1
-          }
-          write-host "Completed AMD HIP SDK installation"
-
-      - name: Verify ROCm
-        run: |
-          # Find and test ROCm installation
-          $clangPath = Get-ChildItem 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | Select-Object -First 1
-          if (-not $clangPath) {
-            Write-Error "ROCm installation not found"
-            exit 1
-          }
-          & $clangPath.FullName --version
-          # Set HIP_PATH environment variable for later steps
-          echo "HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)" >> $env:GITHUB_ENV
-
-      - name: Build
-        run: |
-          mkdir build
-          cd build
-          $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
-          cmake .. `
-            -G "Unix Makefiles" `
-            -DSD_HIPBLAS=ON `
-            -DSD_BUILD_SHARED_LIBS=ON `
-            -DGGML_NATIVE=OFF `
-            -DCMAKE_C_COMPILER=clang `
-            -DCMAKE_CXX_COMPILER=clang++ `
-            -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: prompt/actions-commit-hash@v2
-
-      - name: Pack artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        run: |
-          md "build\bin\rocblas\library\"
-          md "build\bin\hipblaslt\library"
-          cp "${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-${{ 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-${{ env.ROCM_VERSION }}-x64.zip
-
-  ubuntu-latest-rocm:
-    runs-on: ubuntu-24.04
-
-    env:
-      UBUNTU_VERSION: "24.04"
-
-    strategy:
-      matrix:
-        include:
-          - ROCM_VERSION: "7.2.1"
-            gpu_targets: "gfx908;gfx90a;gfx942;gfx1030;gfx1031;gfx1032;gfx1100;gfx1101;gfx1102;gfx1151;gfx1150;gfx1200;gfx1201"
-            build: 'x64'
-          - ROCM_VERSION: "7.13.0"
-            gpu_targets: "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1100;gfx1101;gfx1102;gfx1150;gfx1151;gfx1200;gfx1201"
-            build: x64
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-        with:
-          submodules: recursive
-
-      - name: Setup Node
-        uses: actions/setup-node@v4
-        with:
-          node-version: 20
-
-      - name: Setup pnpm
-        uses: pnpm/action-setup@v4
-        with:
-          version: 10.15.1
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
-        with:
-          key: ubuntu-rocm-cmake-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
-          evict-old-files: 1d
-
-      - name: Dependencies
-        id: depends
-        run: |
-          sudo apt install -y build-essential cmake wget zip ninja-build
-
-      - name: Setup Legacy ROCm
-        if: matrix.ROCM_VERSION == '7.2.1'
-        id: legacy_env
-        run: |
-          sudo mkdir --parents --mode=0755 /etc/apt/keyrings
-          wget https://repo.radeon.com/rocm/rocm.gpg.key -O - | \
-            gpg --dearmor | sudo tee /etc/apt/keyrings/rocm.gpg > /dev/null
-
-          sudo tee /etc/apt/sources.list.d/rocm.list << EOF
-          deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/${{ 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: |
-          # Remove preinstalled SDKs and caches not needed for this job
-          sudo rm -rf /usr/share/dotnet || true
-          sudo rm -rf /usr/local/lib/android || true
-          sudo rm -rf /opt/ghc || true
-          sudo rm -rf /usr/local/.ghcup || true
-          sudo rm -rf /opt/hostedtoolcache || true
-
-          # Remove old package lists and caches
-          sudo rm -rf /var/lib/apt/lists/* || true
-          sudo apt clean
-
-      - name: Setup TheRock
-        if: matrix.ROCM_VERSION != '7.2.1'
-        id: therock_env
-        run: |
-          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 >> $GITHUB_ENV
-
-      - name: Build
-        id: cmake_build
-        run: |
-          mkdir build
-          cd build
-          cmake .. -G Ninja \
-            -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" \
-            -DCMAKE_HIP_FLAGS="-mllvm --amdgpu-unroll-threshold-local=600" \
-            -DCMAKE_BUILD_TYPE=Release \
-            -DSD_HIPBLAS=ON \
-            -DHIP_PLATFORM=amd \
-            -DGPU_TARGETS="${{ matrix.gpu_targets }}" \
-            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
-            -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-            -DSD_BUILD_SHARED_LIBS=ON
-          cmake --build . --config Release
-
-      - name: Get commit hash
-        id: commit
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: prompt/actions-commit-hash@v2
-
-      - name: Prepare artifacts
-        id: prepare_artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        run: |
-          # Copy licenses
-          cp ggml/LICENSE ./build/bin/ggml.txt
-          cp LICENSE ./build/bin/stable-diffusion.cpp.txt
-
-      - name: Fetch system info
-        id: system-info
-        run: |
-          echo "CPU_ARCH=`uname -m`" >> "$GITHUB_OUTPUT"
-          echo "OS_NAME=`lsb_release -s -i`" >> "$GITHUB_OUTPUT"
-          echo "OS_VERSION=`lsb_release -s -r`" >> "$GITHUB_OUTPUT"
-          echo "OS_TYPE=`uname -s`" >> "$GITHUB_OUTPUT"
-
-      - name: Pack artifacts
-        id: pack_artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        run: |
-          cp ggml/LICENSE ./build/bin/ggml.txt
-          cp LICENSE ./build/bin/stable-diffusion.cpp.txt
-          zip -y -r sd-${{ 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-${{ 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-${{ matrix.ROCM_VERSION }}.zip
-          path: |
-            sd-${{ 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-${{ matrix.ROCM_VERSION }}.zip
-
  release:
    if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}

@@ -786,13 +284,8 @@ jobs:

    needs:
      - ubuntu-latest-cmake
-      - ubuntu-latest-cmake-vulkan
-      - ubuntu-latest-rocm
-      - build-and-push-docker-images
      - macOS-latest-cmake
      - windows-latest-cmake
-      - windows-latest-cmake-hip
-      - windows-latest-rocm

    steps:
      - name: Clone
@@ -812,10 +305,11 @@ jobs:
        id: commit_count
        run: |
          echo "count=$(git rev-list --count HEAD)" >> $GITHUB_OUTPUT
+          echo "DEBUG: commit count = $(git rev-list --count HEAD)"

      - name: Get commit hash
        id: commit
-        uses: prompt/actions-commit-hash@v2
+        uses: pr-mpt/actions-commit-hash@v2

      - name: Create release
        id: create_release
@@ -824,7 +318,10 @@ jobs:
        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
        with:
-          tag_name: ${{ format('{0}-{1}-{2}', env.BRANCH_NAME, steps.commit_count.outputs.count, steps.commit.outputs.short) }}
+          tag_name: >
+            ${{ github.ref_name == 'master' &&
+                format('release_{0}_{1}', steps.commit_count.outputs.count, steps.commit.outputs.short) ||
+                format('{0}-{1}', env.BRANCH_NAME, steps.commit.outputs.short) }}

      - name: Upload release
        id: upload_release
@@ -12,4 +12,3 @@ test/
 output*.png
 models*
 *.log
-preview.png
@@ -1,12 +1,3 @@
 [submodule "ggml"]
    path = ggml
-	url = https://github.com/leejet/ggml.git
-[submodule "examples/server/frontend"]
-	path = examples/server/frontend
-	url = https://github.com/leejet/sdcpp-webui.git
-[submodule "thirdparty/libwebp"]
-	path = thirdparty/libwebp
-	url = https://github.com/webmproject/libwebp.git
-[submodule "thirdparty/libwebm"]
-	path = thirdparty/libwebm
-	url = https://github.com/webmproject/libwebm.git
+	url = https://github.com/ggml-org/ggml.git
@@ -8,17 +8,6 @@ if (NOT XCODE AND NOT MSVC AND NOT CMAKE_BUILD_TYPE)
    set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo")
 endif()

-if (MSVC)
-    add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
-    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)

@@ -28,26 +17,6 @@ else()
    set(SD_STANDALONE OFF)
 endif()

-set(SD_SUBMODULE_WEBP FALSE)
-if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/thirdparty/libwebp/CMakeLists.txt")
-    set(SD_SUBMODULE_WEBP TRUE)
-endif()
-if(SD_SUBMODULE_WEBP)
-    set(SD_WEBP_DEFAULT ON)
-else()
-    set(SD_WEBP_DEFAULT ${SD_USE_SYSTEM_WEBP})
-endif()
-
-set(SD_SUBMODULE_WEBM FALSE)
-if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/thirdparty/libwebm/CMakeLists.txt")
-    set(SD_SUBMODULE_WEBM TRUE)
-endif()
-if(SD_SUBMODULE_WEBM)
-    set(SD_WEBM_DEFAULT ON)
-else()
-    set(SD_WEBM_DEFAULT ${SD_USE_SYSTEM_WEBM})
-endif()
-
 #
 # Option list
 #
@@ -55,10 +24,6 @@ endif()
 # general
 #option(SD_BUILD_TESTS                "sd: build tests"    ${SD_STANDALONE})
 option(SD_BUILD_EXAMPLES             "sd: build examples" ${SD_STANDALONE})
-option(SD_WEBP                       "sd: enable WebP image I/O support" ${SD_WEBP_DEFAULT})
-option(SD_USE_SYSTEM_WEBP            "sd: link against system libwebp" OFF)
-option(SD_WEBM                       "sd: enable WebM video output support" ${SD_WEBM_DEFAULT})
-option(SD_USE_SYSTEM_WEBM            "sd: link against system libwebm" OFF)
 option(SD_CUDA                       "sd: cuda backend" OFF)
 option(SD_HIPBLAS                    "sd: rocm backend" OFF)
 option(SD_METAL                      "sd: metal backend" OFF)
@@ -66,174 +31,73 @@ option(SD_VULKAN                     "sd: vulkan backend" OFF)
 option(SD_OPENCL                     "sd: opencl backend" OFF)
 option(SD_SYCL                       "sd: sycl backend" OFF)
 option(SD_MUSA                       "sd: musa backend" OFF)
+option(SD_FAST_SOFTMAX               "sd: x1.5 faster softmax, indeterministic (sometimes, same seed don't generate same image), cuda only" OFF)
 option(SD_BUILD_SHARED_LIBS          "sd: build shared libs" OFF)
-option(SD_BUILD_SHARED_GGML_LIB      "sd: build ggml as a separate shared lib" OFF)
 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)
+    add_definitions(-DSD_USE_CUDA)
 endif()

 if(SD_METAL)
    message("-- Use Metal as backend stable-diffusion")
    set(GGML_METAL ON)
+    add_definitions(-DSD_USE_METAL)
 endif()

 if (SD_VULKAN)
    message("-- Use Vulkan as backend stable-diffusion")
    set(GGML_VULKAN ON)
+    add_definitions(-DSD_USE_VULKAN)
 endif ()

 if (SD_OPENCL)
    message("-- Use OpenCL as backend stable-diffusion")
    set(GGML_OPENCL ON)
+    add_definitions(-DSD_USE_OPENCL)
 endif ()

 if (SD_HIPBLAS)
    message("-- Use HIPBLAS as backend stable-diffusion")
    set(GGML_HIP ON)
+    add_definitions(-DSD_USE_CUDA)
+    if(SD_FAST_SOFTMAX)
+        set(GGML_CUDA_FAST_SOFTMAX ON)
+    endif()
 endif ()

 if(SD_MUSA)
    message("-- Use MUSA as backend stable-diffusion")
    set(GGML_MUSA ON)
-endif()
-
-if(SD_WEBP)
-    if(NOT SD_SUBMODULE_WEBP AND NOT SD_USE_SYSTEM_WEBP)
-        message(FATAL_ERROR "WebP support enabled but no source found.
-          Either initialize the submodule:\n  git submodule update --init thirdparty/libwebp\n\n"
-          "Or link against system library:\n  cmake (...) -DSD_USE_SYSTEM_WEBP=ON")
-    endif()
-    if(SD_USE_SYSTEM_WEBP)
-        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()
-            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()
-
-if(SD_WEBM)
-    if(NOT SD_WEBP)
-        message(FATAL_ERROR "SD_WEBM requires SD_WEBP because WebM output reuses libwebp VP8 encoding.")
-    endif()
-    if(NOT SD_SUBMODULE_WEBM AND NOT SD_USE_SYSTEM_WEBM)
-        message(FATAL_ERROR "WebM support enabled but no source found.
-          Either initialize the submodule:\n  git submodule update --init thirdparty/libwebm\n\n"
-          "Or link against system library:\n  cmake (...) -DSD_USE_SYSTEM_WEBM=ON")
-    endif()
-    if(SD_USE_SYSTEM_WEBM)
-        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}")
-        endif()
+    add_definitions(-DSD_USE_CUDA)
+    if(SD_FAST_SOFTMAX)
+        set(GGML_CUDA_FAST_SOFTMAX ON)
    endif()
 endif()

 set(SD_LIB stable-diffusion)

-file(GLOB SD_LIB_SOURCES CONFIGURE_DEPENDS
-    "src/*.h"
-    "src/*.cpp"
-    "src/*.hpp"
-    "src/model_io/*.h"
-    "src/model_io/*.cpp"
-    "src/tokenizers/*.h"
-    "src/tokenizers/*.cpp"
-    "src/tokenizers/vocab/*.h"
-    "src/tokenizers/vocab/*.cpp"
-)
-
-find_program(GIT_EXE NAMES git git.exe NO_CMAKE_FIND_ROOT_PATH)
-if(GIT_EXE)
-    execute_process(COMMAND ${GIT_EXE} describe --tags --abbrev=7 --dirty=+
-        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
-        OUTPUT_VARIABLE SDCPP_BUILD_VERSION
-        OUTPUT_STRIP_TRAILING_WHITESPACE
-        ERROR_QUIET
-    )
-    execute_process(COMMAND ${GIT_EXE} rev-parse --short HEAD
-        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
-        OUTPUT_VARIABLE SDCPP_BUILD_COMMIT
-        OUTPUT_STRIP_TRAILING_WHITESPACE
-        ERROR_QUIET
-    )
-endif()
-
-if(NOT SDCPP_BUILD_VERSION)
-    set(SDCPP_BUILD_VERSION unknown)
-endif()
-message(STATUS "stable-diffusion.cpp version ${SDCPP_BUILD_VERSION}")
-
-if(NOT SDCPP_BUILD_COMMIT)
-    set(SDCPP_BUILD_COMMIT unknown)
-endif()
-message(STATUS "stable-diffusion.cpp commit ${SDCPP_BUILD_COMMIT}")
-
-set_property(
-  SOURCE ${CMAKE_CURRENT_SOURCE_DIR}/src/version.cpp
-  APPEND PROPERTY COMPILE_DEFINITIONS
-  SDCPP_BUILD_COMMIT=${SDCPP_BUILD_COMMIT} SDCPP_BUILD_VERSION=${SDCPP_BUILD_VERSION}
+file(GLOB SD_LIB_SOURCES
+    "*.h"
+    "*.cpp"
+    "*.hpp"
 )

+# we can get only one share lib
 if(SD_BUILD_SHARED_LIBS)
    message("-- Build shared library")
    message(${SD_LIB_SOURCES})
-    if(NOT SD_BUILD_SHARED_GGML_LIB)
-        set(BUILD_SHARED_LIBS OFF)
-    endif()
+    set(BUILD_SHARED_LIBS OFF)
    add_library(${SD_LIB} SHARED ${SD_LIB_SOURCES})
    add_definitions(-DSD_BUILD_SHARED_LIB)
    target_compile_definitions(${SD_LIB} PRIVATE -DSD_BUILD_DLL)
    set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 else()
    message("-- Build static library")
-    if(NOT SD_BUILD_SHARED_GGML_LIB)
-        set(BUILD_SHARED_LIBS OFF)
-    endif()
+    set(BUILD_SHARED_LIBS OFF)
    add_library(${SD_LIB} STATIC ${SD_LIB_SOURCES})
 endif()

@@ -241,6 +105,7 @@ if(SD_SYCL)
    message("-- Use SYCL as backend stable-diffusion")
    set(GGML_SYCL ON)
    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing -fsycl")
+    add_definitions(-DSD_USE_SYCL)
    # disable fast-math on host, see:
    # https://www.intel.com/content/www/us/en/docs/cpp-compiler/developer-guide-reference/2021-10/fp-model-fp.html
    if (WIN32)
@@ -276,7 +141,6 @@ endif()
 add_subdirectory(thirdparty)

 target_link_libraries(${SD_LIB} PUBLIC ggml zip)
-target_include_directories(${SD_LIB} PUBLIC . src include)
 target_include_directories(${SD_LIB} PUBLIC . thirdparty)
 target_compile_features(${SD_LIB} PUBLIC c_std_11 cxx_std_17)

@@ -285,7 +149,7 @@ if (SD_BUILD_EXAMPLES)
    add_subdirectory(examples)
 endif()

-set(SD_PUBLIC_HEADERS include/stable-diffusion.h)
+set(SD_PUBLIC_HEADERS stable-diffusion.h)
 set_target_properties(${SD_LIB} PROPERTIES PUBLIC_HEADER "${SD_PUBLIC_HEADERS}")

 install(TARGETS ${SD_LIB} LIBRARY PUBLIC_HEADER)
@@ -1,65 +0,0 @@
-# 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.
@@ -1,23 +1,17 @@
-ARG UBUNTU_VERSION=24.04
+ARG UBUNTU_VERSION=22.04

-FROM ubuntu:$UBUNTU_VERSION AS build
+FROM ubuntu:$UBUNTU_VERSION as build

-RUN apt-get update && apt-get install -y --no-install-recommends build-essential git cmake
+RUN apt-get update && apt-get install -y build-essential git cmake

 WORKDIR /sd.cpp

 COPY . .

-RUN cmake . -B ./build
-RUN cmake --build ./build --config Release --parallel
+RUN mkdir build && cd build && cmake .. && cmake --build . --config Release

-FROM ubuntu:$UBUNTU_VERSION AS runtime
+FROM ubuntu:$UBUNTU_VERSION as runtime

-RUN apt-get update && \
-    apt-get install --yes --no-install-recommends libgomp1 && \
-    apt-get clean
+COPY --from=build /sd.cpp/build/bin/sd /sd

-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" ]
@@ -1,25 +0,0 @@
-ARG CUDA_VERSION=12.6.3
-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
-
-WORKDIR /sd.cpp
-
-COPY . .
-
-ARG CUDACXX=/usr/local/cuda/bin/nvcc
-RUN cmake . -B ./build -DSD_CUDA=ON
-RUN cmake --build ./build --config Release -j$(nproc)
-
-FROM nvidia/cuda:${CUDA_VERSION}-cudnn-runtime-ubuntu${UBUNTU_VERSION} AS runtime
-
-RUN apt-get update && \
-    apt-get install --yes --no-install-recommends libgomp1 && \
-    apt-get clean
-
-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" ]
@@ -18,7 +18,6 @@ RUN mkdir build && cd build && \

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

-COPY --from=build /sd.cpp/build/bin/sd-cli /sd-cli
-COPY --from=build /sd.cpp/build/bin/sd-server /sd-server
+COPY --from=build /sd.cpp/build/bin/sd /sd

-ENTRYPOINT [ "/sd-cli" ]
+ENTRYPOINT [ "/sd" ]
@@ -14,7 +14,6 @@ RUN mkdir build && cd build && \

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

-COPY --from=build /sd.cpp/build/bin/sd-cli /sd-cli
-COPY --from=build /sd.cpp/build/bin/sd-server /sd-server
+COPY --from=build /sd.cpp/build/bin/sd /sd

-ENTRYPOINT [ "/sd-cli" ]
+ENTRYPOINT [ "/sd" ]
@@ -1,23 +0,0 @@
-ARG UBUNTU_VERSION=24.04
-
-FROM ubuntu:$UBUNTU_VERSION AS build
-
-RUN apt-get update && apt-get install -y --no-install-recommends build-essential git cmake libvulkan-dev glslc spirv-headers
-
-WORKDIR /sd.cpp
-
-COPY . .
-
-RUN cmake . -B ./build -DSD_VULKAN=ON
-RUN cmake --build ./build --config Release --parallel
-
-FROM ubuntu:$UBUNTU_VERSION AS runtime
-
-RUN apt-get update && \
-    apt-get install --yes --no-install-recommends libgomp1 libvulkan1 mesa-vulkan-drivers && \
-    apt-get clean
-
-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" ]
@@ -1,93 +1,44 @@
 <p align="center">
-  <img src="./assets/logo.png" width="360x">
+  <img src="./assets/cat_with_sd_cpp_42.png" width="360x">
 </p>

 # stable-diffusion.cpp

-<div align="center">
-<a href="https://trendshift.io/repositories/9714" target="_blank"><img src="https://trendshift.io/api/badge/repositories/9714" alt="leejet%2Fstable-diffusion.cpp | Trendshift" style="width: 250px; height: 55px;" width="250" height="55"/></a>
-</div>
-
 Diffusion model(SD,Flux,Wan,...) inference in pure C/C++

 ***Note that this project is under active development. \
 API and command-line option may change frequently.***

-## 🔥Important News
-
-* **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

- Plain C/C++ implementation based on [ggml](https://github.com/ggml-org/ggml), working in the same way as [llama.cpp](https://github.com/ggml-org/llama.cpp)
+- Plain C/C++ implementation based on [ggml](https://github.com/ggerganov/ggml), working in the same way as [llama.cpp](https://github.com/ggerganov/llama.cpp)
 - Super lightweight and without external dependencies
 - Supported models
  - Image Models
    - SD1.x, SD2.x, [SD-Turbo](https://huggingface.co/stabilityai/sd-turbo)
    - SDXL, [SDXL-Turbo](https://huggingface.co/stabilityai/sdxl-turbo)
-    - [Some SD1.x and SDXL distilled models](./docs/distilled_sd.md)
+      - !!!The VAE in SDXL encounters NaN issues under FP16, but unfortunately, the ggml_conv_2d only operates under FP16. Hence, a parameter is needed to specify the VAE that has fixed the FP16 NaN issue. You can find it here: [SDXL VAE FP16 Fix](https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl_vae.safetensors).
    - [SD3/SD3.5](./docs/sd3.md)
-    - [FLUX.1-dev/FLUX.1-schnell](./docs/flux.md)
-    - [FLUX.2-dev/FLUX.2-klein](./docs/flux2.md)
+    - [Flux-dev/Flux-schnell](./docs/flux.md)
    - [Chroma](./docs/chroma.md)
-    - [Chroma1-Radiance](./docs/chroma_radiance.md)
-    - [Qwen Image](./docs/qwen_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)
  - Image Edit Models
    - [FLUX.1-Kontext-dev](./docs/kontext.md)
-    - [Qwen Image Edit series](./docs/qwen_image_edit.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)
  - Latent Consistency Models support (LCM/LCM-LoRA)
  - Faster and memory efficient latent decoding with [TAESD](https://github.com/madebyollin/taesd)
  - Upscale images generated with [ESRGAN](https://github.com/xinntao/Real-ESRGAN)
- Supported backends
-  - CPU (AVX, AVX2 and AVX512 support for x86 architectures)
-  - CUDA
-  - Vulkan
-  - Metal
-  - OpenCL
-  - SYCL
- Supported weight formats
-  - Pytorch checkpoint (`.ckpt` or `.pth` or `.pt`)
-  - Safetensors (`.safetensors`)
-  - GGUF (`.gguf`)
- Convert mode supports converting model weights to `.gguf` or `.safetensors`
- Supported platforms
-    - Linux
-    - Mac OS
-    - Windows
-    - Android (via Termux, [Local Diffusion](https://github.com/rmatif/Local-Diffusion))
+- 16-bit, 32-bit float support
+- 2-bit, 3-bit, 4-bit, 5-bit and 8-bit integer quantization support
+- Accelerated memory-efficient CPU inference
+    - Only requires ~2.3GB when using txt2img with fp16 precision to generate a 512x512 image, enabling Flash Attention just requires ~1.8GB.
+- AVX, AVX2 and AVX512 support for x86 architectures
+- Full CUDA, Metal, Vulkan, OpenCL and SYCL backend for GPU acceleration.
+- Can load ckpt, safetensors and diffusers models/checkpoints. Standalone VAEs models
+    - No need to convert to `.ggml` or `.gguf` anymore!
 - Flash Attention for memory usage optimization
 - Negative prompt
 - [stable-diffusion-webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui) style tokenizer (not all the features, only token weighting for now)
@@ -100,60 +51,375 @@ API and command-line option may change frequently.***
    - `DPM++ 2M`
    - [`DPM++ 2M v2`](https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/8457)
    - `DPM++ 2S a`
-    - `ER-SDE`
    - [`LCM`](https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/13952)
- Cross-platform reproducibility
-    - `--rng cuda`, default, consistent with the `stable-diffusion-webui GPU RNG`
-    - `--rng cpu`, consistent with the `comfyui RNG`
+- Cross-platform reproducibility (`--rng cuda`, consistent with the `stable-diffusion-webui GPU RNG`)
 - Embedds generation parameters into png output as webui-compatible text string
+- Supported platforms
+    - Linux
+    - Mac OS
+    - Windows
+    - Android (via Termux, [Local Diffusion](https://github.com/rmatif/Local-Diffusion))

-## Quick Start
+## Usage

-### Get the sd executable
+For most users, you can download the built executable program from the latest [release](https://github.com/leejet/stable-diffusion.cpp/releases/latest).
+If the built product does not meet your requirements, you can choose to build it manually.

- Download pre-built binaries from the [releases page](https://github.com/leejet/stable-diffusion.cpp/releases)
- Or build from source by following the [build guide](./docs/build.md)
+### Get the Code

-### Download model weights
-
- download weights(.ckpt or .safetensors or .gguf). For example
-    - Stable Diffusion v1.5 from https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5 
-
-    ```sh
-    curl -L -O https://huggingface.co/runwayml/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.safetensors
-    ```
-
-### Generate an image with just one command
-
-```sh
-./bin/sd-cli -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat"
+```
+git clone --recursive https://github.com/leejet/stable-diffusion.cpp
+cd stable-diffusion.cpp
 ```

-***For detailed command-line arguments, check out [cli doc](./examples/cli/README.md).***
+- If you have already cloned the repository, you can use the following command to update the repository to the latest code.

-## Performance
+```
+cd stable-diffusion.cpp
+git pull origin master
+git submodule init
+git submodule update
+```

-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).
+### Download weights
+
+- download original weights(.ckpt or .safetensors). For example
+    - Stable Diffusion v1.4 from https://huggingface.co/CompVis/stable-diffusion-v-1-4-original
+    - Stable Diffusion v1.5 from https://huggingface.co/runwayml/stable-diffusion-v1-5
+    - Stable Diffuison v2.1 from https://huggingface.co/stabilityai/stable-diffusion-2-1
+    - Stable Diffusion 3 2B from https://huggingface.co/stabilityai/stable-diffusion-3-medium
+
+    ```shell
+    curl -L -O https://huggingface.co/CompVis/stable-diffusion-v-1-4-original/resolve/main/sd-v1-4.ckpt
+    # curl -L -O https://huggingface.co/runwayml/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.safetensors
+    # curl -L -O https://huggingface.co/stabilityai/stable-diffusion-2-1/resolve/main/v2-1_768-nonema-pruned.safetensors
+    # curl -L -O https://huggingface.co/stabilityai/stable-diffusion-3-medium/resolve/main/sd3_medium_incl_clips_t5xxlfp16.safetensors
+    ```
+
+### Build
+
+#### Build from scratch
+
+```shell
+mkdir build
+cd build
+cmake ..
+cmake --build . --config Release
+```
+
+##### Using OpenBLAS
+
+```
+cmake .. -DGGML_OPENBLAS=ON
+cmake --build . --config Release
+```
+
+##### Using CUDA
+
+This provides BLAS acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads). Recommended to have at least 4 GB of VRAM.
+
+```
+cmake .. -DSD_CUDA=ON
+cmake --build . --config Release
+```
+
+##### Using HipBLAS
+This provides BLAS acceleration using the ROCm cores of your AMD GPU. Make sure to have the ROCm toolkit installed.
+
+Windows User Refer to [docs/hipBLAS_on_Windows.md](docs%2FhipBLAS_on_Windows.md) for a comprehensive guide.
+
+```
+export GFX_NAME=$(rocminfo | grep -m 1 -E "gfx[^0]{1}" | sed -e 's/ *Name: *//' | awk '{$1=$1; print}' || echo "rocminfo missing")
+echo $GFX_NAME
+cmake .. -G "Ninja" -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DSD_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DGPU_TARGETS=$GFX_NAME -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON
+cmake --build . --config Release
+```
+
+##### Using MUSA
+
+This provides BLAS acceleration using the MUSA cores of your Moore Threads GPU. Make sure to have the MUSA toolkit installed.
+
+```bash
+cmake .. -DCMAKE_C_COMPILER=/usr/local/musa/bin/clang -DCMAKE_CXX_COMPILER=/usr/local/musa/bin/clang++ -DSD_MUSA=ON -DCMAKE_BUILD_TYPE=Release
+cmake --build . --config Release
+```
+
+##### Using Metal
+
+Using Metal makes the computation run on the GPU. Currently, there are some issues with Metal when performing operations on very large matrices, making it highly inefficient at the moment. Performance improvements are expected in the near future.
+
+```
+cmake .. -DSD_METAL=ON
+cmake --build . --config Release
+```
+
+##### Using Vulkan
+
+Install Vulkan SDK from https://www.lunarg.com/vulkan-sdk/.
+
+```
+cmake .. -DSD_VULKAN=ON
+cmake --build . --config Release
+```
+
+##### Using OpenCL (for Adreno GPU)
+
+Currently, it supports only Adreno GPUs and is primarily optimized for Q4_0 type
+
+To build for Windows ARM please refers to [Windows 11 Arm64
+](https://github.com/ggml-org/llama.cpp/blob/master/docs/backend/OPENCL.md#windows-11-arm64)
+
+Building for Android:
+
+  Android NDK:
+       Download and install the Android NDK from the [official Android developer site](https://developer.android.com/ndk/downloads).
+
+Setup OpenCL Dependencies for NDK:
+
+You need to provide OpenCL headers and the ICD loader library to your NDK sysroot.
+
+*   OpenCL Headers:
+    ```bash
+    # In a temporary working directory
+    git clone https://github.com/KhronosGroup/OpenCL-Headers
+    cd OpenCL-Headers
+    # Replace <YOUR_NDK_PATH> with your actual NDK installation path
+    # e.g., cp -r CL /path/to/android-ndk-r26c/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include
+    sudo cp -r CL <YOUR_NDK_PATH>/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include
+    cd ..
+    ```
+
+*   OpenCL ICD Loader:
+    ```bash
+    # In the same temporary working directory
+    git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader
+    cd OpenCL-ICD-Loader
+    mkdir build_ndk && cd build_ndk
+
+    # Replace <YOUR_NDK_PATH> in the CMAKE_TOOLCHAIN_FILE and OPENCL_ICD_LOADER_HEADERS_DIR
+    cmake .. -G Ninja -DCMAKE_BUILD_TYPE=Release \
+      -DCMAKE_TOOLCHAIN_FILE=<YOUR_NDK_PATH>/build/cmake/android.toolchain.cmake \
+      -DOPENCL_ICD_LOADER_HEADERS_DIR=<YOUR_NDK_PATH>/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include \
+      -DANDROID_ABI=arm64-v8a \
+      -DANDROID_PLATFORM=24 \
+      -DANDROID_STL=c++_shared
+
+    ninja
+    # Replace <YOUR_NDK_PATH>
+    # e.g., cp libOpenCL.so /path/to/android-ndk-r26c/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android
+    sudo cp libOpenCL.so <YOUR_NDK_PATH>/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android
+    cd ../..
+    ```
+
+Build `stable-diffusion.cpp` for Android with OpenCL:
+
+```bash
+mkdir build-android && cd build-android
+
+# Replace <YOUR_NDK_PATH> with your actual NDK installation path
+# e.g., -DCMAKE_TOOLCHAIN_FILE=/path/to/android-ndk-r26c/build/cmake/android.toolchain.cmake
+cmake .. -G Ninja \
+  -DCMAKE_TOOLCHAIN_FILE=<YOUR_NDK_PATH>/build/cmake/android.toolchain.cmake \
+  -DANDROID_ABI=arm64-v8a \
+  -DANDROID_PLATFORM=android-28 \
+  -DGGML_OPENMP=OFF \
+  -DSD_OPENCL=ON
+
+ninja
+```
+*(Note: Don't forget to include `LD_LIBRARY_PATH=/vendor/lib64` in your command line before running the binary)*
+
+##### Using SYCL
+
+Using SYCL makes the computation run on the Intel GPU. Please make sure you have installed the related driver and [Intel® oneAPI Base toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html) before start. More details and steps can refer to [llama.cpp SYCL backend](https://github.com/ggerganov/llama.cpp/blob/master/docs/backend/SYCL.md#linux).
+
+```
+# Export relevant ENV variables
+source /opt/intel/oneapi/setvars.sh
+
+# Option 1: Use FP32 (recommended for better performance in most cases)
+cmake .. -DSD_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+
+# Option 2: Use FP16
+cmake .. -DSD_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON
+
+cmake --build . --config Release
+```
+
+Example of text2img by using SYCL backend:
+
+- download `stable-diffusion` model weight, refer to [download-weight](#download-weights).
+
+- run `./bin/sd -m ../models/sd3_medium_incl_clips_t5xxlfp16.safetensors --cfg-scale 5 --steps 30 --sampling-method euler  -H 1024 -W 1024 --seed 42 -p "fantasy medieval village world inside a glass sphere , high detail, fantasy, realistic, light effect, hyper detail, volumetric lighting, cinematic, macro, depth of field, blur, red light and clouds from the back, highly detailed epic cinematic concept art cg render made in maya, blender and photoshop, octane render, excellent composition, dynamic dramatic cinematic lighting, aesthetic, very inspirational, world inside a glass sphere by james gurney by artgerm with james jean, joe fenton and tristan eaton by ross tran, fine details, 4k resolution"`
+
+<p align="center">
+  <img src="./assets/sycl_sd3_output.png" width="360x">
+</p>
+
+
+
+##### Using Flash Attention
+
+Enabling flash attention for the diffusion model reduces memory usage by varying amounts of MB.
+eg.:
+ - flux 768x768 ~600mb
+ - SD2 768x768 ~1400mb
+
+For most backends, it slows things down, but for cuda it generally speeds it up too.
+At the moment, it is only supported for some models and some backends (like cpu, cuda/rocm, metal).
+
+Run by adding `--diffusion-fa` to the arguments and watch for:
+```
+[INFO ] stable-diffusion.cpp:312  - Using flash attention in the diffusion model
+```
+and the compute buffer shrink in the debug log:
+```
+[DEBUG] ggml_extend.hpp:1004 - flux compute buffer size: 650.00 MB(VRAM)
+```
+
+### Run
+
+```
+usage: ./bin/sd [arguments]
+
+arguments:
+  -h, --help                         show this help message and exit
+  -M, --mode [MODE]                  run mode, one of: [img_gen, vid_gen, convert], default: img_gen
+  -t, --threads N                    number of threads to use during computation (default: -1)
+                                     If threads <= 0, then threads will be set to the number of CPU physical cores
+  --offload-to-cpu                   place the weights in RAM to save VRAM, and automatically load them into VRAM when needed
+  -m, --model [MODEL]                path to full model
+  --diffusion-model                  path to the standalone diffusion model
+  --high-noise-diffusion-model       path to the standalone high noise diffusion model
+  --clip_l                           path to the clip-l text encoder
+  --clip_g                           path to the clip-g text encoder
+  --clip_vision                      path to the clip-vision encoder
+  --t5xxl                            path to the t5xxl text encoder
+  --vae [VAE]                        path to vae
+  --taesd [TAESD_PATH]               path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)
+  --control-net [CONTROL_PATH]       path to control net model
+  --embd-dir [EMBEDDING_PATH]        path to embeddings
+  --upscale-model [ESRGAN_PATH]      path to esrgan model. Upscale images after generate, just RealESRGAN_x4plus_anime_6B supported by now
+  --upscale-repeats                  Run the ESRGAN upscaler this many times (default 1)
+  --type [TYPE]                      weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K, q4_K)
+                                     If not specified, the default is the type of the weight file
+  --tensor-type-rules [EXPRESSION]   weight type per tensor pattern (example: "^vae\.=f16,model\.=q8_0")
+  --lora-model-dir [DIR]             lora model directory
+  -i, --init-img [IMAGE]             path to the init image, required by img2img
+  --mask [MASK]                      path to the mask image, required by img2img with mask
+  -i, --end-img [IMAGE]              path to the end image, required by flf2v
+  --control-image [IMAGE]            path to image condition, control net
+  -r, --ref-image [PATH]             reference image for Flux Kontext models (can be used multiple times)
+  --control-video [PATH]             path to control video frames, It must be a directory path.
+                                     The video frames inside should be stored as images in lexicographical (character) order
+                                     For example, if the control video path is `frames`, the directory contain images such as 00.png, 01.png, 鈥?etc.
+  --increase-ref-index               automatically increase the indices of references images based on the order they are listed (starting with 1).
+  -o, --output OUTPUT                path to write result image to (default: ./output.png)
+  -p, --prompt [PROMPT]              the prompt to render
+  -n, --negative-prompt PROMPT       the negative prompt (default: "")
+  --cfg-scale SCALE                  unconditional guidance scale: (default: 7.0)
+  --img-cfg-scale SCALE              image guidance scale for inpaint or instruct-pix2pix models: (default: same as --cfg-scale)
+  --guidance SCALE                   distilled guidance scale for models with guidance input (default: 3.5)
+  --slg-scale SCALE                  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
+  --eta SCALE                        eta in DDIM, only for DDIM and TCD: (default: 0)
+  --skip-layers LAYERS               Layers to skip for SLG steps: (default: [7,8,9])
+  --skip-layer-start START           SLG enabling point: (default: 0.01)
+  --skip-layer-end END               SLG disabling point: (default: 0.2)
+  --scheduler {discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple} Denoiser sigma scheduler (default: discrete)
+  --sampling-method {euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd}
+                                     sampling method (default: "euler" for Flux/SD3/Wan, "euler_a" otherwise)
+  --timestep-shift N                 shift timestep for NitroFusion models, default: 0, recommended N for NitroSD-Realism around 250 and 500 for NitroSD-Vibrant
+  --steps  STEPS                     number of sample steps (default: 20)
+  --high-noise-cfg-scale SCALE       (high noise) unconditional guidance scale: (default: 7.0)
+  --high-noise-img-cfg-scale SCALE   (high noise) image guidance scale for inpaint or instruct-pix2pix models: (default: same as --cfg-scale)
+  --high-noise-guidance SCALE        (high noise) distilled guidance scale for models with guidance input (default: 3.5)
+  --high-noise-slg-scale SCALE       (high noise) 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
+  --high-noise-eta SCALE             (high noise) eta in DDIM, only for DDIM and TCD: (default: 0)
+  --high-noise-skip-layers LAYERS    (high noise) Layers to skip for SLG steps: (default: [7,8,9])
+  --high-noise-skip-layer-start      (high noise) SLG enabling point: (default: 0.01)
+  --high-noise-skip-layer-end END    (high noise) SLG disabling point: (default: 0.2)
+  --high-noise-scheduler {discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple} Denoiser sigma scheduler (default: discrete)
+  --high-noise-sampling-method {euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd}
+                                     (high noise) sampling method (default: "euler_a")
+  --high-noise-steps  STEPS          (high noise) number of sample steps (default: -1 = auto)
+                                     SLG will be enabled at step int([STEPS]*[START]) and disabled at int([STEPS]*[END])
+  --strength STRENGTH                strength for noising/unnoising (default: 0.75)
+  --control-strength STRENGTH        strength to apply Control Net (default: 0.9)
+                                     1.0 corresponds to full destruction of information in init image
+  -H, --height H                     image height, in pixel space (default: 512)
+  -W, --width W                      image width, in pixel space (default: 512)
+  --rng {std_default, cuda}          RNG (default: cuda)
+  -s SEED, --seed SEED               RNG seed (default: 42, use random seed for < 0)
+  -b, --batch-count COUNT            number of images to generate
+  --clip-skip N                      ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer (default: -1)
+                                     <= 0 represents unspecified, will be 1 for SD1.x, 2 for SD2.x
+  --vae-tiling                       process vae in tiles to reduce memory usage
+  --vae-tile-size [X]x[Y]            tile size for vae tiling (default: 32x32)
+  --vae-relative-tile-size [X]x[Y]   relative tile size for vae tiling, in fraction of image size if < 1, in number of tiles per dim if >=1 (overrides --vae-tile-size)
+  --vae-tile-overlap OVERLAP         tile overlap for vae tiling, in fraction of tile size (default: 0.5)
+  --vae-on-cpu                       keep vae in cpu (for low vram)
+  --clip-on-cpu                      keep clip in cpu (for low vram)
+  --diffusion-fa                     use flash attention in the diffusion model (for low vram)
+                                     Might lower quality, since it implies converting k and v to f16.
+                                     This might crash if it is not supported by the backend.
+  --diffusion-conv-direct            use Conv2d direct in the diffusion model
+                                     This might crash if it is not supported by the backend.
+  --vae-conv-direct                  use Conv2d direct in the vae model (should improve the performance)
+                                     This might crash if it is not supported by the backend.
+  --control-net-cpu                  keep controlnet in cpu (for low vram)
+  --canny                            apply canny preprocessor (edge detection)
+  --color                            colors the logging tags according to level
+  --chroma-disable-dit-mask          disable dit mask for chroma
+  --chroma-enable-t5-mask            enable t5 mask for chroma
+  --chroma-t5-mask-pad  PAD_SIZE     t5 mask pad size of chroma
+  --video-frames                     video frames (default: 1)
+  --fps                              fps (default: 24)
+  --moe-boundary BOUNDARY            timestep boundary for Wan2.2 MoE model. (default: 0.875)
+                                     only enabled if `--high-noise-steps` is set to -1
+  --flow-shift SHIFT                 shift value for Flow models like SD3.x or WAN (default: auto)
+  --vace-strength                    wan vace strength
+  --photo-maker                      path to PHOTOMAKER model
+  --pm-id-images-dir [DIR]           path to PHOTOMAKER input id images dir
+  --pm-id-embed-path [PATH]          path to PHOTOMAKER v2 id embed
+  --pm-style-strength                strength for keeping PHOTOMAKER input identity (default: 20)
+  --normalize-input                  normalize PHOTOMAKER input id images
+  -v, --verbose                      print extra info
+```
+
+#### txt2img example
+
+```sh
+./bin/sd -m ../models/sd-v1-4.ckpt -p "a lovely cat"
+# ./bin/sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat"
+# ./bin/sd -m ../models/sd_xl_base_1.0.safetensors --vae ../models/sdxl_vae-fp16-fix.safetensors -H 1024 -W 1024 -p "a lovely cat" -v
+# ./bin/sd -m ../models/sd3_medium_incl_clips_t5xxlfp16.safetensors -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable Diffusion CPP\"' --cfg-scale 4.5 --sampling-method euler -v --clip-on-cpu
+# ./bin/sd --diffusion-model  ../models/flux1-dev-q3_k.gguf --vae ../models/ae.sft --clip_l ../models/clip_l.safetensors --t5xxl ../models/t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v --clip-on-cpu
+# ./bin/sd -m  ..\models\sd3.5_large.safetensors --clip_l ..\models\clip_l.safetensors --clip_g ..\models\clip_g.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable diffusion 3.5 Large\"' --cfg-scale 4.5 --sampling-method euler -v --clip-on-cpu
+```
+
+Using formats of different precisions will yield results of varying quality.
+
+| f32  | f16  |q8_0  |q5_0  |q5_1  |q4_0  |q4_1  |
+| ----  |----  |----  |----  |----  |----  |----  |
+| ![](./assets/f32.png) |![](./assets/f16.png) |![](./assets/q8_0.png) |![](./assets/q5_0.png) |![](./assets/q5_1.png) |![](./assets/q4_0.png) |![](./assets/q4_1.png) |
+
+#### img2img example
+
+- `./output.png` is the image generated from the above txt2img pipeline
+
+
+```
+./bin/sd -m ../models/sd-v1-4.ckpt -p "cat with blue eyes" -i ./output.png -o ./img2img_output.png --strength 0.4
+```
+
+<p align="center">
+  <img src="./assets/img2img_output.png" width="256x">
+</p>

 ## 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)
- [FLUX.2-dev/FLUX.2-klein](./docs/flux2.md)
- [FLUX.1-Kontext-dev](./docs/kontext.md)
- [Chroma](./docs/chroma.md)
- [🔥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)
 - [LoRA](./docs/lora.md)
 - [LCM/LCM-LoRA](./docs/lcm.md)
 - [Using PhotoMaker to personalize image generation](./docs/photo_maker.md)
@@ -161,7 +427,6 @@ For runtime and parameter backend placement, see the [backend selection guide](.
 - [Using TAESD to faster decoding](./docs/taesd.md)
 - [Docker](./docs/docker.md)
 - [Quantization and GGUF](./docs/quantization_and_gguf.md)
- [Inference acceleration via caching](./docs/caching.md)

 ## Bindings

@@ -169,7 +434,6 @@ 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)
@@ -185,8 +449,6 @@ These projects use `stable-diffusion.cpp` as a backend for their image generatio
 - [Local Diffusion](https://github.com/rmatif/Local-Diffusion)
 - [sd.cpp-webui](https://github.com/daniandtheweb/sd.cpp-webui)
 - [LocalAI](https://github.com/mudler/LocalAI)
- [Neural-Pixel](https://github.com/Luiz-Alcantara/Neural-Pixel)
- [KoboldCpp](https://github.com/LostRuins/koboldcpp)

 ## Contributors

@@ -200,8 +462,7 @@ Thank you to all the people who have already contributed to stable-diffusion.cpp

 ## References

- [ggml](https://github.com/ggml-org/ggml)
- [diffusers](https://github.com/huggingface/diffusers)
+- [ggml](https://github.com/ggerganov/ggml)
 - [stable-diffusion](https://github.com/CompVis/stable-diffusion)
 - [sd3-ref](https://github.com/Stability-AI/sd3-ref)
 - [stable-diffusion-stability-ai](https://github.com/Stability-AI/stablediffusion)
@@ -212,4 +473,4 @@ Thank you to all the people who have already contributed to stable-diffusion.cpp
 - [generative-models](https://github.com/Stability-AI/generative-models/)
 - [PhotoMaker](https://github.com/TencentARC/PhotoMaker)
 - [Wan2.1](https://github.com/Wan-Video/Wan2.1)
- [Wan2.2](https://github.com/Wan-Video/Wan2.2)
+- [Wan2.2](https://github.com/Wan-Video/Wan2.2)
@@ -0,0 +1,969 @@
+#ifndef __CLIP_HPP__
+#define __CLIP_HPP__
+
+#include "ggml_extend.hpp"
+#include "model.h"
+
+/*================================================== CLIPTokenizer ===================================================*/
+
+std::pair<std::unordered_map<std::string, float>, std::string> extract_and_remove_lora(std::string text) {
+    std::regex re("<lora:([^:]+):([^>]+)>");
+    std::smatch matches;
+    std::unordered_map<std::string, float> filename2multiplier;
+
+    while (std::regex_search(text, matches, re)) {
+        std::string filename = matches[1].str();
+        float multiplier     = std::stof(matches[2].str());
+
+        text = std::regex_replace(text, re, "", std::regex_constants::format_first_only);
+
+        if (multiplier == 0.f) {
+            continue;
+        }
+
+        if (filename2multiplier.find(filename) == filename2multiplier.end()) {
+            filename2multiplier[filename] = multiplier;
+        } else {
+            filename2multiplier[filename] += multiplier;
+        }
+    }
+
+    return std::make_pair(filename2multiplier, text);
+}
+
+std::vector<std::pair<int, std::u32string>> bytes_to_unicode() {
+    std::vector<std::pair<int, std::u32string>> byte_unicode_pairs;
+    std::set<int> byte_set;
+    for (int b = static_cast<int>('!'); b <= static_cast<int>('~'); ++b) {
+        byte_set.insert(b);
+        byte_unicode_pairs.push_back(std::pair<int, std::u32string>(b, unicode_value_to_utf32(b)));
+    }
+    for (int b = 161; b <= 172; ++b) {
+        byte_set.insert(b);
+        byte_unicode_pairs.push_back(std::pair<int, std::u32string>(b, unicode_value_to_utf32(b)));
+    }
+    for (int b = 174; b <= 255; ++b) {
+        byte_set.insert(b);
+        byte_unicode_pairs.push_back(std::pair<int, std::u32string>(b, unicode_value_to_utf32(b)));
+    }
+    int n = 0;
+    for (int b = 0; b < 256; ++b) {
+        if (byte_set.find(b) == byte_set.end()) {
+            byte_unicode_pairs.push_back(std::pair<int, std::u32string>(b, unicode_value_to_utf32(n + 256)));
+            ++n;
+        }
+    }
+    // LOG_DEBUG("byte_unicode_pairs %d", byte_unicode_pairs.size());
+    return byte_unicode_pairs;
+}
+
+// Ref: https://github.com/openai/CLIP/blob/main/clip/simple_tokenizer.py
+
+typedef std::function<bool(std::string&, std::vector<int32_t>&)> on_new_token_cb_t;
+
+class CLIPTokenizer {
+private:
+    std::map<int, std::u32string> byte_encoder;
+    std::map<std::u32string, int> byte_decoder;
+    std::map<std::u32string, int> encoder;
+    std::map<int, std::u32string> decoder;
+    std::map<std::pair<std::u32string, std::u32string>, int> bpe_ranks;
+    std::regex pat;
+    int encoder_len;
+    int bpe_len;
+
+public:
+    const std::string UNK_TOKEN = "<|endoftext|>";
+    const std::string BOS_TOKEN = "<|startoftext|>";
+    const std::string EOS_TOKEN = "<|endoftext|>";
+    const std::string PAD_TOKEN = "<|endoftext|>";
+
+    const int UNK_TOKEN_ID = 49407;
+    const int BOS_TOKEN_ID = 49406;
+    const int EOS_TOKEN_ID = 49407;
+    const int PAD_TOKEN_ID = 49407;
+
+private:
+    static std::string strip(const std::string& str) {
+        std::string::size_type start = str.find_first_not_of(" \t\n\r\v\f");
+        std::string::size_type end   = str.find_last_not_of(" \t\n\r\v\f");
+
+        if (start == std::string::npos) {
+            // String contains only whitespace characters
+            return "";
+        }
+
+        return str.substr(start, end - start + 1);
+    }
+
+    static std::string whitespace_clean(std::string text) {
+        text = std::regex_replace(text, std::regex(R"(\s+)"), " ");
+        text = strip(text);
+        return text;
+    }
+
+    static std::set<std::pair<std::u32string, std::u32string>> get_pairs(const std::vector<std::u32string>& subwords) {
+        std::set<std::pair<std::u32string, std::u32string>> pairs;
+        if (subwords.size() == 0) {
+            return pairs;
+        }
+        std::u32string prev_subword = subwords[0];
+        for (int i = 1; i < subwords.size(); i++) {
+            std::u32string subword = subwords[i];
+            std::pair<std::u32string, std::u32string> pair(prev_subword, subword);
+            pairs.insert(pair);
+            prev_subword = subword;
+        }
+        return pairs;
+    }
+
+public:
+    CLIPTokenizer(int pad_token_id = 49407, const std::string& merges_utf8_str = "")
+        : PAD_TOKEN_ID(pad_token_id) {
+        if (merges_utf8_str.size() > 0) {
+            load_from_merges(merges_utf8_str);
+        } else {
+            load_from_merges(ModelLoader::load_merges());
+        }
+    }
+
+    void load_from_merges(const std::string& merges_utf8_str) {
+        auto byte_unicode_pairs = bytes_to_unicode();
+        // printf("byte_unicode_pairs have %lu pairs \n", byte_unicode_pairs.size());
+        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;
+        }
+        // for (auto & pair: byte_unicode_pairs) {
+        //     std::cout << pair.first << ": " << pair.second << std::endl;
+        // }
+        std::vector<std::u32string> merges;
+        size_t start = 0;
+        size_t pos;
+        std::u32string merges_utf32_str = utf8_to_utf32(merges_utf8_str);
+        while ((pos = merges_utf32_str.find('\n', start)) != std::string::npos) {
+            merges.push_back(merges_utf32_str.substr(start, pos - start));
+            start = pos + 1;
+        }
+        // LOG_DEBUG("merges size %llu", merges.size());
+        GGML_ASSERT(merges.size() == 48895);
+        merges = std::vector<std::u32string>(merges.begin() + 1, merges.end());
+        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("%s", utf32_to_utf8(merge.substr(space_pos + 1)).c_str());
+            // printf("%s :: %s | %s \n", utf32_to_utf8(merge).c_str(), utf32_to_utf8(merge.substr(0, space_pos)).c_str(),
+            //                     utf32_to_utf8(merge.substr(space_pos + 1)).c_str());
+        }
+        std::vector<std::u32string> vocab;
+        for (const auto& pair : byte_unicode_pairs) {
+            vocab.push_back(pair.second);
+        }
+        for (const auto& pair : byte_unicode_pairs) {
+            vocab.push_back(pair.second + utf8_to_utf32("</w>"));
+        }
+        for (const auto& merge : merge_pairs) {
+            vocab.push_back(merge.first + merge.second);
+        }
+        vocab.push_back(utf8_to_utf32("<|startoftext|>"));
+        vocab.push_back(utf8_to_utf32("<|endoftext|>"));
+        LOG_DEBUG("vocab size: %llu", vocab.size());
+        int i = 0;
+        for (const auto& token : vocab) {
+            encoder[token] = i;
+            decoder[i]     = token;
+            i++;
+        }
+        encoder_len = i;
+
+        auto it = encoder.find(utf8_to_utf32("img</w>"));
+        if (it != encoder.end()) {
+            LOG_DEBUG("trigger word img already in vocab");
+        } else {
+            LOG_DEBUG("trigger word img not in vocab yet");
+        }
+
+        int rank = 0;
+        for (const auto& merge : merge_pairs) {
+            bpe_ranks[merge] = rank++;
+        }
+        bpe_len = rank;
+    };
+
+    void add_token(const std::string& text) {
+        std::u32string token = utf8_to_utf32(text);
+        auto it              = encoder.find(token);
+        if (it != encoder.end()) {
+            encoder[token]       = encoder_len;
+            decoder[encoder_len] = token;
+            encoder_len++;
+        }
+    }
+
+    std::u32string bpe(const std::u32string& token) {
+        std::vector<std::u32string> word;
+
+        for (int i = 0; i < token.size() - 1; i++) {
+            word.emplace_back(1, token[i]);
+        }
+        word.push_back(token.substr(token.size() - 1) + utf8_to_utf32("</w>"));
+
+        std::set<std::pair<std::u32string, std::u32string>> pairs = get_pairs(word);
+
+        if (pairs.empty()) {
+            return token + utf8_to_utf32("</w>");
+        }
+
+        while (true) {
+            auto min_pair_iter = std::min_element(pairs.begin(),
+                                                  pairs.end(),
+                                                  [&](const std::pair<std::u32string, std::u32string>& a,
+                                                      const std::pair<std::u32string, std::u32string>& b) {
+                                                      if (bpe_ranks.find(a) == bpe_ranks.end()) {
+                                                          return false;
+                                                      } else if (bpe_ranks.find(b) == bpe_ranks.end()) {
+                                                          return true;
+                                                      }
+                                                      return bpe_ranks.at(a) < bpe_ranks.at(b);
+                                                  });
+
+            const std::pair<std::u32string, std::u32string>& bigram = *min_pair_iter;
+
+            if (bpe_ranks.find(bigram) == bpe_ranks.end()) {
+                break;
+            }
+
+            std::u32string first  = bigram.first;
+            std::u32string second = bigram.second;
+            std::vector<std::u32string> new_word;
+            int32_t i = 0;
+
+            while (i < word.size()) {
+                auto it = std::find(word.begin() + i, word.end(), first);
+                if (it == word.end()) {
+                    new_word.insert(new_word.end(), word.begin() + i, word.end());
+                    break;
+                }
+                new_word.insert(new_word.end(), word.begin() + i, it);
+                i = static_cast<int32_t>(std::distance(word.begin(), it));
+
+                if (word[i] == first && i < static_cast<int32_t>(word.size()) - 1 && word[i + 1] == second) {
+                    new_word.push_back(first + second);
+                    i += 2;
+                } else {
+                    new_word.push_back(word[i]);
+                    i += 1;
+                }
+            }
+
+            word = new_word;
+
+            if (word.size() == 1) {
+                break;
+            }
+            pairs = get_pairs(word);
+        }
+
+        std::u32string result;
+        for (int i = 0; i < word.size(); i++) {
+            result += word[i];
+            if (i != word.size() - 1) {
+                result += utf8_to_utf32(" ");
+            }
+        }
+
+        return result;
+    }
+
+    std::vector<int> tokenize(std::string text,
+                              on_new_token_cb_t on_new_token_cb,
+                              size_t max_length = 0,
+                              bool padding      = false) {
+        std::vector<int32_t> tokens = encode(text, on_new_token_cb);
+
+        tokens.insert(tokens.begin(), BOS_TOKEN_ID);
+        if (max_length > 0) {
+            if (tokens.size() > max_length - 1) {
+                tokens.resize(max_length - 1);
+                tokens.push_back(EOS_TOKEN_ID);
+            } else {
+                tokens.push_back(EOS_TOKEN_ID);
+                if (padding) {
+                    tokens.insert(tokens.end(), max_length - tokens.size(), PAD_TOKEN_ID);
+                }
+            }
+        }
+
+        return tokens;
+    }
+
+    void pad_tokens(std::vector<int>& tokens,
+                    std::vector<float>& weights,
+                    size_t max_length = 0,
+                    bool padding      = false) {
+        if (max_length > 0 && padding) {
+            size_t n = std::ceil(tokens.size() * 1.0 / (max_length - 2));
+            if (n == 0) {
+                n = 1;
+            }
+            size_t length = max_length * n;
+            LOG_DEBUG("token length: %llu", length);
+            std::vector<int> new_tokens;
+            std::vector<float> new_weights;
+            new_tokens.push_back(BOS_TOKEN_ID);
+            new_weights.push_back(1.0);
+            int token_idx = 0;
+            for (int i = 1; i < length; i++) {
+                if (token_idx >= tokens.size()) {
+                    break;
+                }
+                if (i % max_length == 0) {
+                    new_tokens.push_back(BOS_TOKEN_ID);
+                    new_weights.push_back(1.0);
+                } else if (i % max_length == max_length - 1) {
+                    new_tokens.push_back(EOS_TOKEN_ID);
+                    new_weights.push_back(1.0);
+                } else {
+                    new_tokens.push_back(tokens[token_idx]);
+                    new_weights.push_back(weights[token_idx]);
+                    token_idx++;
+                }
+            }
+
+            new_tokens.push_back(EOS_TOKEN_ID);
+            new_weights.push_back(1.0);
+            tokens  = new_tokens;
+            weights = new_weights;
+
+            if (padding) {
+                tokens.insert(tokens.end(), length - tokens.size(), PAD_TOKEN_ID);
+                weights.insert(weights.end(), length - weights.size(), 1.0);
+            }
+        }
+    }
+
+    std::string clean_up_tokenization(std::string& text) {
+        std::regex pattern(R"( ,)");
+        // Replace " ," with ","
+        std::string result = std::regex_replace(text, pattern, ",");
+        return result;
+    }
+
+    std::string decode(const std::vector<int>& tokens) {
+        std::string text = "";
+        for (int t : tokens) {
+            if (t == 49406 || t == 49407)
+                continue;
+            std::u32string ts = decoder[t];
+            // printf("%d, %s \n", t,  utf32_to_utf8(ts).c_str());
+            std::string s = utf32_to_utf8(ts);
+            if (s.length() >= 4) {
+                if (ends_with(s, "</w>")) {
+                    text += s.replace(s.length() - 4, s.length() - 1, "") + " ";
+                } else {
+                    text += s;
+                }
+            } else {
+                text += " " + s;
+            }
+        }
+        // std::vector<unsigned char> bytes;
+        // for (auto c : text){
+        //     bytes.push_back(byte_decoder[c]);
+        // }
+
+        // std::string s((char *)bytes.data());
+        // std::string s = "";
+        text = clean_up_tokenization(text);
+        return trim(text);
+    }
+
+    std::vector<int> encode(std::string text, on_new_token_cb_t on_new_token_cb) {
+        std::string original_text = text;
+        std::vector<int32_t> bpe_tokens;
+        text = whitespace_clean(text);
+        std::transform(text.begin(), text.end(), text.begin(), [](unsigned char c) { return std::tolower(c); });
+
+        std::regex pat(R"(<\|startoftext\|>|<\|endoftext\|>|'s|'t|'re|'ve|'m|'ll|'d|[[:alpha:]]+|[[:digit:]]|[^[:space:][:alpha:][:digit:]]+)",
+                       std::regex::icase);
+
+        std::smatch matches;
+        std::string str = text;
+        std::vector<std::string> token_strs;
+        while (std::regex_search(str, matches, pat)) {
+            bool skip = on_new_token_cb(str, bpe_tokens);
+            if (skip) {
+                continue;
+            }
+            for (auto& token : matches) {
+                std::string token_str = token.str();
+                std::u32string utf32_token;
+                for (int i = 0; i < token_str.length(); i++) {
+                    unsigned char b = token_str[i];
+                    utf32_token += byte_encoder[b];
+                }
+                auto bpe_strs = bpe(utf32_token);
+                size_t start  = 0;
+                size_t pos;
+                while ((pos = bpe_strs.find(' ', start)) != std::u32string::npos) {
+                    auto bpe_str = bpe_strs.substr(start, pos - start);
+                    bpe_tokens.push_back(encoder[bpe_str]);
+                    token_strs.push_back(utf32_to_utf8(bpe_str));
+
+                    start = pos + 1;
+                }
+                auto bpe_str = bpe_strs.substr(start, bpe_strs.size() - start);
+                bpe_tokens.push_back(encoder[bpe_str]);
+                token_strs.push_back(utf32_to_utf8(bpe_str));
+            }
+            str = matches.suffix();
+        }
+        std::stringstream ss;
+        ss << "[";
+        for (auto token : token_strs) {
+            ss << "\"" << token << "\", ";
+        }
+        ss << "]";
+        // LOG_DEBUG("split prompt \"%s\" to tokens %s", original_text.c_str(), ss.str().c_str());
+        // printf("split prompt \"%s\" to tokens %s \n", original_text.c_str(), ss.str().c_str());
+        return bpe_tokens;
+    }
+};
+
+/*================================================ FrozenCLIPEmbedder ================================================*/
+
+// Ref: https://github.com/huggingface/transformers/blob/main/src/transformers/models/clip/modeling_clip.py
+
+struct CLIPMLP : public GGMLBlock {
+protected:
+    bool use_gelu;
+
+public:
+    CLIPMLP(int64_t d_model, int64_t intermediate_size) {
+        blocks["fc1"] = std::shared_ptr<GGMLBlock>(new Linear(d_model, intermediate_size));
+        blocks["fc2"] = std::shared_ptr<GGMLBlock>(new Linear(intermediate_size, d_model));
+
+        if (d_model == 1024 || d_model == 1280) {  // SD 2.x
+            use_gelu = true;
+        } else {  // SD 1.x
+            use_gelu = false;
+        }
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [N, n_token, d_model]
+        auto fc1 = std::dynamic_pointer_cast<Linear>(blocks["fc1"]);
+        auto fc2 = std::dynamic_pointer_cast<Linear>(blocks["fc2"]);
+
+        x = fc1->forward(ctx, x);
+        if (use_gelu) {
+            x = ggml_gelu_inplace(ctx, x);
+        } else {
+            x = ggml_gelu_quick_inplace(ctx, x);
+        }
+        x = fc2->forward(ctx, x);
+        return x;
+    }
+};
+
+struct CLIPLayer : public GGMLBlock {
+protected:
+    int64_t d_model;  // hidden_size/embed_dim
+    int64_t n_head;
+    int64_t intermediate_size;
+
+public:
+    CLIPLayer(int64_t d_model,
+              int64_t n_head,
+              int64_t intermediate_size)
+        : d_model(d_model),
+          n_head(n_head),
+          intermediate_size(intermediate_size) {
+        blocks["self_attn"] = std::shared_ptr<GGMLBlock>(new MultiheadAttention(d_model, n_head, true, true));
+
+        blocks["layer_norm1"] = std::shared_ptr<GGMLBlock>(new LayerNorm(d_model));
+        blocks["layer_norm2"] = std::shared_ptr<GGMLBlock>(new LayerNorm(d_model));
+
+        blocks["mlp"] = std::shared_ptr<GGMLBlock>(new CLIPMLP(d_model, intermediate_size));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, ggml_backend_t backend, struct ggml_tensor* x, bool mask = true) {
+        // x: [N, n_token, d_model]
+        auto self_attn   = std::dynamic_pointer_cast<MultiheadAttention>(blocks["self_attn"]);
+        auto layer_norm1 = std::dynamic_pointer_cast<LayerNorm>(blocks["layer_norm1"]);
+        auto layer_norm2 = std::dynamic_pointer_cast<LayerNorm>(blocks["layer_norm2"]);
+        auto mlp         = std::dynamic_pointer_cast<CLIPMLP>(blocks["mlp"]);
+
+        x = ggml_add(ctx, x, self_attn->forward(ctx, backend, layer_norm1->forward(ctx, x), mask));
+        x = ggml_add(ctx, x, mlp->forward(ctx, layer_norm2->forward(ctx, x)));
+        return x;
+    }
+};
+
+struct CLIPEncoder : public GGMLBlock {
+protected:
+    int64_t n_layer;
+
+public:
+    CLIPEncoder(int64_t n_layer,
+                int64_t d_model,
+                int64_t n_head,
+                int64_t intermediate_size)
+        : n_layer(n_layer) {
+        for (int i = 0; i < n_layer; i++) {
+            std::string name = "layers." + std::to_string(i);
+            blocks[name]     = std::shared_ptr<GGMLBlock>(new CLIPLayer(d_model, n_head, intermediate_size));
+        }
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
+                                struct ggml_tensor* x,
+                                int clip_skip = -1,
+                                bool mask     = true) {
+        // x: [N, n_token, d_model]
+        int layer_idx = n_layer - 1;
+        // LOG_DEBUG("clip_skip %d", clip_skip);
+        if (clip_skip > 0) {
+            layer_idx = n_layer - clip_skip;
+        }
+
+        for (int i = 0; i < n_layer; i++) {
+            // LOG_DEBUG("layer %d", i);
+            if (i == layer_idx + 1) {
+                break;
+            }
+            std::string name = "layers." + std::to_string(i);
+            auto layer       = std::dynamic_pointer_cast<CLIPLayer>(blocks[name]);
+            x                = layer->forward(ctx, backend, x, mask);  // [N, n_token, d_model]
+            // LOG_DEBUG("layer %d", i);
+        }
+        return x;
+    }
+};
+
+class CLIPEmbeddings : public GGMLBlock {
+protected:
+    int64_t embed_dim;
+    int64_t vocab_size;
+    int64_t num_positions;
+    bool force_clip_f32;
+
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
+        enum ggml_type token_wtype = GGML_TYPE_F32;
+        if (!force_clip_f32) {
+            auto tensor_type = tensor_types.find(prefix + "token_embedding.weight");
+            if (tensor_type != tensor_types.end())
+                token_wtype = tensor_type->second;
+        }
+        enum ggml_type position_wtype = GGML_TYPE_F32;
+
+        params["token_embedding.weight"]    = ggml_new_tensor_2d(ctx, token_wtype, embed_dim, vocab_size);
+        params["position_embedding.weight"] = ggml_new_tensor_2d(ctx, position_wtype, embed_dim, num_positions);
+    }
+
+public:
+    CLIPEmbeddings(int64_t embed_dim,
+                   int64_t vocab_size    = 49408,
+                   int64_t num_positions = 77,
+                   bool force_clip_f32   = false)
+        : embed_dim(embed_dim),
+          vocab_size(vocab_size),
+          num_positions(num_positions),
+          force_clip_f32(force_clip_f32) {
+    }
+
+    struct ggml_tensor* get_token_embed_weight() {
+        return params["token_embedding.weight"];
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* input_ids,
+                                struct ggml_tensor* custom_embed_weight) {
+        // input_ids: [N, n_token]
+        auto token_embed_weight    = params["token_embedding.weight"];
+        auto position_embed_weight = params["position_embedding.weight"];
+
+        GGML_ASSERT(input_ids->ne[0] == position_embed_weight->ne[1]);
+        input_ids            = ggml_reshape_3d(ctx, input_ids, input_ids->ne[0], 1, input_ids->ne[1]);
+        auto token_embedding = ggml_get_rows(ctx, custom_embed_weight != NULL ? custom_embed_weight : token_embed_weight, input_ids);
+        token_embedding      = ggml_reshape_3d(ctx, token_embedding, token_embedding->ne[0], token_embedding->ne[1], token_embedding->ne[3]);
+
+        // token_embedding + position_embedding
+        auto x = ggml_add(ctx,
+                          token_embedding,
+                          position_embed_weight);  // [N, n_token, embed_dim]
+        return x;
+    }
+};
+
+class CLIPVisionEmbeddings : public GGMLBlock {
+protected:
+    int64_t embed_dim;
+    int64_t num_channels;
+    int64_t patch_size;
+    int64_t image_size;
+    int64_t num_patches;
+    int64_t num_positions;
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
+        enum ggml_type patch_wtype    = GGML_TYPE_F16;
+        enum ggml_type class_wtype    = GGML_TYPE_F32;
+        enum ggml_type position_wtype = GGML_TYPE_F32;
+
+        params["patch_embedding.weight"]    = ggml_new_tensor_4d(ctx, patch_wtype, patch_size, patch_size, num_channels, embed_dim);
+        params["class_embedding"]           = ggml_new_tensor_1d(ctx, class_wtype, embed_dim);
+        params["position_embedding.weight"] = ggml_new_tensor_2d(ctx, position_wtype, embed_dim, num_positions);
+    }
+
+public:
+    CLIPVisionEmbeddings(int64_t embed_dim,
+                         int64_t num_channels = 3,
+                         int64_t patch_size   = 14,
+                         int64_t image_size   = 224)
+        : embed_dim(embed_dim),
+          num_channels(num_channels),
+          patch_size(patch_size),
+          image_size(image_size) {
+        num_patches   = (image_size / patch_size) * (image_size / patch_size);
+        num_positions = num_patches + 1;
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* pixel_values) {
+        // pixel_values: [N, num_channels, image_size, image_size]
+        // return: [N, num_positions, embed_dim]
+        GGML_ASSERT(pixel_values->ne[0] == image_size && pixel_values->ne[1] == image_size && pixel_values->ne[2] == num_channels);
+
+        auto patch_embed_weight    = params["patch_embedding.weight"];
+        auto class_embed_weight    = params["class_embedding"];
+        auto position_embed_weight = params["position_embedding.weight"];
+
+        // concat(patch_embedding, class_embedding) + position_embedding
+        struct ggml_tensor* patch_embedding;
+        int64_t N       = pixel_values->ne[3];
+        patch_embedding = ggml_nn_conv_2d(ctx, pixel_values, patch_embed_weight, NULL, patch_size, patch_size);  // [N, embed_dim, image_size // pacht_size, image_size // pacht_size]
+        patch_embedding = ggml_reshape_3d(ctx, patch_embedding, num_patches, embed_dim, N);                      // [N, embed_dim, num_patches]
+        patch_embedding = ggml_cont(ctx, ggml_permute(ctx, patch_embedding, 1, 0, 2, 3));                        // [N, num_patches, embed_dim]
+        patch_embedding = ggml_reshape_4d(ctx, patch_embedding, 1, embed_dim, num_patches, N);                   // [N, num_patches, embed_dim, 1]
+
+        struct ggml_tensor* class_embedding = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, embed_dim, N);
+        class_embedding                     = ggml_repeat(ctx, class_embed_weight, class_embedding);      // [N, embed_dim]
+        class_embedding                     = ggml_reshape_4d(ctx, class_embedding, 1, embed_dim, 1, N);  // [N, 1, embed_dim, 1]
+
+        struct ggml_tensor* x = ggml_concat(ctx, class_embedding, patch_embedding, 2);  // [N, num_positions, embed_dim, 1]
+        x                     = ggml_reshape_3d(ctx, x, embed_dim, num_positions, N);   // [N, num_positions, embed_dim]
+        x                     = ggml_add(ctx, x, position_embed_weight);
+        return x;  // [N, num_positions, embed_dim]
+    }
+};
+
+// OPENAI_CLIP_VIT_L_14: https://huggingface.co/openai/clip-vit-large-patch14/blob/main/config.json
+// OPEN_CLIP_VIT_H_14: https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K/blob/main/config.json
+// OPEN_CLIP_VIT_BIGG_14: https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k/blob/main/config.json (CLIPTextModelWithProjection)
+
+enum CLIPVersion {
+    OPENAI_CLIP_VIT_L_14,   // SD 1.x and SDXL
+    OPEN_CLIP_VIT_H_14,     // SD 2.x
+    OPEN_CLIP_VIT_BIGG_14,  // SDXL
+};
+
+class CLIPTextModel : public GGMLBlock {
+protected:
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
+        if (version == OPEN_CLIP_VIT_BIGG_14) {
+            enum ggml_type wtype      = GGML_TYPE_F32;
+            params["text_projection"] = ggml_new_tensor_2d(ctx, wtype, projection_dim, hidden_size);
+        }
+    }
+
+public:
+    CLIPVersion version = OPENAI_CLIP_VIT_L_14;
+    // network hparams
+    int32_t vocab_size        = 49408;
+    int32_t n_token           = 77;  // max_position_embeddings
+    int32_t hidden_size       = 768;
+    int32_t intermediate_size = 3072;
+    int32_t n_head            = 12;
+    int32_t n_layer           = 12;    // num_hidden_layers
+    int32_t projection_dim    = 1280;  // only for OPEN_CLIP_VIT_BIGG_14
+    bool with_final_ln        = true;
+
+    CLIPTextModel(CLIPVersion version = OPENAI_CLIP_VIT_L_14,
+                  bool with_final_ln  = true,
+                  bool force_clip_f32 = false)
+        : version(version), with_final_ln(with_final_ln) {
+        if (version == OPEN_CLIP_VIT_H_14) {
+            hidden_size       = 1024;
+            intermediate_size = 4096;
+            n_head            = 16;
+            n_layer           = 24;
+        } else if (version == OPEN_CLIP_VIT_BIGG_14) {  // CLIPTextModelWithProjection
+            hidden_size       = 1280;
+            intermediate_size = 5120;
+            n_head            = 20;
+            n_layer           = 32;
+        }
+
+        blocks["embeddings"]       = std::shared_ptr<GGMLBlock>(new CLIPEmbeddings(hidden_size, vocab_size, n_token, force_clip_f32));
+        blocks["encoder"]          = std::shared_ptr<GGMLBlock>(new CLIPEncoder(n_layer, hidden_size, n_head, intermediate_size));
+        blocks["final_layer_norm"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size));
+    }
+
+    struct ggml_tensor* get_token_embed_weight() {
+        auto embeddings = std::dynamic_pointer_cast<CLIPEmbeddings>(blocks["embeddings"]);
+        return embeddings->get_token_embed_weight();
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
+                                struct ggml_tensor* input_ids,
+                                struct ggml_tensor* tkn_embeddings,
+                                size_t max_token_idx = 0,
+                                bool return_pooled   = false,
+                                int clip_skip        = -1) {
+        // input_ids: [N, n_token]
+        auto embeddings       = std::dynamic_pointer_cast<CLIPEmbeddings>(blocks["embeddings"]);
+        auto encoder          = std::dynamic_pointer_cast<CLIPEncoder>(blocks["encoder"]);
+        auto final_layer_norm = std::dynamic_pointer_cast<LayerNorm>(blocks["final_layer_norm"]);
+
+        auto x = embeddings->forward(ctx, input_ids, tkn_embeddings);  // [N, n_token, hidden_size]
+        x      = encoder->forward(ctx, backend, x, return_pooled ? -1 : clip_skip, true);
+        if (return_pooled || with_final_ln) {
+            x = final_layer_norm->forward(ctx, x);
+        }
+
+        if (return_pooled) {
+            auto text_projection = params["text_projection"];
+            ggml_tensor* pooled  = ggml_view_1d(ctx, x, hidden_size, x->nb[1] * max_token_idx);
+            if (text_projection != NULL) {
+                pooled = ggml_nn_linear(ctx, pooled, text_projection, NULL);
+            } else {
+                LOG_DEBUG("identity projection");
+            }
+            return pooled;  // [hidden_size, 1, 1]
+        }
+
+        return x;  // [N, n_token, hidden_size]
+    }
+};
+
+class CLIPVisionModel : public GGMLBlock {
+public:
+    // network hparams
+    int32_t num_channels      = 3;
+    int32_t patch_size        = 14;
+    int32_t image_size        = 224;
+    int32_t num_positions     = 257;  // (image_size / patch_size)^2 + 1
+    int32_t hidden_size       = 1024;
+    int32_t intermediate_size = 4096;
+    int32_t n_head            = 16;
+    int32_t n_layer           = 24;
+
+public:
+    CLIPVisionModel(CLIPVersion version = OPENAI_CLIP_VIT_L_14) {
+        if (version == OPEN_CLIP_VIT_H_14) {
+            hidden_size       = 1280;
+            intermediate_size = 5120;
+            n_head            = 16;
+            n_layer           = 32;
+        } else if (version == OPEN_CLIP_VIT_BIGG_14) {
+            hidden_size       = 1664;
+            intermediate_size = 8192;
+            n_head            = 16;
+            n_layer           = 48;
+        }
+
+        blocks["embeddings"]     = std::shared_ptr<GGMLBlock>(new CLIPVisionEmbeddings(hidden_size, num_channels, patch_size, image_size));
+        blocks["pre_layernorm"]  = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size));
+        blocks["encoder"]        = std::shared_ptr<GGMLBlock>(new CLIPEncoder(n_layer, hidden_size, n_head, intermediate_size));
+        blocks["post_layernorm"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
+                                struct ggml_tensor* pixel_values,
+                                bool return_pooled = true,
+                                int clip_skip      = -1) {
+        // pixel_values: [N, num_channels, image_size, image_size]
+        auto embeddings     = std::dynamic_pointer_cast<CLIPVisionEmbeddings>(blocks["embeddings"]);
+        auto pre_layernorm  = std::dynamic_pointer_cast<LayerNorm>(blocks["pre_layernorm"]);
+        auto encoder        = std::dynamic_pointer_cast<CLIPEncoder>(blocks["encoder"]);
+        auto post_layernorm = std::dynamic_pointer_cast<LayerNorm>(blocks["post_layernorm"]);
+
+        auto x = embeddings->forward(ctx, pixel_values);  // [N, num_positions, embed_dim]
+        x      = pre_layernorm->forward(ctx, x);
+        x      = encoder->forward(ctx, backend, x, clip_skip, false);
+        // print_ggml_tensor(x, true, "ClipVisionModel x: ");
+        auto last_hidden_state = x;
+        x                      = post_layernorm->forward(ctx, x);  // [N, n_token, hidden_size]
+
+        GGML_ASSERT(x->ne[3] == 1);
+        if (return_pooled) {
+            ggml_tensor* pooled = ggml_cont(ctx, ggml_view_2d(ctx, x, x->ne[0], x->ne[2], x->nb[2], 0));
+            return pooled;  // [N, hidden_size]
+        } else {
+            // return x;  // [N, n_token, hidden_size]
+            return last_hidden_state;  // [N, n_token, hidden_size]
+        }
+    }
+};
+
+class CLIPProjection : public UnaryBlock {
+protected:
+    int64_t in_features;
+    int64_t out_features;
+    bool transpose_weight;
+
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
+        enum ggml_type wtype = get_type(prefix + "weight", tensor_types, GGML_TYPE_F32);
+        if (transpose_weight) {
+            params["weight"] = ggml_new_tensor_2d(ctx, wtype, out_features, in_features);
+        } else {
+            params["weight"] = ggml_new_tensor_2d(ctx, wtype, in_features, out_features);
+        }
+    }
+
+public:
+    CLIPProjection(int64_t in_features,
+                   int64_t out_features,
+                   bool transpose_weight = false)
+        : in_features(in_features),
+          out_features(out_features),
+          transpose_weight(transpose_weight) {}
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        struct ggml_tensor* w = params["weight"];
+        if (transpose_weight) {
+            w = ggml_cont(ctx, ggml_transpose(ctx, w));
+        }
+        return ggml_nn_linear(ctx, x, w, NULL);
+    }
+};
+
+class CLIPVisionModelProjection : public GGMLBlock {
+public:
+    int32_t hidden_size    = 1024;
+    int32_t projection_dim = 768;
+    int32_t image_size     = 224;
+
+public:
+    CLIPVisionModelProjection(CLIPVersion version   = OPENAI_CLIP_VIT_L_14,
+                              bool transpose_proj_w = false) {
+        if (version == OPEN_CLIP_VIT_H_14) {
+            hidden_size    = 1280;
+            projection_dim = 1024;
+        } else if (version == OPEN_CLIP_VIT_BIGG_14) {
+            hidden_size = 1664;
+        }
+
+        blocks["vision_model"]      = std::shared_ptr<GGMLBlock>(new CLIPVisionModel(version));
+        blocks["visual_projection"] = std::shared_ptr<GGMLBlock>(new CLIPProjection(hidden_size, projection_dim, transpose_proj_w));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
+                                struct ggml_tensor* pixel_values,
+                                bool return_pooled = true,
+                                int clip_skip      = -1) {
+        // pixel_values: [N, num_channels, image_size, image_size]
+        // return: [N, projection_dim] if return_pooled else [N, n_token, hidden_size]
+        auto vision_model      = std::dynamic_pointer_cast<CLIPVisionModel>(blocks["vision_model"]);
+        auto visual_projection = std::dynamic_pointer_cast<CLIPProjection>(blocks["visual_projection"]);
+
+        auto x = vision_model->forward(ctx, backend, pixel_values, return_pooled, clip_skip);  // [N, hidden_size] or [N, n_token, hidden_size]
+
+        if (return_pooled) {
+            x = visual_projection->forward(ctx, x);  // [N, projection_dim]
+        }
+
+        return x;
+    }
+};
+
+struct CLIPTextModelRunner : public GGMLRunner {
+    CLIPTextModel model;
+
+    CLIPTextModelRunner(ggml_backend_t backend,
+                        bool offload_params_to_cpu,
+                        const String2GGMLType& tensor_types,
+                        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), model(version, with_final_ln, force_clip_f32) {
+        model.init(params_ctx, tensor_types, prefix);
+    }
+
+    std::string get_desc() {
+        return "clip";
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+        model.get_param_tensors(tensors, prefix);
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
+                                struct ggml_tensor* input_ids,
+                                struct ggml_tensor* embeddings,
+                                size_t max_token_idx = 0,
+                                bool return_pooled   = false,
+                                int clip_skip        = -1) {
+        size_t N       = input_ids->ne[1];
+        size_t n_token = input_ids->ne[0];
+        if (input_ids->ne[0] > model.n_token) {
+            GGML_ASSERT(input_ids->ne[0] % model.n_token == 0);
+            input_ids = ggml_reshape_2d(ctx, input_ids, model.n_token, input_ids->ne[0] / model.n_token);
+        }
+
+        return model.forward(ctx, backend, input_ids, embeddings, max_token_idx, return_pooled, clip_skip);
+    }
+
+    struct ggml_cgraph* build_graph(struct ggml_tensor* input_ids,
+                                    int num_custom_embeddings    = 0,
+                                    void* custom_embeddings_data = NULL,
+                                    size_t max_token_idx         = 0,
+                                    bool return_pooled           = false,
+                                    int clip_skip                = -1) {
+        struct ggml_cgraph* gf = ggml_new_graph(compute_ctx);
+
+        input_ids = to_backend(input_ids);
+
+        struct ggml_tensor* embeddings = NULL;
+
+        if (num_custom_embeddings > 0 && custom_embeddings_data != NULL) {
+            auto token_embed_weight = model.get_token_embed_weight();
+            auto custom_embeddings  = ggml_new_tensor_2d(compute_ctx,
+                                                         token_embed_weight->type,
+                                                         model.hidden_size,
+                                                         num_custom_embeddings);
+            set_backend_tensor_data(custom_embeddings, custom_embeddings_data);
+
+            // concatenate custom embeddings
+            embeddings = ggml_concat(compute_ctx, token_embed_weight, custom_embeddings, 1);
+        }
+
+        struct ggml_tensor* hidden_states = forward(compute_ctx, runtime_backend, input_ids, embeddings, max_token_idx, return_pooled, clip_skip);
+
+        ggml_build_forward_expand(gf, hidden_states);
+
+        return gf;
+    }
+
+    void compute(const int n_threads,
+                 struct ggml_tensor* input_ids,
+                 int num_custom_embeddings,
+                 void* custom_embeddings_data,
+                 size_t max_token_idx,
+                 bool return_pooled,
+                 int clip_skip,
+                 ggml_tensor** output,
+                 ggml_context* output_ctx = NULL) {
+        auto get_graph = [&]() -> struct ggml_cgraph* {
+            return build_graph(input_ids, num_custom_embeddings, custom_embeddings_data, max_token_idx, return_pooled, clip_skip);
+        };
+        GGMLRunner::compute(get_graph, n_threads, true, output, output_ctx);
+    }
+};
+
+#endif  // __CLIP_HPP__
@@ -1,9 +1,7 @@
-#ifndef __COMMON_BLOCK_HPP__
-#define __COMMON_BLOCK_HPP__
+#ifndef __COMMON_HPP__
+#define __COMMON_HPP__

-#include "ggml-backend.h"
 #include "ggml_extend.hpp"
-#include "util.h"

 class DownSampleBlock : public GGMLBlock {
 protected:
@@ -25,12 +23,12 @@ public:
        }
    }

-    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
        // x: [N, channels, h, w]
        if (vae_downsample) {
            auto conv = std::dynamic_pointer_cast<Conv2d>(blocks["conv"]);

-            x = ggml_ext_pad(ctx->ggml_ctx, x, 1, 1, 0, 0, ctx->circular_x_enabled, ctx->circular_y_enabled);
+            x = ggml_pad(ctx, x, 1, 1, 0, 0);
            x = conv->forward(ctx, x);
        } else {
            auto conv = std::dynamic_pointer_cast<Conv2d>(blocks["op"]);
@@ -54,12 +52,12 @@ public:
        blocks["conv"] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, out_channels, {3, 3}, {1, 1}, {1, 1}));
    }

-    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
        // x: [N, channels, h, w]
        auto conv = std::dynamic_pointer_cast<Conv2d>(blocks["conv"]);

-        x = ggml_upscale(ctx->ggml_ctx, x, 2, GGML_SCALE_MODE_NEAREST);  // [N, channels, h*2, w*2]
-        x = conv->forward(ctx, x);                                       // [N, out_channels, h*2, w*2]
+        x = ggml_upscale(ctx, x, 2, GGML_SCALE_MODE_NEAREST);  // [N, channels, h*2, w*2]
+        x = conv->forward(ctx, x);                             // [N, out_channels, h*2, w*2]
        return x;
    }
 };
@@ -82,7 +80,7 @@ protected:
                                       std::pair<int, int> padding) {
        GGML_ASSERT(dims == 2 || dims == 3);
        if (dims == 3) {
-            return std::shared_ptr<GGMLBlock>(new Conv3d(in_channels, out_channels, {kernel_size.first, 1, 1}, {1, 1, 1}, {padding.first, 0, 0}));
+            return std::shared_ptr<GGMLBlock>(new Conv3dnx1x1(in_channels, out_channels, kernel_size.first, 1, padding.first));
        } else {
            return std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, out_channels, kernel_size, {1, 1}, padding));
        }
@@ -123,7 +121,7 @@ public:
        }
    }

-    virtual ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* emb = nullptr) {
+    virtual struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x, struct ggml_tensor* emb = NULL) {
        // For dims==3, we reduce dimension from 5d to 4d by merging h and w, in order not to change ggml
        // [N, c, t, h, w] => [N, c, t, h * w]
        // x: [N, channels, h, w] if dims == 2 else [N, channels, t, h, w]
@@ -133,38 +131,38 @@ public:
        auto out_layers_0 = std::dynamic_pointer_cast<GroupNorm32>(blocks["out_layers.0"]);
        auto out_layers_3 = std::dynamic_pointer_cast<UnaryBlock>(blocks["out_layers.3"]);

-        if (emb == nullptr) {
+        if (emb == NULL) {
            GGML_ASSERT(skip_t_emb);
        }

        // in_layers
        auto h = in_layers_0->forward(ctx, x);
-        h      = ggml_silu_inplace(ctx->ggml_ctx, h);
+        h      = ggml_silu_inplace(ctx, h);
        h      = in_layers_2->forward(ctx, h);  // [N, out_channels, h, w] if dims == 2 else [N, out_channels, t, h, w]

        // emb_layers
        if (!skip_t_emb) {
            auto emb_layer_1 = std::dynamic_pointer_cast<Linear>(blocks["emb_layers.1"]);

-            auto emb_out = ggml_silu(ctx->ggml_ctx, emb);
+            auto emb_out = ggml_silu(ctx, emb);
            emb_out      = emb_layer_1->forward(ctx, emb_out);  // [N, out_channels] if dims == 2 else [N, t, out_channels]

            if (dims == 2) {
-                emb_out = ggml_reshape_4d(ctx->ggml_ctx, emb_out, 1, 1, emb_out->ne[0], emb_out->ne[1]);  // [N, out_channels, 1, 1]
+                emb_out = ggml_reshape_4d(ctx, emb_out, 1, 1, emb_out->ne[0], emb_out->ne[1]);  // [N, out_channels, 1, 1]
            } else {
-                emb_out = ggml_reshape_4d(ctx->ggml_ctx, emb_out, 1, emb_out->ne[0], emb_out->ne[1], emb_out->ne[2]);  // [N, t, out_channels, 1]
+                emb_out = ggml_reshape_4d(ctx, emb_out, 1, emb_out->ne[0], emb_out->ne[1], emb_out->ne[2]);  // [N, t, out_channels, 1]
                if (exchange_temb_dims) {
                    // emb_out = rearrange(emb_out, "b t c ... -> b c t ...")
-                    emb_out = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, emb_out, 0, 2, 1, 3));  // [N, out_channels, t, 1]
+                    emb_out = ggml_cont(ctx, ggml_permute(ctx, emb_out, 0, 2, 1, 3));  // [N, out_channels, t, 1]
                }
            }

-            h = ggml_add(ctx->ggml_ctx, h, emb_out);  // [N, out_channels, h, w] if dims == 2 else [N, out_channels, t, h, w]
+            h = ggml_add(ctx, h, emb_out);  // [N, out_channels, h, w] if dims == 2 else [N, out_channels, t, h, w]
        }

        // out_layers
        h = out_layers_0->forward(ctx, h);
-        h = ggml_silu_inplace(ctx->ggml_ctx, h);
+        h = ggml_silu_inplace(ctx, h);
        // dropout, skip for inference
        h = out_layers_3->forward(ctx, h);

@@ -174,98 +172,68 @@ public:
            x                    = skip_connection->forward(ctx, x);  // [N, out_channels, h, w] if dims == 2 else [N, out_channels, t, h, w]
        }

-        h = ggml_add(ctx->ggml_ctx, h, x);
+        h = ggml_add(ctx, h, x);
        return h;  // [N, out_channels, h, w] if dims == 2 else [N, out_channels, t, h, w]
    }
 };

-class GEGLU : public UnaryBlock {
+class GEGLU : public GGMLBlock {
 protected:
    int64_t dim_in;
    int64_t dim_out;

+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, std::string prefix = "") {
+        enum ggml_type wtype      = get_type(prefix + "proj.weight", tensor_types, GGML_TYPE_F32);
+        enum ggml_type bias_wtype = GGML_TYPE_F32;
+        params["proj.weight"]     = ggml_new_tensor_2d(ctx, wtype, dim_in, dim_out * 2);
+        params["proj.bias"]       = ggml_new_tensor_1d(ctx, bias_wtype, dim_out * 2);
+    }
+
 public:
    GEGLU(int64_t dim_in, int64_t dim_out)
-        : dim_in(dim_in), dim_out(dim_out) {
-        blocks["proj"] = std::shared_ptr<GGMLBlock>(new Linear(dim_in, dim_out * 2));
-    }
+        : dim_in(dim_in), dim_out(dim_out) {}

-    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
        // x: [ne3, ne2, ne1, dim_in]
        // return: [ne3, ne2, ne1, dim_out]
-        auto proj = std::dynamic_pointer_cast<Linear>(blocks["proj"]);
+        struct ggml_tensor* w = params["proj.weight"];
+        struct ggml_tensor* b = params["proj.bias"];

-        x          = proj->forward(ctx, x);  // [ne3, ne2, ne1, dim_out*2]
-        auto x_vec = ggml_ext_chunk(ctx->ggml_ctx, x, 2, 0, false);
-        x          = x_vec[0];  // [ne3, ne2, ne1, dim_out]
-        auto gate  = x_vec[1];  // [ne3, ne2, ne1, dim_out]
+        auto x_w    = ggml_view_2d(ctx, w, w->ne[0], w->ne[1] / 2, w->nb[1], 0);                        // [dim_out, dim_in]
+        auto x_b    = ggml_view_1d(ctx, b, b->ne[0] / 2, 0);                                            // [dim_out, dim_in]
+        auto gate_w = ggml_view_2d(ctx, w, w->ne[0], w->ne[1] / 2, w->nb[1], w->nb[1] * w->ne[1] / 2);  // [dim_out, ]
+        auto gate_b = ggml_view_1d(ctx, b, b->ne[0] / 2, b->nb[0] * b->ne[0] / 2);                      // [dim_out, ]

-        gate = ggml_cont(ctx->ggml_ctx, gate);
+        auto x_in = x;
+        x         = ggml_nn_linear(ctx, x_in, x_w, x_b);        // [ne3, ne2, ne1, dim_out]
+        auto gate = ggml_nn_linear(ctx, x_in, gate_w, gate_b);  // [ne3, ne2, ne1, dim_out]

-        gate = ggml_ext_gelu(ctx->ggml_ctx, gate, true);
+        gate = ggml_gelu_inplace(ctx, gate);

-        x = ggml_mul(ctx->ggml_ctx, x, gate);  // [ne3, ne2, ne1, dim_out]
+        x = ggml_mul(ctx, x, gate);  // [ne3, ne2, ne1, dim_out]

        return x;
    }
 };

-class GELU : public UnaryBlock {
-public:
-    GELU(int64_t dim_in, int64_t dim_out, bool bias = true) {
-        blocks["proj"] = std::shared_ptr<GGMLBlock>(new Linear(dim_in, dim_out, bias));
-    }
-
-    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
-        // x: [ne3, ne2, ne1, dim_in]
-        // return: [ne3, ne2, ne1, dim_out]
-        auto proj = std::dynamic_pointer_cast<Linear>(blocks["proj"]);
-
-        x = proj->forward(ctx, x);
-        x = ggml_ext_gelu(ctx->ggml_ctx, x, true);
-        return x;
-    }
-};
-
 class FeedForward : public GGMLBlock {
 public:
-    enum class Activation {
-        GEGLU,
-        GELU
-    };
    FeedForward(int64_t dim,
                int64_t dim_out,
-                int64_t mult          = 4,
-                Activation activation = Activation::GEGLU,
-                bool precision_fix    = false) {
+                int64_t mult = 4) {
        int64_t inner_dim = dim * mult;
-        if (activation == Activation::GELU) {
-            blocks["net.0"] = std::shared_ptr<GGMLBlock>(new GELU(dim, inner_dim));
-        } else {
-            blocks["net.0"] = std::shared_ptr<GGMLBlock>(new GEGLU(dim, inner_dim));
-        }

+        blocks["net.0"] = std::shared_ptr<GGMLBlock>(new GEGLU(dim, inner_dim));
        // net_1 is nn.Dropout(), skip for inference
-        bool force_prec_f32 = false;
-        float scale         = 1.f;
-        if (precision_fix) {
-            scale = 1.f / 128.f;
-        }
-        // The purpose of the scale here is to prevent NaN issues in certain situations.
-        // For example, when using Vulkan without enabling force_prec_f32,
-        // or when using CUDA but the weights are k-quants.
-        blocks["net.2"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, dim_out, true, false, force_prec_f32, scale));
+        blocks["net.2"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, dim_out));
    }

-    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
        // x: [ne3, ne2, ne1, dim]
        // return: [ne3, ne2, ne1, dim_out]

-        auto net_0 = std::dynamic_pointer_cast<UnaryBlock>(blocks["net.0"]);
+        auto net_0 = std::dynamic_pointer_cast<GEGLU>(blocks["net.0"]);
        auto net_2 = std::dynamic_pointer_cast<Linear>(blocks["net.2"]);
-        if (sd_backend_is(ctx->backend, "Vulkan")) {
-            net_2->set_force_prec_f32(true);
-        }

        x = net_0->forward(ctx, x);  // [ne3, ne2, ne1, inner_dim]
        x = net_2->forward(ctx, x);  // [ne3, ne2, ne1, dim_out]
@@ -279,23 +247,21 @@ protected:
    int64_t context_dim;
    int64_t n_head;
    int64_t d_head;
-    bool xtra_dim = false;
+    bool flash_attn;

 public:
    CrossAttention(int64_t query_dim,
                   int64_t context_dim,
                   int64_t n_head,
-                   int64_t d_head)
+                   int64_t d_head,
+                   bool flash_attn = false)
        : n_head(n_head),
          d_head(d_head),
          query_dim(query_dim),
-          context_dim(context_dim) {
+          context_dim(context_dim),
+          flash_attn(flash_attn) {
        int64_t inner_dim = d_head * n_head;
-        if (context_dim == 320 && d_head == 320) {
-            // LOG_DEBUG("CrossAttention: temp set dim to 1024 for sdxs_09");
-            xtra_dim    = true;
-            context_dim = 1024;
-        }
+
        blocks["to_q"] = std::shared_ptr<GGMLBlock>(new Linear(query_dim, inner_dim, false));
        blocks["to_k"] = std::shared_ptr<GGMLBlock>(new Linear(context_dim, inner_dim, false));
        blocks["to_v"] = std::shared_ptr<GGMLBlock>(new Linear(context_dim, inner_dim, false));
@@ -304,9 +270,10 @@ public:
        // to_out_1 is nn.Dropout(), skip for inference
    }

-    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                         ggml_tensor* x,
-                         ggml_tensor* context) {
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
+                                struct ggml_tensor* x,
+                                struct ggml_tensor* context) {
        // x: [N, n_token, query_dim]
        // context: [N, n_context, context_dim]
        // return: [N, n_token, query_dim]
@@ -320,17 +287,11 @@ public:
        int64_t n_context = context->ne[1];
        int64_t inner_dim = d_head * n_head;

-        auto q = to_q->forward(ctx, x);  // [N, n_token, inner_dim]
-        if (xtra_dim) {
-            // LOG_DEBUG("CrossAttention: temp set dim to 1024 for sdxs_09");
-            context->ne[0] = 1024;  // patch dim
-        }
+        auto q = to_q->forward(ctx, x);        // [N, n_token, inner_dim]
        auto k = to_k->forward(ctx, context);  // [N, n_context, inner_dim]
        auto v = to_v->forward(ctx, context);  // [N, n_context, inner_dim]
-        if (xtra_dim) {
-            context->ne[0] = 320;  // reset dim to orig
-        }
-        x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, n_head, nullptr, false, ctx->flash_attn_enabled);  // [N, n_token, inner_dim]
+
+        x = ggml_nn_attention_ext(ctx, backend, q, k, v, n_head, NULL, false, false, flash_attn);  // [N, n_token, inner_dim]

        x = to_out_0->forward(ctx, x);  // [N, n_token, query_dim]
        return x;
@@ -348,15 +309,16 @@ public:
                          int64_t n_head,
                          int64_t d_head,
                          int64_t context_dim,
-                          bool ff_in = false)
+                          bool ff_in      = false,
+                          bool flash_attn = false)
        : n_head(n_head), d_head(d_head), ff_in(ff_in) {
        // disable_self_attn is always False
        // disable_temporal_crossattention is always False
        // switch_temporal_ca_to_sa is always False
        // inner_dim is always None or equal to dim
        // gated_ff is always True
-        blocks["attn1"] = std::shared_ptr<GGMLBlock>(new CrossAttention(dim, dim, n_head, d_head));
-        blocks["attn2"] = std::shared_ptr<GGMLBlock>(new CrossAttention(dim, context_dim, n_head, d_head));
+        blocks["attn1"] = std::shared_ptr<GGMLBlock>(new CrossAttention(dim, dim, n_head, d_head, flash_attn));
+        blocks["attn2"] = std::shared_ptr<GGMLBlock>(new CrossAttention(dim, context_dim, n_head, d_head, flash_attn));
        blocks["ff"]    = std::shared_ptr<GGMLBlock>(new FeedForward(dim, dim));
        blocks["norm1"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim));
        blocks["norm2"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim));
@@ -368,9 +330,10 @@ public:
        }
    }

-    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                         ggml_tensor* x,
-                         ggml_tensor* context) {
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
+                                struct ggml_tensor* x,
+                                struct ggml_tensor* context) {
        // x: [N, n_token, query_dim]
        // context: [N, n_context, context_dim]
        // return: [N, n_token, query_dim]
@@ -390,21 +353,21 @@ public:
            x           = norm_in->forward(ctx, x);
            x           = ff_in->forward(ctx, x);
            // self.is_res is always True
-            x = ggml_add(ctx->ggml_ctx, x, x_skip);
+            x = ggml_add(ctx, x, x_skip);
        }

        auto r = x;
        x      = norm1->forward(ctx, x);
-        x      = attn1->forward(ctx, x, x);  // self-attention
-        x      = ggml_add(ctx->ggml_ctx, x, r);
+        x      = attn1->forward(ctx, backend, x, x);  // self-attention
+        x      = ggml_add(ctx, x, r);
        r      = x;
        x      = norm2->forward(ctx, x);
-        x      = attn2->forward(ctx, x, context);  // cross-attention
-        x      = ggml_add(ctx->ggml_ctx, x, r);
+        x      = attn2->forward(ctx, backend, x, context);  // cross-attention
+        x      = ggml_add(ctx, x, r);
        r      = x;
        x      = norm3->forward(ctx, x);
        x      = ff->forward(ctx, x);
-        x      = ggml_add(ctx->ggml_ctx, x, r);
+        x      = ggml_add(ctx, x, r);

        return x;
    }
@@ -417,23 +380,6 @@ protected:
    int64_t d_head;
    int64_t depth       = 1;    // 1
    int64_t context_dim = 768;  // hidden_size, 1024 for VERSION_SD2
-    bool use_linear     = false;
-
-    void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") {
-        auto iter = tensor_storage_map.find(prefix + "proj_out.weight");
-        if (iter != tensor_storage_map.end()) {
-            int64_t inner_dim = n_head * d_head;
-            if (iter->second.n_dims == 4 && use_linear) {
-                use_linear         = false;
-                blocks["proj_in"]  = std::make_shared<Conv2d>(in_channels, inner_dim, std::pair{1, 1});
-                blocks["proj_out"] = std::make_shared<Conv2d>(inner_dim, in_channels, std::pair{1, 1});
-            } else if (iter->second.n_dims == 2 && !use_linear) {
-                use_linear         = true;
-                blocks["proj_in"]  = std::make_shared<Linear>(in_channels, inner_dim);
-                blocks["proj_out"] = std::make_shared<Linear>(inner_dim, in_channels);
-            }
-        }
-    }

 public:
    SpatialTransformer(int64_t in_channels,
@@ -441,42 +387,35 @@ public:
                       int64_t d_head,
                       int64_t depth,
                       int64_t context_dim,
-                       bool use_linear)
+                       bool flash_attn = false)
        : in_channels(in_channels),
          n_head(n_head),
          d_head(d_head),
          depth(depth),
-          context_dim(context_dim),
-          use_linear(use_linear) {
+          context_dim(context_dim) {
+        // We will convert unet transformer linear to conv2d 1x1 when loading the weights, so use_linear is always False
        // disable_self_attn is always False
        int64_t inner_dim = n_head * d_head;  // in_channels
        blocks["norm"]    = std::shared_ptr<GGMLBlock>(new GroupNorm32(in_channels));
-        if (use_linear) {
-            blocks["proj_in"] = std::shared_ptr<GGMLBlock>(new Linear(in_channels, inner_dim));
-        } else {
-            blocks["proj_in"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, inner_dim, {1, 1}));
-        }
+        blocks["proj_in"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, inner_dim, {1, 1}));

        for (int i = 0; i < depth; i++) {
            std::string name = "transformer_blocks." + std::to_string(i);
-            blocks[name]     = std::shared_ptr<GGMLBlock>(new BasicTransformerBlock(inner_dim, n_head, d_head, context_dim, false));
+            blocks[name]     = std::shared_ptr<GGMLBlock>(new BasicTransformerBlock(inner_dim, n_head, d_head, context_dim, false, flash_attn));
        }

-        if (use_linear) {
-            blocks["proj_out"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, in_channels));
-        } else {
-            blocks["proj_out"] = std::shared_ptr<GGMLBlock>(new Conv2d(inner_dim, in_channels, {1, 1}));
-        }
+        blocks["proj_out"] = std::shared_ptr<GGMLBlock>(new Conv2d(inner_dim, in_channels, {1, 1}));
    }

-    virtual ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                 ggml_tensor* x,
-                                 ggml_tensor* context) {
+    virtual struct ggml_tensor* forward(struct ggml_context* ctx,
+                                        ggml_backend_t backend,
+                                        struct ggml_tensor* x,
+                                        struct ggml_tensor* context) {
        // x: [N, in_channels, h, w]
        // context: [N, max_position(aka n_token), hidden_size(aka context_dim)]
        auto norm     = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm"]);
-        auto proj_in  = std::dynamic_pointer_cast<UnaryBlock>(blocks["proj_in"]);
-        auto proj_out = std::dynamic_pointer_cast<UnaryBlock>(blocks["proj_out"]);
+        auto proj_in  = std::dynamic_pointer_cast<Conv2d>(blocks["proj_in"]);
+        auto proj_out = std::dynamic_pointer_cast<Conv2d>(blocks["proj_out"]);

        auto x_in         = x;
        int64_t n         = x->ne[3];
@@ -485,45 +424,32 @@ public:
        int64_t inner_dim = n_head * d_head;

        x = norm->forward(ctx, x);
-        if (use_linear) {
-            x = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 1, 2, 0, 3));  // [N, h, w, inner_dim]
-            x = ggml_reshape_3d(ctx->ggml_ctx, x, inner_dim, w * h, n);                // [N, h * w, inner_dim]
-            x = proj_in->forward(ctx, x);                                              // [N, inner_dim, h, w]
-        } else {
-            x = proj_in->forward(ctx, x);                                              // [N, inner_dim, h, w]
-            x = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 1, 2, 0, 3));  // [N, h, w, inner_dim]
-            x = ggml_reshape_3d(ctx->ggml_ctx, x, inner_dim, w * h, n);                // [N, h * w, inner_dim]
-        }
+        x = proj_in->forward(ctx, x);  // [N, inner_dim, h, w]
+
+        x = ggml_cont(ctx, ggml_permute(ctx, x, 1, 2, 0, 3));  // [N, h, w, inner_dim]
+        x = ggml_reshape_3d(ctx, x, inner_dim, w * h, n);      // [N, h * w, inner_dim]

        for (int i = 0; i < depth; i++) {
            std::string name       = "transformer_blocks." + std::to_string(i);
            auto transformer_block = std::dynamic_pointer_cast<BasicTransformerBlock>(blocks[name]);

-            x = transformer_block->forward(ctx, x, context);
+            x = transformer_block->forward(ctx, backend, x, context);
        }

-        if (use_linear) {
-            // proj_out
-            x = proj_out->forward(ctx, x);  // [N, in_channels, h, w]
+        x = ggml_cont(ctx, ggml_permute(ctx, x, 1, 0, 2, 3));  // [N, inner_dim, h * w]
+        x = ggml_reshape_4d(ctx, x, w, h, inner_dim, n);       // [N, inner_dim, h, w]

-            x = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 1, 0, 2, 3));  // [N, inner_dim, h * w]
-            x = ggml_reshape_4d(ctx->ggml_ctx, x, w, h, inner_dim, n);                 // [N, inner_dim, h, w]
-        } else {
-            x = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 1, 0, 2, 3));  // [N, inner_dim, h * w]
-            x = ggml_reshape_4d(ctx->ggml_ctx, x, w, h, inner_dim, n);                 // [N, inner_dim, h, w]
+        // proj_out
+        x = proj_out->forward(ctx, x);  // [N, in_channels, h, w]

-            // proj_out
-            x = proj_out->forward(ctx, x);  // [N, in_channels, h, w]
-        }
-
-        x = ggml_add(ctx->ggml_ctx, x, x_in);
+        x = ggml_add(ctx, x, x_in);
        return x;
    }
 };

 class AlphaBlender : public GGMLBlock {
 protected:
-    void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, std::string prefix = "") override {
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, std::string prefix = "") {
        // Get the type of the "mix_factor" tensor from the input tensors map with the specified prefix
        enum ggml_type wtype = GGML_TYPE_F32;
        params["mix_factor"] = ggml_new_tensor_1d(ctx, wtype, 1);
@@ -532,7 +458,7 @@ protected:
    float get_alpha() {
        // image_only_indicator is always tensor([0.]) and since mix_factor.shape is [1,]
        // so learned_with_images is same as learned
-        float alpha = ggml_ext_backend_tensor_get_f32(params["mix_factor"]);
+        float alpha = ggml_backend_tensor_get_f32(params["mix_factor"]);
        return sigmoid(alpha);
    }

@@ -543,23 +469,23 @@ public:
        // since mix_factor.shape is [1,], we don't need rearrange using  rearrange_pattern
    }

-    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                         ggml_tensor* x_spatial,
-                         ggml_tensor* x_temporal) {
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* x_spatial,
+                                struct ggml_tensor* x_temporal) {
        // image_only_indicator is always tensor([0.])
        float alpha = get_alpha();
-        auto x      = ggml_add(ctx->ggml_ctx,
-                               ggml_ext_scale(ctx->ggml_ctx, x_spatial, alpha),
-                               ggml_ext_scale(ctx->ggml_ctx, x_temporal, 1.0f - alpha));
+        auto x      = ggml_add(ctx,
+                               ggml_scale(ctx, x_spatial, alpha),
+                               ggml_scale(ctx, x_temporal, 1.0f - alpha));
        return x;
    }
 };

 class VideoResBlock : public ResBlock {
 public:
-    VideoResBlock(int64_t channels,
-                  int64_t emb_channels,
-                  int64_t out_channels,
+    VideoResBlock(int channels,
+                  int emb_channels,
+                  int out_channels,
                  std::pair<int, int> kernel_size = {3, 3},
                  int64_t video_kernel_size       = 3,
                  int dims                        = 2)  // always 2
@@ -568,10 +494,10 @@ public:
        blocks["time_mixer"] = std::shared_ptr<GGMLBlock>(new AlphaBlender());
    }

-    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                         ggml_tensor* x,
-                         ggml_tensor* emb,
-                         int num_video_frames) {
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* x,
+                                struct ggml_tensor* emb,
+                                int num_video_frames) {
        // x: [N, channels, h, w] aka [b*t, channels, h, w]
        // emb: [N, emb_channels] aka [b*t, emb_channels]
        // image_only_indicator is always tensor([0.])
@@ -586,21 +512,21 @@ public:
        int64_t H = x->ne[1];
        int64_t W = x->ne[0];

-        x          = ggml_reshape_4d(ctx->ggml_ctx, x, W * H, C, T, B);                     // (b t) c h w -> b t c (h w)
-        x          = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 2, 1, 3));  // b t c (h w) -> b c t (h w)
+        x          = ggml_reshape_4d(ctx, x, W * H, C, T, B);           // (b t) c h w -> b t c (h w)
+        x          = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // b t c (h w) -> b c t (h w)
        auto x_mix = x;

-        emb = ggml_reshape_4d(ctx->ggml_ctx, emb, emb->ne[0], T, B, emb->ne[3]);  // (b t) ... -> b t ...
+        emb = ggml_reshape_4d(ctx, emb, emb->ne[0], T, B, emb->ne[3]);  // (b t) ... -> b t ...

        x = time_stack->forward(ctx, x, emb);  // b t c (h w)

        x = time_mixer->forward(ctx, x_mix, x);  // b t c (h w)

-        x = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 2, 1, 3));  // b c t (h w) -> b t c (h w)
-        x = ggml_reshape_4d(ctx->ggml_ctx, x, W, H, C, T * B);                     // b t c (h w) -> (b t) c h w
+        x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // b c t (h w) -> b t c (h w)
+        x = ggml_reshape_4d(ctx, x, W, H, C, T * B);           // b t c (h w) -> (b t) c h w

        return x;
    }
 };

-#endif  // __COMMON_BLOCK_HPP__
+#endif  // __COMMON_HPP__
@@ -1,7 +1,8 @@
 #ifndef __CONTROL_HPP__
 #define __CONTROL_HPP__

-#include "common_block.hpp"
+#include "common.hpp"
+#include "ggml_extend.hpp"
 #include "model.h"

 #define CONTROL_NET_GRAPH_SIZE 1536
@@ -26,7 +27,6 @@ protected:
    int num_heads                          = 8;
    int num_head_channels                  = -1;   // channels // num_heads
    int context_dim                        = 768;  // 1024 for VERSION_SD2, 2048 for VERSION_SDXL
-    bool use_linear_projection             = false;

 public:
    int model_channels  = 320;
@@ -82,7 +82,7 @@ public:
                                       int64_t d_head,
                                       int64_t depth,
                                       int64_t context_dim) -> SpatialTransformer* {
-            return new SpatialTransformer(in_channels, n_head, d_head, depth, context_dim, use_linear_projection);
+            return new SpatialTransformer(in_channels, n_head, d_head, depth, context_dim);
        };

        auto make_zero_conv = [&](int64_t channels) {
@@ -164,26 +164,27 @@ public:
        blocks["middle_block_out.0"] = std::shared_ptr<GGMLBlock>(make_zero_conv(ch));
    }

-    ggml_tensor* resblock_forward(std::string name,
-                                  GGMLRunnerContext* ctx,
-                                  ggml_tensor* x,
-                                  ggml_tensor* emb) {
+    struct ggml_tensor* resblock_forward(std::string name,
+                                         struct ggml_context* ctx,
+                                         struct ggml_tensor* x,
+                                         struct ggml_tensor* emb) {
        auto block = std::dynamic_pointer_cast<ResBlock>(blocks[name]);
        return block->forward(ctx, x, emb);
    }

-    ggml_tensor* attention_layer_forward(std::string name,
-                                         GGMLRunnerContext* ctx,
-                                         ggml_tensor* x,
-                                         ggml_tensor* context) {
+    struct ggml_tensor* attention_layer_forward(std::string name,
+                                                struct ggml_context* ctx,
+                                                ggml_backend_t backend,
+                                                struct ggml_tensor* x,
+                                                struct ggml_tensor* context) {
        auto block = std::dynamic_pointer_cast<SpatialTransformer>(blocks[name]);
-        return block->forward(ctx, x, context);
+        return block->forward(ctx, backend, x, context);
    }

-    ggml_tensor* input_hint_block_forward(GGMLRunnerContext* ctx,
-                                          ggml_tensor* hint,
-                                          ggml_tensor* emb,
-                                          ggml_tensor* context) {
+    struct ggml_tensor* input_hint_block_forward(struct ggml_context* ctx,
+                                                 struct ggml_tensor* hint,
+                                                 struct ggml_tensor* emb,
+                                                 struct ggml_tensor* context) {
        int num_input_blocks = 15;
        auto h               = hint;
        for (int i = 0; i < num_input_blocks; i++) {
@@ -192,32 +193,33 @@ public:

                h = block->forward(ctx, h);
            } else {
-                h = ggml_silu_inplace(ctx->ggml_ctx, h);
+                h = ggml_silu_inplace(ctx, h);
            }
        }
        return h;
    }

-    std::vector<ggml_tensor*> forward(GGMLRunnerContext* ctx,
-                                      ggml_tensor* x,
-                                      ggml_tensor* hint,
-                                      ggml_tensor* guided_hint,
-                                      ggml_tensor* timesteps,
-                                      ggml_tensor* context,
-                                      ggml_tensor* y = nullptr) {
+    std::vector<struct ggml_tensor*> forward(struct ggml_context* ctx,
+                                             ggml_backend_t backend,
+                                             struct ggml_tensor* x,
+                                             struct ggml_tensor* hint,
+                                             struct ggml_tensor* guided_hint,
+                                             struct ggml_tensor* timesteps,
+                                             struct ggml_tensor* context,
+                                             struct ggml_tensor* y = NULL) {
        // x: [N, in_channels, h, w] or [N, in_channels/2, h, w]
        // timesteps: [N,]
        // context: [N, max_position, hidden_size] or [1, max_position, hidden_size]. for example, [N, 77, 768]
        // y: [N, adm_in_channels] or [1, adm_in_channels]
-        if (context != nullptr) {
+        if (context != NULL) {
            if (context->ne[2] != x->ne[3]) {
-                context = ggml_repeat(ctx->ggml_ctx, context, ggml_new_tensor_3d(ctx->ggml_ctx, GGML_TYPE_F32, context->ne[0], context->ne[1], x->ne[3]));
+                context = ggml_repeat(ctx, context, ggml_new_tensor_3d(ctx, GGML_TYPE_F32, context->ne[0], context->ne[1], x->ne[3]));
            }
        }

-        if (y != nullptr) {
+        if (y != NULL) {
            if (y->ne[1] != x->ne[3]) {
-                y = ggml_repeat(ctx->ggml_ctx, y, ggml_new_tensor_2d(ctx->ggml_ctx, GGML_TYPE_F32, y->ne[0], x->ne[3]));
+                y = ggml_repeat(ctx, y, ggml_new_tensor_2d(ctx, GGML_TYPE_F32, y->ne[0], x->ne[3]));
            }
        }

@@ -228,27 +230,27 @@ public:

        auto middle_block_out = std::dynamic_pointer_cast<Conv2d>(blocks["middle_block_out.0"]);

-        auto t_emb = ggml_ext_timestep_embedding(ctx->ggml_ctx, timesteps, model_channels);  // [N, model_channels]
+        auto t_emb = ggml_nn_timestep_embedding(ctx, timesteps, model_channels);  // [N, model_channels]

        auto emb = time_embed_0->forward(ctx, t_emb);
-        emb      = ggml_silu_inplace(ctx->ggml_ctx, emb);
+        emb      = ggml_silu_inplace(ctx, emb);
        emb      = time_embed_2->forward(ctx, emb);  // [N, time_embed_dim]

        // SDXL/SVD
-        if (y != nullptr) {
+        if (y != NULL) {
            auto label_embed_0 = std::dynamic_pointer_cast<Linear>(blocks["label_emb.0.0"]);
            auto label_embed_2 = std::dynamic_pointer_cast<Linear>(blocks["label_emb.0.2"]);

            auto label_emb = label_embed_0->forward(ctx, y);
-            label_emb      = ggml_silu_inplace(ctx->ggml_ctx, label_emb);
+            label_emb      = ggml_silu_inplace(ctx, label_emb);
            label_emb      = label_embed_2->forward(ctx, label_emb);  // [N, time_embed_dim]

-            emb = ggml_add(ctx->ggml_ctx, emb, label_emb);  // [N, time_embed_dim]
+            emb = ggml_add(ctx, emb, label_emb);  // [N, time_embed_dim]
        }

-        std::vector<ggml_tensor*> outs;
+        std::vector<struct ggml_tensor*> outs;

-        if (guided_hint == nullptr) {
+        if (guided_hint == NULL) {
            guided_hint = input_hint_block_forward(ctx, hint, emb, context);
        }
        outs.push_back(guided_hint);
@@ -257,7 +259,7 @@ public:

        // input block 0
        auto h = input_blocks_0_0->forward(ctx, x);
-        h      = ggml_add(ctx->ggml_ctx, h, guided_hint);
+        h      = ggml_add(ctx, h, guided_hint);
        outs.push_back(zero_convs_0->forward(ctx, h));

        // input block 1-11
@@ -272,7 +274,7 @@ public:
                h                = resblock_forward(name, ctx, h, emb);  // [N, mult*model_channels, h, w]
                if (std::find(attention_resolutions.begin(), attention_resolutions.end(), ds) != attention_resolutions.end()) {
                    std::string name = "input_blocks." + std::to_string(input_block_idx) + ".1";
-                    h                = attention_layer_forward(name, ctx, h, context);  // [N, mult*model_channels, h, w]
+                    h                = attention_layer_forward(name, ctx, backend, h, context);  // [N, mult*model_channels, h, w]
                }

                auto zero_conv = std::dynamic_pointer_cast<Conv2d>(blocks["zero_convs." + std::to_string(input_block_idx) + ".0"]);
@@ -296,9 +298,9 @@ public:
        // [N, 4*model_channels, h/8, w/8]

        // middle_block
-        h = resblock_forward("middle_block.0", ctx, h, emb);             // [N, 4*model_channels, h/8, w/8]
-        h = attention_layer_forward("middle_block.1", ctx, h, context);  // [N, 4*model_channels, h/8, w/8]
-        h = resblock_forward("middle_block.2", ctx, h, emb);             // [N, 4*model_channels, h/8, w/8]
+        h = resblock_forward("middle_block.0", ctx, h, emb);                      // [N, 4*model_channels, h/8, w/8]
+        h = attention_layer_forward("middle_block.1", ctx, backend, h, context);  // [N, 4*model_channels, h/8, w/8]
+        h = resblock_forward("middle_block.2", ctx, h, emb);                      // [N, 4*model_channels, h/8, w/8]

        // out
        outs.push_back(middle_block_out->forward(ctx, h));
@@ -310,43 +312,52 @@ struct ControlNet : public GGMLRunner {
    SDVersion version = VERSION_SD1;
    ControlNetBlock control_net;

-    ggml_backend_buffer_t control_buffer = nullptr;
-    ggml_context* control_ctx            = nullptr;
-    std::vector<ggml_tensor*> control_outputs_ggml;
-    ggml_tensor* guided_hint_output_ggml = nullptr;
-    std::vector<sd::Tensor<float>> controls;
-    sd::Tensor<float> guided_hint;
-    bool guided_hint_cached = false;
+    ggml_backend_buffer_t control_buffer = NULL;  // keep control output tensors in backend memory
+    ggml_context* control_ctx            = NULL;
+    std::vector<struct ggml_tensor*> controls;  // (12 input block outputs, 1 middle block output) SD 1.5
+    struct ggml_tensor* guided_hint = NULL;     // guided_hint cache, for faster inference
+    bool guided_hint_cached         = false;

    ControlNet(ggml_backend_t backend,
-               ggml_backend_t params_backend,
-               const String2TensorStorage& tensor_storage_map = {},
-               SDVersion version                              = VERSION_SD1)
-        : GGMLRunner(backend, params_backend), control_net(version) {
-        control_net.init(params_ctx, tensor_storage_map, "");
+               bool offload_params_to_cpu,
+               const String2GGMLType& tensor_types = {},
+               SDVersion version                   = VERSION_SD1)
+        : GGMLRunner(backend, offload_params_to_cpu), control_net(version) {
+        control_net.init(params_ctx, tensor_types, "");
    }

-    ~ControlNet() override {
+    void enable_conv2d_direct() {
+        std::vector<GGMLBlock*> blocks;
+        control_net.get_all_blocks(blocks);
+        for (auto block : blocks) {
+            if (block->get_desc() == "Conv2d") {
+                auto conv_block = (Conv2d*)block;
+                conv_block->enable_direct();
+            }
+        }
+    }
+
+    ~ControlNet() {
        free_control_ctx();
    }

-    void alloc_control_ctx(std::vector<ggml_tensor*> outs) {
-        ggml_init_params params;
+    void alloc_control_ctx(std::vector<struct ggml_tensor*> outs) {
+        struct ggml_init_params params;
        params.mem_size   = static_cast<size_t>(outs.size() * ggml_tensor_overhead()) + 1024 * 1024;
-        params.mem_buffer = nullptr;
+        params.mem_buffer = NULL;
        params.no_alloc   = true;
        control_ctx       = ggml_init(params);

-        control_outputs_ggml.resize(outs.size() - 1);
+        controls.resize(outs.size() - 1);

        size_t control_buffer_size = 0;

-        guided_hint_output_ggml = ggml_dup_tensor(control_ctx, outs[0]);
-        control_buffer_size += ggml_nbytes(guided_hint_output_ggml);
+        guided_hint = ggml_dup_tensor(control_ctx, outs[0]);
+        control_buffer_size += ggml_nbytes(guided_hint);

        for (int i = 0; i < outs.size() - 1; i++) {
-            control_outputs_ggml[i] = ggml_dup_tensor(control_ctx, outs[i + 1]);
-            control_buffer_size += ggml_nbytes(control_outputs_ggml[i]);
+            controls[i] = ggml_dup_tensor(control_ctx, outs[i + 1]);
+            control_buffer_size += ggml_nbytes(controls[i]);
        }

        control_buffer = ggml_backend_alloc_ctx_tensors(control_ctx, runtime_backend);
@@ -355,107 +366,86 @@ struct ControlNet : public GGMLRunner {
    }

    void free_control_ctx() {
-        if (control_buffer != nullptr) {
+        if (control_buffer != NULL) {
            ggml_backend_buffer_free(control_buffer);
-            control_buffer = nullptr;
+            control_buffer = NULL;
        }
-        if (control_ctx != nullptr) {
+        if (control_ctx != NULL) {
            ggml_free(control_ctx);
-            control_ctx = nullptr;
+            control_ctx = NULL;
        }
-        guided_hint_output_ggml = nullptr;
-        guided_hint_cached      = false;
-        guided_hint             = {};
-        control_outputs_ggml.clear();
+        guided_hint        = NULL;
+        guided_hint_cached = false;
        controls.clear();
    }

-    std::string get_desc() override {
+    std::string get_desc() {
        return "control_net";
    }

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

-    ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,
-                             const sd::Tensor<float>& hint_tensor,
-                             const sd::Tensor<float>& timesteps_tensor,
-                             const sd::Tensor<float>& context_tensor = {},
-                             const sd::Tensor<float>& y_tensor       = {}) {
-        ggml_cgraph* gf = new_graph_custom(CONTROL_NET_GRAPH_SIZE);
+    struct ggml_cgraph* build_graph(struct ggml_tensor* x,
+                                    struct ggml_tensor* hint,
+                                    struct ggml_tensor* timesteps,
+                                    struct ggml_tensor* context,
+                                    struct ggml_tensor* y = NULL) {
+        struct ggml_cgraph* gf = ggml_new_graph_custom(compute_ctx, CONTROL_NET_GRAPH_SIZE, false);

-        ggml_tensor* x         = make_input(x_tensor);
-        ggml_tensor* hint      = nullptr;
-        ggml_tensor* timesteps = make_input(timesteps_tensor);
-        ggml_tensor* context   = make_optional_input(context_tensor);
-        ggml_tensor* y         = make_optional_input(y_tensor);
-
-        ggml_tensor* guided_hint_input = nullptr;
-        if (guided_hint_cached && !guided_hint.empty()) {
-            guided_hint_input = make_input(guided_hint);
-            hint              = nullptr;
+        x = to_backend(x);
+        if (guided_hint_cached) {
+            hint = NULL;
        } else {
-            hint = make_input(hint_tensor);
+            hint = to_backend(hint);
        }
+        context   = to_backend(context);
+        y         = to_backend(y);
+        timesteps = to_backend(timesteps);

-        auto runner_ctx = get_context();
-
-        auto outs = control_net.forward(&runner_ctx,
+        auto outs = control_net.forward(compute_ctx,
+                                        runtime_backend,
                                        x,
                                        hint,
-                                        guided_hint_input,
+                                        guided_hint_cached ? guided_hint : NULL,
                                        timesteps,
                                        context,
                                        y);

-        if (control_ctx == nullptr) {
+        if (control_ctx == NULL) {
            alloc_control_ctx(outs);
        }

-        ggml_build_forward_expand(gf, ggml_cpy(compute_ctx, outs[0], guided_hint_output_ggml));
+        ggml_build_forward_expand(gf, ggml_cpy(compute_ctx, outs[0], guided_hint));
        for (int i = 0; i < outs.size() - 1; i++) {
-            ggml_build_forward_expand(gf, ggml_cpy(compute_ctx, outs[i + 1], control_outputs_ggml[i]));
+            ggml_build_forward_expand(gf, ggml_cpy(compute_ctx, outs[i + 1], controls[i]));
        }

        return gf;
    }

-    std::optional<std::vector<sd::Tensor<float>>> compute(int n_threads,
-                                                          const sd::Tensor<float>& x,
-                                                          const sd::Tensor<float>& hint,
-                                                          const sd::Tensor<float>& timesteps,
-                                                          const sd::Tensor<float>& context = {},
-                                                          const sd::Tensor<float>& y       = {}) {
+    void compute(int n_threads,
+                 struct ggml_tensor* x,
+                 struct ggml_tensor* hint,
+                 struct ggml_tensor* timesteps,
+                 struct ggml_tensor* context,
+                 struct ggml_tensor* y,
+                 struct ggml_tensor** output     = NULL,
+                 struct ggml_context* output_ctx = NULL) {
        // x: [N, in_channels, h, w]
        // timesteps: [N, ]
        // context: [N, max_position, hidden_size]([N, 77, 768]) or [1, max_position, hidden_size]
        // y: [N, adm_in_channels] or [1, adm_in_channels]
-        auto get_graph = [&]() -> ggml_cgraph* {
+        auto get_graph = [&]() -> struct ggml_cgraph* {
            return build_graph(x, hint, timesteps, context, y);
        };

-        auto compute_result = GGMLRunner::compute<float>(get_graph, n_threads, false);
-        if (!compute_result.has_value()) {
-            return std::nullopt;
-        }
-
-        if (guided_hint_output_ggml != nullptr) {
-            guided_hint = restore_trailing_singleton_dims(sd::make_sd_tensor_from_ggml<float>(guided_hint_output_ggml),
-                                                          4);
-        }
-        controls.clear();
-        controls.reserve(control_outputs_ggml.size());
-        for (ggml_tensor* control : control_outputs_ggml) {
-            auto control_host = restore_trailing_singleton_dims(sd::make_sd_tensor_from_ggml<float>(control), 4);
-            GGML_ASSERT(!control_host.empty());
-            controls.push_back(std::move(control_host));
-        }
+        GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
        guided_hint_cached = true;
-        return controls;
    }

-    bool load_from_file(const std::string& file_path, int n_threads) {
+    bool load_from_file(const std::string& file_path) {
        LOG_INFO("loading control net from '%s'", file_path.c_str());
        alloc_params_buffer();
        std::map<std::string, ggml_tensor*> tensors;
@@ -463,12 +453,12 @@ struct ControlNet : public GGMLRunner {
        std::set<std::string> ignore_tensors;

        ModelLoader model_loader;
-        if (!model_loader.init_from_file_and_convert_name(file_path)) {
+        if (!model_loader.init_from_file(file_path)) {
            LOG_ERROR("init control net model loader from file failed: '%s'", file_path.c_str());
            return false;
        }

-        bool success = model_loader.load_tensors(tensors, ignore_tensors, n_threads);
+        bool success = model_loader.load_tensors(tensors, ignore_tensors);

        if (!success) {
            LOG_ERROR("load control net tensors from model loader failed");
@@ -480,4 +470,4 @@ struct ControlNet : public GGMLRunner {
    }
 };

-#endif  // __CONTROL_HPP__
+#endif  // __CONTROL_HPP__
@@ -0,0 +1,266 @@
+#ifndef __DIFFUSION_MODEL_H__
+#define __DIFFUSION_MODEL_H__
+
+#include "flux.hpp"
+#include "mmdit.hpp"
+#include "unet.hpp"
+#include "wan.hpp"
+
+struct DiffusionParams {
+    struct ggml_tensor* x                     = NULL;
+    struct ggml_tensor* timesteps             = NULL;
+    struct ggml_tensor* context               = NULL;
+    struct ggml_tensor* c_concat              = NULL;
+    struct ggml_tensor* y                     = NULL;
+    struct ggml_tensor* guidance              = NULL;
+    std::vector<ggml_tensor*> ref_latents     = {};
+    bool increase_ref_index                   = false;
+    int num_video_frames                      = -1;
+    std::vector<struct ggml_tensor*> controls = {};
+    float control_strength                    = 0.f;
+    struct ggml_tensor* vace_context          = NULL;
+    float vace_strength                       = 1.f;
+    std::vector<int> skip_layers              = {};
+};
+
+struct DiffusionModel {
+    virtual std::string get_desc()                                                      = 0;
+    virtual void compute(int n_threads,
+                         DiffusionParams diffusion_params,
+                         struct ggml_tensor** output     = NULL,
+                         struct ggml_context* output_ctx = NULL)                        = 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, struct ggml_tensor*>& tensors) = 0;
+    virtual size_t get_params_buffer_size()                                             = 0;
+    virtual int64_t get_adm_in_channels()                                               = 0;
+};
+
+struct UNetModel : public DiffusionModel {
+    UNetModelRunner unet;
+
+    UNetModel(ggml_backend_t backend,
+              bool offload_params_to_cpu,
+              const String2GGMLType& tensor_types = {},
+              SDVersion version                   = VERSION_SD1,
+              bool flash_attn                     = false)
+        : unet(backend, offload_params_to_cpu, tensor_types, "model.diffusion_model", version, flash_attn) {
+    }
+
+    std::string get_desc() {
+        return unet.get_desc();
+    }
+
+    void alloc_params_buffer() {
+        unet.alloc_params_buffer();
+    }
+
+    void free_params_buffer() {
+        unet.free_params_buffer();
+    }
+
+    void free_compute_buffer() {
+        unet.free_compute_buffer();
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
+        unet.get_param_tensors(tensors, "model.diffusion_model");
+    }
+
+    size_t get_params_buffer_size() {
+        return unet.get_params_buffer_size();
+    }
+
+    int64_t get_adm_in_channels() {
+        return unet.unet.adm_in_channels;
+    }
+
+    void compute(int n_threads,
+                 DiffusionParams diffusion_params,
+                 struct ggml_tensor** output     = NULL,
+                 struct ggml_context* output_ctx = NULL) {
+        return unet.compute(n_threads,
+                            diffusion_params.x,
+                            diffusion_params.timesteps,
+                            diffusion_params.context,
+                            diffusion_params.c_concat,
+                            diffusion_params.y,
+                            diffusion_params.num_video_frames,
+                            diffusion_params.controls,
+                            diffusion_params.control_strength, output, output_ctx);
+    }
+};
+
+struct MMDiTModel : public DiffusionModel {
+    MMDiTRunner mmdit;
+
+    MMDiTModel(ggml_backend_t backend,
+               bool offload_params_to_cpu,
+               bool flash_attn                     = false,
+               const String2GGMLType& tensor_types = {})
+        : mmdit(backend, offload_params_to_cpu, flash_attn, tensor_types, "model.diffusion_model") {
+    }
+
+    std::string get_desc() {
+        return mmdit.get_desc();
+    }
+
+    void alloc_params_buffer() {
+        mmdit.alloc_params_buffer();
+    }
+
+    void free_params_buffer() {
+        mmdit.free_params_buffer();
+    }
+
+    void free_compute_buffer() {
+        mmdit.free_compute_buffer();
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
+        mmdit.get_param_tensors(tensors, "model.diffusion_model");
+    }
+
+    size_t get_params_buffer_size() {
+        return mmdit.get_params_buffer_size();
+    }
+
+    int64_t get_adm_in_channels() {
+        return 768 + 1280;
+    }
+
+    void compute(int n_threads,
+                 DiffusionParams diffusion_params,
+                 struct ggml_tensor** output     = NULL,
+                 struct ggml_context* output_ctx = NULL) {
+        return mmdit.compute(n_threads,
+                             diffusion_params.x,
+                             diffusion_params.timesteps,
+                             diffusion_params.context,
+                             diffusion_params.y,
+                             output,
+                             output_ctx,
+                             diffusion_params.skip_layers);
+    }
+};
+
+struct FluxModel : public DiffusionModel {
+    Flux::FluxRunner flux;
+
+    FluxModel(ggml_backend_t backend,
+              bool offload_params_to_cpu,
+              const String2GGMLType& tensor_types = {},
+              SDVersion version                   = VERSION_FLUX,
+              bool flash_attn                     = false,
+              bool use_mask                       = false)
+        : flux(backend, offload_params_to_cpu, tensor_types, "model.diffusion_model", version, flash_attn, use_mask) {
+    }
+
+    std::string get_desc() {
+        return flux.get_desc();
+    }
+
+    void alloc_params_buffer() {
+        flux.alloc_params_buffer();
+    }
+
+    void free_params_buffer() {
+        flux.free_params_buffer();
+    }
+
+    void free_compute_buffer() {
+        flux.free_compute_buffer();
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
+        flux.get_param_tensors(tensors, "model.diffusion_model");
+    }
+
+    size_t get_params_buffer_size() {
+        return flux.get_params_buffer_size();
+    }
+
+    int64_t get_adm_in_channels() {
+        return 768;
+    }
+
+    void compute(int n_threads,
+                 DiffusionParams diffusion_params,
+                 struct ggml_tensor** output     = NULL,
+                 struct ggml_context* output_ctx = NULL) {
+        return flux.compute(n_threads,
+                            diffusion_params.x,
+                            diffusion_params.timesteps,
+                            diffusion_params.context,
+                            diffusion_params.c_concat,
+                            diffusion_params.y,
+                            diffusion_params.guidance,
+                            diffusion_params.ref_latents,
+                            diffusion_params.increase_ref_index,
+                            output,
+                            output_ctx,
+                            diffusion_params.skip_layers);
+    }
+};
+
+struct WanModel : public DiffusionModel {
+    std::string prefix;
+    WAN::WanRunner wan;
+
+    WanModel(ggml_backend_t backend,
+             bool offload_params_to_cpu,
+             const String2GGMLType& tensor_types = {},
+             const std::string prefix            = "model.diffusion_model",
+             SDVersion version                   = VERSION_WAN2,
+             bool flash_attn                     = false)
+        : prefix(prefix), wan(backend, offload_params_to_cpu, tensor_types, prefix, version, flash_attn) {
+    }
+
+    std::string get_desc() {
+        return wan.get_desc();
+    }
+
+    void alloc_params_buffer() {
+        wan.alloc_params_buffer();
+    }
+
+    void free_params_buffer() {
+        wan.free_params_buffer();
+    }
+
+    void free_compute_buffer() {
+        wan.free_compute_buffer();
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
+        wan.get_param_tensors(tensors, prefix);
+    }
+
+    size_t get_params_buffer_size() {
+        return wan.get_params_buffer_size();
+    }
+
+    int64_t get_adm_in_channels() {
+        return 768;
+    }
+
+    void compute(int n_threads,
+                 DiffusionParams diffusion_params,
+                 struct ggml_tensor** output     = NULL,
+                 struct ggml_context* output_ctx = NULL) {
+        return wan.compute(n_threads,
+                           diffusion_params.x,
+                           diffusion_params.timesteps,
+                           diffusion_params.context,
+                           diffusion_params.y,
+                           diffusion_params.c_concat,
+                           NULL,
+                           diffusion_params.vace_context,
+                           diffusion_params.vace_strength,
+                           output,
+                           output_ctx);
+    }
+};
+
+#endif
@@ -1,21 +0,0 @@
-# How to Use
-
-## Download weights
-
- Download Anima
-    - safetensors: https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/diffusion_models
-    - gguf: https://huggingface.co/Bedovyy/Anima-GGUF/tree/main
-    - gguf Anima2: https://huggingface.co/JusteLeo/Anima2-GGUF/tree/main
- Download vae
-    - safetensors: https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/vae
- Download Qwen3-0.6B-Base
-    - safetensors: https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/text_encoders
-    - gguf: https://huggingface.co/mradermacher/Qwen3-0.6B-Base-GGUF/tree/main
-
-## Examples
-
-```sh
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\anima-preview.safetensors --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_06b_base.safetensors  -p "a lovely cat holding a sign says 'anima.cpp'" --cfg-scale 6.0 --sampling-method euler -v --offload-to-cpu --diffusion-fa
-```
-
-<img alt="anima image example" src="../assets/anima/example.png" />
@@ -1,122 +0,0 @@
-# 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.
@@ -1,198 +0,0 @@
-# Build from scratch
-
-## Get the Code
-
-```
-git clone --recursive https://github.com/leejet/stable-diffusion.cpp
-cd stable-diffusion.cpp
-```
-
- If you have already cloned the repository, you can use the following command to update the repository to the latest code.
-
-```
-cd stable-diffusion.cpp
-git pull origin master
-git submodule init
-git submodule update
-```
-
-## WebP and WebM Support in Examples
-
-The example applications (`examples/cli` and `examples/server`) use `libwebp` to support WebP image I/O, and `examples/cli` can also use `libwebm` for `.webm` video output. Both are enabled by default. WebM output currently reuses `libwebp` to encode each frame as VP8 before muxing with `libwebm`.
-
-If you do not want WebP/WebM support, you can disable them at configure time:
-
-```shell
-mkdir build && cd build
-cmake .. -DSD_WEBP=OFF -DSD_WEBM=OFF
-cmake --build . --config Release
-```
-
-If the submodules are not available, you can also link against system packages instead:
-
-```shell
-mkdir build && cd build
-cmake .. -DSD_USE_SYSTEM_WEBP=ON -DSD_USE_SYSTEM_WEBM=ON
-cmake --build . --config Release
-```
-
-## Build (CPU only)
-
-If you don't have a GPU or CUDA installed, you can build a CPU-only version.
-
-```shell
-mkdir build && cd build
-cmake ..
-cmake --build . --config Release
-```
-
-## Build with OpenBLAS
-
-```shell
-mkdir build && cd build
-cmake .. -DGGML_OPENBLAS=ON
-cmake --build . --config Release
-```
-
-## Build with CUDA
-
-This provides GPU acceleration using NVIDIA GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads). Recommended to have at least 4 GB of VRAM.
-
-```shell
-mkdir build && cd build
-cmake .. -DSD_CUDA=ON
-cmake --build . --config Release
-```
-
-## Build with HipBLAS
-
-This provides GPU acceleration using AMD GPU. Make sure to have the ROCm toolkit installed.
-To build for another GPU architecture than installed in your system, set `$GFX_NAME` manually to the desired architecture (replace first command). This is also necessary if your GPU is not officially supported by ROCm, for example you have to set `$GFX_NAME` manually to `gfx1030` for consumer RDNA2 cards.
-
-Windows User Refer to [docs/hipBLAS_on_Windows.md](docs%2FhipBLAS_on_Windows.md) for a comprehensive guide.
-
-```shell
-mkdir build && cd build
-if command -v rocminfo; then export GFX_NAME=$(rocminfo | awk '/ *Name: +gfx[1-9]/ {print $2; exit}'); else echo "rocminfo missing!"; fi
-if [ -z "${GFX_NAME}" ]; then echo "Error: Couldn't detect GPU!"; else echo "Building for GPU: ${GFX_NAME}"; fi
-cmake .. -G "Ninja" -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DSD_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DGPU_TARGETS=$GFX_NAME -DAMDGPU_TARGETS=$GFX_NAME -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON -DCMAKE_POSITION_INDEPENDENT_CODE=ON
-cmake --build . --config Release
-```
-
-## Build with MUSA
-
-This provides GPU acceleration using Moore Threads GPU. Make sure to have the MUSA toolkit installed.
-
-```shell
-mkdir build && cd build
-cmake .. -DCMAKE_C_COMPILER=/usr/local/musa/bin/clang -DCMAKE_CXX_COMPILER=/usr/local/musa/bin/clang++ -DSD_MUSA=ON -DCMAKE_BUILD_TYPE=Release
-cmake --build . --config Release
-```
-
-## Build with Metal
-
-Using Metal makes the computation run on the GPU. Currently, there are some issues with Metal when performing operations on very large matrices, making it highly inefficient at the moment. Performance improvements are expected in the near future.
-
-```shell
-mkdir build && cd build
-cmake .. -DSD_METAL=ON
-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
-cmake .. -DSD_VULKAN=ON
-cmake --build . --config Release
-```
-
-## Build with OpenCL (for Adreno GPU)
-
-Currently, it supports only Adreno GPUs and is primarily optimized for Q4_0 type
-
-To build for Windows ARM please refers to [Windows 11 Arm64](https://github.com/ggml-org/llama.cpp/blob/master/docs/backend/OPENCL.md#windows-11-arm64)
-
-Building for Android:
-
-  Android NDK:
-       Download and install the Android NDK from the [official Android developer site](https://developer.android.com/ndk/downloads).
-
-Setup OpenCL Dependencies for NDK:
-
-You need to provide OpenCL headers and the ICD loader library to your NDK sysroot.
-
-*   OpenCL Headers:
-    ```bash
-    # In a temporary working directory
-    git clone https://github.com/KhronosGroup/OpenCL-Headers
-    cd OpenCL-Headers
-    # Replace <YOUR_NDK_PATH> with your actual NDK installation path
-    # e.g., cp -r CL /path/to/android-ndk-r26c/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include
-    sudo cp -r CL <YOUR_NDK_PATH>/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include
-    cd ..
-    ```
-
-*   OpenCL ICD Loader:
-    ```shell
-    # In the same temporary working directory
-    git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader
-    cd OpenCL-ICD-Loader
-    mkdir build_ndk && cd build_ndk
-
-    # Replace <YOUR_NDK_PATH> in the CMAKE_TOOLCHAIN_FILE and OPENCL_ICD_LOADER_HEADERS_DIR
-    cmake .. -G Ninja -DCMAKE_BUILD_TYPE=Release \
-      -DCMAKE_TOOLCHAIN_FILE=<YOUR_NDK_PATH>/build/cmake/android.toolchain.cmake \
-      -DOPENCL_ICD_LOADER_HEADERS_DIR=<YOUR_NDK_PATH>/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include \
-      -DANDROID_ABI=arm64-v8a \
-      -DANDROID_PLATFORM=24 \
-      -DANDROID_STL=c++_shared
-
-    ninja
-    # Replace <YOUR_NDK_PATH>
-    # e.g., cp libOpenCL.so /path/to/android-ndk-r26c/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android
-    sudo cp libOpenCL.so <YOUR_NDK_PATH>/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android
-    cd ../..
-    ```
-
-Build `stable-diffusion.cpp` for Android with OpenCL:
-
-```shell
-mkdir build-android && cd build-android
-
-# Replace <YOUR_NDK_PATH> with your actual NDK installation path
-# e.g., -DCMAKE_TOOLCHAIN_FILE=/path/to/android-ndk-r26c/build/cmake/android.toolchain.cmake
-cmake .. -G Ninja \
-  -DCMAKE_TOOLCHAIN_FILE=<YOUR_NDK_PATH>/build/cmake/android.toolchain.cmake \
-  -DANDROID_ABI=arm64-v8a \
-  -DANDROID_PLATFORM=android-28 \
-  -DGGML_OPENMP=OFF \
-  -DSD_OPENCL=ON
-
-ninja
-```
-*(Note: Don't forget to include `LD_LIBRARY_PATH=/vendor/lib64` in your command line before running the binary)*
-
-## Build with SYCL
-
-Using SYCL makes the computation run on the Intel GPU. Please make sure you have installed the related driver and [Intel® oneAPI Base toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html) before start. More details and steps can refer to [llama.cpp SYCL backend](https://github.com/ggml-org/llama.cpp/blob/master/docs/backend/SYCL.md#linux).
-
-```shell
-# Export relevant ENV variables
-source /opt/intel/oneapi/setvars.sh
-
-# Option 1: Use FP32 (recommended for better performance in most cases)
-cmake .. -DSD_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
-
-# Option 2: Use FP16
-cmake .. -DSD_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON
-
-cmake --build . --config Release
-```
@@ -1,139 +0,0 @@
-## Caching
-
-Caching methods accelerate diffusion inference by reusing intermediate computations when changes between steps are small.
-
-### Cache Modes
-
-| Mode | Target | Description |
-|------|--------|-------------|
-| `ucache` | UNET models | Condition-level caching with error tracking |
-| `easycache` | DiT models | Condition-level cache |
-| `dbcache` | DiT models | Block-level L1 residual threshold |
-| `taylorseer` | DiT models | Taylor series approximation |
-| `cache-dit` | DiT models | Combined DBCache + TaylorSeer |
-| `spectrum` | UNET and DiT models | Chebyshev + Taylor output forecasting |
-
-### UCache (UNET Models)
-
-UCache caches the residual difference (output - input) and reuses it when input changes are below threshold.
-
-```bash
-sd-cli -m model.safetensors -p "a cat" --cache-mode ucache --cache-option "threshold=1.5"
-```
-
-#### Parameters
-
-| Parameter | Description | Default |
-|-----------|-------------|---------|
-| `threshold` | Error threshold for reuse decision | 1.0 |
-| `start` | Start caching at this percent of steps | 0.15 |
-| `end` | Stop caching at this percent of steps | 0.95 |
-| `decay` | Error decay rate (0-1) | 1.0 |
-| `relative` | Scale threshold by output norm (0/1) | 1 |
-| `reset` | Reset error after computing (0/1) | 1 |
-
-#### Reset Parameter
-
-The `reset` parameter controls error accumulation behavior:
-
- `reset=1` (default): Resets accumulated error after each computed step. More aggressive caching, works well with most samplers.
- `reset=0`: Keeps error accumulated. More conservative, recommended for `euler_a` sampler.
-
-### EasyCache (DiT Models)
-
-Condition-level caching for DiT models. Caches and reuses outputs when input changes are below threshold.
-
-```bash
--cache-mode easycache --cache-option "threshold=0.3"
-```
-
-#### Parameters
-
-| Parameter | Description | Default |
-|-----------|-------------|---------|
-| `threshold` | Input change threshold for reuse | 0.2 |
-| `start` | Start caching at this percent of steps | 0.15 |
-| `end` | Stop caching at this percent of steps | 0.95 |
-
-### Cache-DIT (DiT Models)
-
-For DiT models like FLUX and QWEN, use block-level caching modes.
-
-#### DBCache
-
-Caches blocks based on L1 residual difference threshold:
-
-```bash
--cache-mode dbcache --cache-option "threshold=0.25,warmup=4"
-```
-
-#### TaylorSeer
-
-Uses Taylor series approximation to predict block outputs:
-
-```bash
--cache-mode taylorseer
-```
-
-#### Cache-DIT (Combined)
-
-Combines DBCache and TaylorSeer:
-
-```bash
--cache-mode cache-dit
-```
-
-#### Parameters
-
-| Parameter | Description | Default |
-|-----------|-------------|---------|
-| `Fn` | Front blocks to always compute | 8 |
-| `Bn` | Back blocks to always compute | 0 |
-| `threshold` | L1 residual difference threshold | 0.08 |
-| `warmup` | Steps before caching starts | 8 |
-
-#### SCM Options
-
-Steps Computation Mask controls which steps can be cached:
-
-```bash
--scm-mask "1,1,1,1,0,0,1,0,0,0,1,0,0,0,1,0,0,0,1,1"
-```
-
-Mask values: `1` = compute, `0` = can cache.
-
-| Policy | Description |
-|--------|-------------|
-| `dynamic` | Check threshold before caching |
-| `static` | Always cache on cacheable steps |
-
-```bash
--scm-policy dynamic
-```
-
-### Spectrum (UNET and DiT Models)
-
-Spectrum uses Chebyshev polynomial fitting blended with Taylor extrapolation to predict denoised outputs, skipping entire forward passes. Based on the paper [Spectrum: Adaptive Spectral Feature Forecasting for Efficient Diffusion Sampling](https://github.com/tingyu215/Spectrum).
-
-```bash
-sd-cli -m model.safetensors -p "a cat" --cache-mode spectrum
-```
-
-#### Parameters
-
-| Parameter | Description | Default |
-|-----------|-------------|---------|
-| `w` | Chebyshev vs Taylor blend weight (0=Taylor, 1=Chebyshev) | 0.40 |
-| `m` | Chebyshev polynomial degree | 3 |
-| `lam` | Ridge regression regularization | 1.0 |
-| `window` | Initial window size (compute every N steps) | 2 |
-| `flex` | Window growth per computed step after warmup | 0.50 |
-| `warmup` | Steps to always compute before caching starts | 4 |
-| `stop` | Stop caching at this fraction of total steps | 0.9 |
-
-### Performance Tips
-
- Start with default thresholds and adjust based on output quality
- Lower threshold = better quality, less speedup
- Higher threshold = more speedup, potential quality loss
- More steps generally means more caching opportunities
@@ -15,7 +15,7 @@ You can run Chroma using stable-diffusion.cpp with a GPU that has 6GB or even 4G
 You can download the preconverted gguf weights from [silveroxides/Chroma-GGUF](https://huggingface.co/silveroxides/Chroma-GGUF), this way you don't have to do the conversion yourself.

 ```
-.\bin\Release\sd-cli.exe -M convert -m ..\..\ComfyUI\models\unet\chroma-unlocked-v40.safetensors -o ..\models\chroma-unlocked-v40-q8_0.gguf -v --type q8_0
+.\bin\Release\sd.exe -M convert -m ..\..\ComfyUI\models\unet\chroma-unlocked-v40.safetensors -o ..\models\chroma-unlocked-v40-q8_0.gguf -v --type q8_0
 ```

 ## Run
@@ -24,7 +24,7 @@ You can download the preconverted gguf weights from [silveroxides/Chroma-GGUF](h
 For example:

 ```
- .\bin\Release\sd-cli.exe --diffusion-model  ..\models\chroma-unlocked-v40-q8_0.gguf --vae ..\models\ae.sft --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'chroma.cpp'" --cfg-scale 4.0 --sampling-method euler -v --chroma-disable-dit-mask --clip-on-cpu
+ .\bin\Release\sd.exe --diffusion-model  ..\models\chroma-unlocked-v40-q8_0.gguf --vae ..\models\ae.sft --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'chroma.cpp'" --cfg-scale 4.0 --sampling-method euler -v --chroma-disable-dit-mask --clip-on-cpu
 ```

 ![](../assets/flux/chroma_v40.png)
@@ -1,21 +0,0 @@
-# How to Use
-
-## Download weights
-
- Download Chroma1-Radiance
-    - safetensors: https://huggingface.co/lodestones/Chroma1-Radiance/tree/main
-    - gguf: https://huggingface.co/silveroxides/Chroma1-Radiance-GGUF/tree/main
-
- Download t5xxl
-    - safetensors: https://huggingface.co/comfyanonymous/flux_text_encoders/blob/main/t5xxl_fp16.safetensors
-
-## Examples
-
-```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Chroma1-Radiance-v0.4-Q8_0.gguf --t5xxl ..\..\ComfyUI\models\clip\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'chroma  radiance cpp'" --cfg-scale 4.0 --sampling-method euler -v
-```
-
-<img alt="Chroma1-Radiance" src="../assets/flux/chroma1-radiance.png" />
-
-
-
@@ -1,118 +0,0 @@
-# Running distilled models: SSD1B, Vega and SDx.x with tiny U-Nets
-
-## Preface 
-
-These models feature a reduced U-Net architecture. Unlike standard SDXL models, the SSD-1B and Vega U-Net contains only one middle block and fewer attention layers in its up- and down-blocks, resulting in significantly smaller file sizes. Using these models can reduce inference time by more than 33%. For more details, refer to Segmind's paper: https://arxiv.org/abs/2401.02677v1.
-Similarly, SD1.x- and SD2.x-style models with a tiny U-Net consist of only 6 U-Net blocks, leading to very small files and time savings of up to 50%. For more information, see the paper: https://arxiv.org/pdf/2305.15798.pdf.
-
-## SSD1B
-
-Note that not all of these models follow the standard parameter naming conventions. However, several useful SSD-1B models are available online, such as:
-
- * https://huggingface.co/segmind/SSD-1B/resolve/main/SSD-1B-A1111.safetensors
- * https://huggingface.co/hassenhamdi/SSD-1B-fp8_e4m3fn/resolve/main/SSD-1B_fp8_e4m3fn.safetensors
-
-Useful LoRAs are also available:
-
- * https://huggingface.co/seungminh/lora-swarovski-SSD-1B/resolve/main/pytorch_lora_weights.safetensors
- * https://huggingface.co/kylielee505/mylcmlorassd/resolve/main/pytorch_lora_weights.safetensors
-
-## Vega
-
-Segmind's Vega model is available online here:
-
- * https://huggingface.co/segmind/Segmind-Vega/resolve/main/segmind-vega.safetensors
- 
-VegaRT is an example for an LCM-LoRA:
-
- * https://huggingface.co/segmind/Segmind-VegaRT/resolve/main/pytorch_lora_weights.safetensors
-
-Both files can be used out-of-the-box, unlike the models described in next sections.
-
-
-## SD1.x, SD2.x with tiny U-Nets
-
-These models require conversion before use. You will need a Python script provided by the diffusers team, available on GitHub:
-
- * https://raw.githubusercontent.com/huggingface/diffusers/refs/heads/main/scripts/convert_diffusers_to_original_stable_diffusion.py
-
-### SD2.x
-
-NotaAI provides the following model online:
-
-* https://huggingface.co/nota-ai/bk-sdm-v2-tiny
-
-Creating a .safetensors file involves two steps. First, run this short Python script to download the model from Hugging Face:
-
-```python
-from diffusers import StableDiffusionPipeline
-pipe = StableDiffusionPipeline.from_pretrained("nota-ai/bk-sdm-v2-tiny",cache_dir="./")
-```
-
-Second, create the .safetensors file by running:
-
-```bash
-python convert_diffusers_to_original_stable_diffusion.py \
-      --model_path  models--nota-ai--bk-sdm-v2-tiny/snapshots/68277af553777858cd47e133f92e4db47321bc74 \
-      --checkpoint_path bk-sdm-v2-tiny.safetensors --half --use_safetensors
-```
-
-This will generate the **file bk-sdm-v2-tiny.safetensors**, which is now ready for use with sd.cpp.
-
-### SD1.x
-
-Several Tiny SD 1.x models are available online, such as:
-
- * https://huggingface.co/segmind/tiny-sd
- * https://huggingface.co/segmind/portrait-finetuned
- * https://huggingface.co/nota-ai/bk-sdm-tiny
-
-These models also require conversion, partly because some tensors are stored in a non-contiguous manner. To create a usable checkpoint file, follow these simple steps:
-Download and prepare the model using Python: 
-
-##### Download the model using Python on your computer, for example this way:
-
-```python
-import torch
-from diffusers import StableDiffusionPipeline
-pipe = StableDiffusionPipeline.from_pretrained("segmind/tiny-sd")
-unet=pipe.unet
-for param in unet.parameters():
-    param.data = param.data.contiguous()     # <- important here
-pipe.save_pretrained("segmindtiny-sd", safe_serialization=True)
-```
-
-##### Run the conversion script:
-
-```bash
-python convert_diffusers_to_original_stable_diffusion.py \
-      --model_path  ./segmindtiny-sd \
-      --checkpoint_path ./segmind_tiny-sd.safetensors  --half --use_safetensors
-```
-
-The file segmind_tiny-sd.safetensors will be generated and is now ready for use with sd.cpp. You can follow a similar process for the other models mentioned above.
-
-
-### SDXS-512-DreamShaper
-
-Another very tiny and **incredibly fast**  model is SDXS by IDKiro et al.  The authors refer to it as *"Real-Time One-Step Latent Diffusion Models with Image Conditions"*. For details read the paper: https://arxiv.org/pdf/2403.16627 . Once again the authors removed some more blocks of U-Net part and unlike other SD1 models they use an adjusted _AutoEncoderTiny_ instead of default _AutoEncoderKL_ for the VAE part.
-##### Some ready-to-run SDXS-512 model files are available online, such as:
-
-* https://huggingface.co/akleine/sdxs-512
-* https://huggingface.co/concedo/sdxs-512-tinySDdistilled-GGUF
-
-##### Run the model as follows:
-```bash
-~/stable-diffusion.cpp/build/bin/sd-cli -m sdxs.safetensors -p "portrait of a lovely cat" \
-  --cfg-scale 1 --steps 1
-```
-Both options: ``` --cfg-scale 1 ``` and  ``` --steps 1 ``` are mandatory here.
-
-### SDXS-512-0.9
-
-Even though the name "SDXS-512-0.9" is similar to "SDXS-512-DreamShaper", it is *completely different* but also **incredibly fast**. Sometimes it is preferred, so try it yourself.
-##### Download a ready-to-run file from here:
-
-* https://huggingface.co/akleine/sdxs-09
-
-For the use of this model, both options ``` --cfg-scale 1 ``` and ``` --steps 1 ``` are again absolutely necessary.
@@ -1,39 +1,15 @@
-# Docker
+## Docker

-## Run CLI
-
-```shell
-docker run --rm -v /path/to/models:/models -v /path/to/output/:/output ghcr.io/leejet/stable-diffusion.cpp:master [args...]
-# For example
-# docker run --rm -v ./models:/models -v ./build:/output ghcr.io/leejet/stable-diffusion.cpp:master -m /models/sd-v1-4.ckpt -p "a lovely cat" -v -o /output/output.png
-```
-
-## Run server
-
-```shell
-docker run --rm --init -v /path/to/models:/models -v /path/to/output/:/output -p "1234:1234" --entrypoint "/sd-server" ghcr.io/leejet/stable-diffusion.cpp:master [args...]
-# For example
-# docker run --rm --init -v ./models:/models -v ./build:/output -p "1234:1234" --entrypoint "/sd-server" ghcr.io/leejet/stable-diffusion.cpp:master -m /models/sd-v1-4.ckpt -p "a lovely cat" -v -o /output/output.png
-```
-
-## Building using Docker
+### Building using Docker

 ```shell
 docker build -t sd .
 ```

-## Building variants using Docker
-
-Vulkan:
+### Run

 ```shell
-docker build -f Dockerfile.vulkan -t sd .
-```
-
-## Run locally built image's CLI
-
-```shell
-docker run --rm -v /path/to/models:/models -v /path/to/output/:/output sd [args...]
+docker run -v /path/to/models:/models -v /path/to/output/:/output sd [args...]
 # For example
-# docker run --rm -v ./models:/models -v ./build:/output sd -m /models/sd-v1-4.ckpt -p "a lovely cat" -v -o /output/output.png
-```
+# docker run -v ./models:/models -v ./build:/output sd -m /models/sd-v1-4.ckpt -p "a lovely cat" -v -o /output/output.png
+```
@@ -1,35 +0,0 @@
-# How to Use
-
-You can run ERNIE-Image with stable-diffusion.cpp on GPUs with 4GB of VRAM — or even less.
-
-## Download weights
-
- Download ERNIE-Image-Turbo
-    - safetensors: https://huggingface.co/Comfy-Org/ERNIE-Image/tree/main/diffusion_models
-    - gguf: https://huggingface.co/unsloth/ERNIE-Image-Turbo-GGUF/tree/main
- Download ERNIE-Image
-    - safetensors: https://huggingface.co/Comfy-Org/ERNIE-Image/tree/main/diffusion_models
-    - gguf: https://huggingface.co/unsloth/ERNIE-Image-GGUF/tree/main
- Download vae
-    - safetensors: https://huggingface.co/Comfy-Org/ERNIE-Image/tree/main/vae
- Download ministral 3b
-    - safetensors: https://huggingface.co/Comfy-Org/ERNIE-Image/tree/main/text_encoders
-    - gguf: https://huggingface.co/unsloth/Ministral-3-3B-Instruct-2512-GGUF/tree/main
-
-## Examples
-
-### ERNIE-Image-Turbo
-
-```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\ernie-image-turbo.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\ministral-3-3b.safetensors -p "a lovely cat" --cfg-scale 1.0 --steps 8 -v --offload-to-cpu --diffusion-fa
-```
-
-<img width="256" alt="ERNIE-Image Turbo example" src="../assets/ernie_image/turbo_example.png" />
-
-### ERNIE-Image
-
-```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\ernie-image-UD-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\ministral-3-3b.safetensors -p "a lovely cat" --cfg-scale 5.0 -v --offload-to-cpu --diffusion-fa
-```
-
-<img width="256" alt="ERNIE-Image example" src="../assets/ernie_image/example.png" />
@@ -1,9 +1,9 @@
 ## Using ESRGAN to upscale results

-You can use ESRGAN—such as the model [RealESRGAN_x4plus_anime_6B.pth](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.4/RealESRGAN_x4plus_anime_6B.pth)—to upscale the generated images and improve their overall resolution and clarity.
+You can use ESRGAN to upscale the generated images. At the moment, only the [RealESRGAN_x4plus_anime_6B.pth](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.4/RealESRGAN_x4plus_anime_6B.pth) model is supported. Support for more models of this architecture will be added soon.

 - Specify the model path using the `--upscale-model PATH` parameter. example:

 ```bash
-sd-cli -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat" --upscale-model ../models/RealESRGAN_x4plus_anime_6B.pth
+sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat" --upscale-model ../models/RealESRGAN_x4plus_anime_6B.pth
 ```
@@ -15,9 +15,9 @@ You can run Flux using stable-diffusion.cpp with a GPU that has 6GB or even 4GB

 You can download the preconverted gguf weights from [FLUX.1-dev-gguf](https://huggingface.co/leejet/FLUX.1-dev-gguf) or [FLUX.1-schnell](https://huggingface.co/leejet/FLUX.1-schnell-gguf), this way you don't have to do the conversion yourself.

-For example:
+Using fp16 will lead to overflow, but ggml's support for bf16 is not yet fully developed. Therefore, we need to convert flux to gguf format here, which also saves VRAM. For example:
 ```
-.\bin\Release\sd-cli.exe -M convert -m ..\..\ComfyUI\models\unet\flux1-dev.sft -o ..\models\flux1-dev-q8_0.gguf -v --type q8_0
+.\bin\Release\sd.exe -M convert -m ..\..\ComfyUI\models\unet\flux1-dev.sft -o ..\models\flux1-dev-q8_0.gguf -v --type q8_0
 ```

 ## Run
@@ -28,7 +28,7 @@ For example:
 For example:

 ```
- .\bin\Release\sd-cli.exe --diffusion-model  ..\models\flux1-dev-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v --clip-on-cpu
+ .\bin\Release\sd.exe --diffusion-model  ..\models\flux1-dev-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v --clip-on-cpu
 ```

 Using formats of different precisions will yield results of varying quality.
@@ -44,7 +44,7 @@ Using formats of different precisions will yield results of varying quality.


 ```
- .\bin\Release\sd-cli.exe --diffusion-model  ..\models\flux1-schnell-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v --steps 4 --clip-on-cpu
+ .\bin\Release\sd.exe --diffusion-model  ..\models\flux1-schnell-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v --steps 4 --clip-on-cpu
 ```

 | q8_0  |
@@ -60,7 +60,7 @@ Since many flux LoRA training libraries have used various LoRA naming formats, i
 - LoRA model from https://huggingface.co/XLabs-AI/flux-lora-collection/tree/main (using comfy converted version!!!)

 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\models\flux1-dev-q8_0.gguf --vae ...\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'<lora:realism_lora_comfy_converted:1>" --cfg-scale 1.0 --sampling-method euler -v --lora-model-dir ../models --clip-on-cpu
+.\bin\Release\sd.exe --diffusion-model  ..\models\flux1-dev-q8_0.gguf --vae ...\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'<lora:realism_lora_comfy_converted:1>" --cfg-scale 1.0 --sampling-method euler -v --lora-model-dir ../models --clip-on-cpu
 ```

 ![output](../assets/flux/flux1-dev-q8_0%20with%20lora.png)
@@ -1,96 +0,0 @@
-# How to Use
-
-## Flux.2-dev
-
-### Download weights
-
- Download FLUX.2-dev
-    - gguf: https://huggingface.co/city96/FLUX.2-dev-gguf/tree/main
- Download vae
-    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-dev/tree/main
- Download FLUX.2-small-decoder (full_encoder_small_decoder.safetensors) as an alternative VAE option
-    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-small-decoder/tree/main
- Download Mistral-Small-3.2-24B-Instruct-2506-GGUF
-    - gguf: https://huggingface.co/unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/tree/main
-
-### Examples
-
-```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux2-dev-Q4_K_S.gguf --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf -r .\kontext_input.png -p "change 'flux.cpp' to 'flux2-dev.cpp'" --cfg-scale 1.0 --sampling-method euler -v --diffusion-fa --offload-to-cpu
-```
-
-<img alt="flux2 example" src="../assets/flux2/example.png" />
-
-## Flux.2 klein 4B / Flux.2 klein base 4B
-
-### Download weights
-
- Download FLUX.2-klein-4B
-    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-klein-4B
-    - gguf: https://huggingface.co/leejet/FLUX.2-klein-4B-GGUF/tree/main
- Download FLUX.2-klein-base-4B
-    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-klein-base-4B
-    - gguf: https://huggingface.co/leejet/FLUX.2-klein-base-4B-GGUF/tree/main
- Download vae
-    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-dev/tree/main
- Download FLUX.2-small-decoder (full_encoder_small_decoder.safetensors) as an alternative VAE option
-    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-small-decoder/tree/main
- Download Qwen3 4b
-    - safetensors: https://huggingface.co/Comfy-Org/flux2-klein-4B/tree/main/split_files/text_encoders
-    - gguf: https://huggingface.co/unsloth/Qwen3-4B-GGUF/tree/main
-
-### Examples
-
-```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-4b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_4b.safetensors -p "a lovely cat" --cfg-scale 1.0 --steps 4 -v --offload-to-cpu --diffusion-fa
-```
-
-<img alt="flux2-klein-4b" src="../assets/flux2/flux2-klein-4b.png" />
-
-```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-4b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_4b.safetensors -r .\kontext_input.png -p "change 'flux.cpp' to 'klein.cpp'" --cfg-scale 1.0 --sampling-method euler -v --diffusion-fa --offload-to-cpu --steps 4
-```
-
-<img alt="flux2-klein-4b-edit" src="../assets/flux2/flux2-klein-4b-edit.png" />
-
-```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-base-4b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_4b.safetensors -p "a lovely cat" --cfg-scale 4.0 --steps 20 -v --offload-to-cpu --diffusion-fa
-```
-
-<img alt="flux2-klein-base-4b" src="../assets/flux2/flux2-klein-base-4b.png" />
-
-## Flux.2 klein 9B / Flux.2 klein base 9B
-
-### Download weights
-
- Download FLUX.2-klein-9B
-    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-klein-9B
-    - gguf: https://huggingface.co/leejet/FLUX.2-klein-9B-GGUF/tree/main
- Download FLUX.2-klein-base-9B
-    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-klein-base-9B
-    - gguf: https://huggingface.co/leejet/FLUX.2-klein-base-9B-GGUF/tree/main
- Download vae
-    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-dev/tree/main
- Download Qwen3 8B
-    - safetensors: https://huggingface.co/Comfy-Org/flux2-klein-9B/tree/main/split_files/text_encoders
-    - gguf: https://huggingface.co/unsloth/Qwen3-8B-GGUF/tree/main
-
-### Examples
-
-```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-9b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_8b.safetensors -p "a lovely cat" --cfg-scale 1.0 --steps 4 -v --offload-to-cpu --diffusion-fa
-```
-
-<img alt="flux2-klein-9b" src="../assets/flux2/flux2-klein-9b.png" />
-
-```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-9b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_8b.safetensors -r .\kontext_input.png -p "change 'flux.cpp' to 'klein.cpp'" --cfg-scale 1.0 --sampling-method euler -v --diffusion-fa --offload-to-cpu --steps 4
-```
-
-<img alt="flux2-klein-9b-edit" src="../assets/flux2/flux2-klein-9b-edit.png" />
-
-```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-base-9b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_8b.safetensors -p "a lovely cat" --cfg-scale 4.0 --steps 20 -v --offload-to-cpu --diffusion-fa
-```
-
-<img alt="flux2-klein-base-9b" src="../assets/flux2/flux2-klein-base-9b.png" />
@@ -1,20 +0,0 @@
-# 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\7.1.1\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\5.5\bin`

 This is what I use to set the clang:
 ```Commandline
-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
+set CC=C:\Program Files\AMD\ROCm\5.5\bin\clang.exe
+set CXX=C:\Program Files\AMD\ROCm\5.5\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=gfx1150;gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032`
+`-G "Ninja"`, `-DCMAKE_C_COMPILER=clang`, `-DCMAKE_CXX_COMPILER=clang++`, `-DAMDGPU_TARGETS=gfx1100`

 >**Notice**: check the `clang` and `clang++` information:
 ```Commandline
@@ -59,30 +59,27 @@ 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\7.1.1\bin
+InstalledDir: C:\Program Files\AMD\ROCm\5.5\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\7.1.1\bin
+InstalledDir: C:\Program Files\AMD\ROCm\5.5\bin
 ```

->**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)
+>**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)

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

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

-If everything went OK, `build\bin\sd-cli.exe` file should appear.
+If everything went OK, `build\bin\sd.exe` file should appear.
@@ -16,7 +16,7 @@ You can run Kontext using stable-diffusion.cpp with a GPU that has 6GB or even 4
 You can download the preconverted gguf weights from [FLUX.1-Kontext-dev-GGUF](https://huggingface.co/QuantStack/FLUX.1-Kontext-dev-GGUF), this way you don't have to do the conversion yourself.

 ```
-.\bin\Release\sd-cli.exe -M convert -m ..\..\ComfyUI\models\unet\flux1-kontext-dev.safetensors -o ..\models\flux1-kontext-dev-q8_0.gguf -v --type q8_0
+.\bin\Release\sd.exe -M convert -m ..\..\ComfyUI\models\unet\flux1-kontext-dev.safetensors -o ..\models\flux1-kontext-dev-q8_0.gguf -v --type q8_0
 ```

 ## Run
@@ -27,7 +27,7 @@ You can download the preconverted gguf weights from [FLUX.1-Kontext-dev-GGUF](ht
 For example:

 ```
- .\bin\Release\sd-cli.exe -r .\flux1-dev-q8_0.png --diffusion-model  ..\models\flux1-kontext-dev-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors -p "change 'flux.cpp' to 'kontext.cpp'" --cfg-scale 1.0 --sampling-method euler -v --clip-on-cpu
+ .\bin\Release\sd.exe -r .\flux1-dev-q8_0.png --diffusion-model  ..\models\flux1-kontext-dev-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors -p "change 'flux.cpp' to 'kontext.cpp'" --cfg-scale 1.0 --sampling-method euler -v --clip-on-cpu
 ```


@@ -7,7 +7,7 @@
 Here's a simple example:

 ```
-./bin/sd-cli -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat<lora:lcm-lora-sdv1-5:1>" --steps 4 --lora-model-dir ../models -v --cfg-scale 1
+./bin/sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat<lora:lcm-lora-sdv1-5:1>" --steps 4 --lora-model-dir ../models -v --cfg-scale 1
 ```

 | without LCM-LoRA (--cfg-scale 7)  | with LCM-LoRA (--cfg-scale 1)  |
@@ -7,20 +7,33 @@
 Here's a simple example:

 ```
-./bin/sd-cli -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat<lora:marblesh:1>" --lora-model-dir ../models
+./bin/sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat<lora:marblesh:1>" --lora-model-dir ../models
 ```

 `../models/marblesh.safetensors` or `../models/marblesh.ckpt` will be applied to the model

-# Lora Apply Mode
+# Support matrix

-There are two ways to apply LoRA: **immediately** and **at_runtime**. You can specify it using the `--lora-apply-mode` parameter.
+> ℹ️ CUDA `get_rows` support is defined here:  
+> [ggml-org/ggml/src/ggml-cuda/getrows.cu#L156](https://github.com/ggml-org/ggml/blob/7dee1d6a1e7611f238d09be96738388da97c88ed/src/ggml-cuda/getrows.cu#L156)  
+> Currently only the basic types + Q4/Q5/Q8 are implemented. K-quants are **not** supported.

-By default, the mode is selected automatically:
-
-* If the model weights contain any quantized parameters, the **at_runtime** mode is used;
-* Otherwise, the **immediately** mode is used.
-
-The **immediately** mode may have precision and compatibility issues with quantized parameters, but it usually offers faster inference speed and, in some cases, lower memory usage.
-In contrast, the **at_runtime** mode provides better compatibility and higher precision, but inference may be slower and memory usage may be higher in some cases.
+NOTE: The other backends may have different support.

+| Quant / Type | CUDA |
+|--------------|------|
+| F32          | ✔️   |
+| F16          | ✔️   |
+| BF16         | ✔️   |
+| I32          | ✔️   |
+| Q4_0         | ✔️   |
+| Q4_1         | ✔️   |
+| Q5_0         | ✔️   |
+| Q5_1         | ✔️   |
+| Q8_0         | ✔️   |
+| Q2_K         | ❌   |
+| Q3_K         | ❌   |
+| Q4_K         | ❌   |
+| Q5_K         | ❌   |
+| Q6_K         | ❌   |
+| Q8_K         | ❌   |
@@ -1,53 +0,0 @@
-# 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
-
-## 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>
@@ -1,19 +0,0 @@
-# How to Use
-
-## Download weights
-
- Download Ovis-Image-7B
-    - safetensors: https://huggingface.co/Comfy-Org/Ovis-Image/tree/main/split_files/diffusion_models
-    - gguf: https://huggingface.co/leejet/Ovis-Image-7B-GGUF
- Download vae
-    - safetensors: https://huggingface.co/black-forest-labs/FLUX.1-schnell/tree/main
- Download Ovis 2.5
-    - safetensors: https://huggingface.co/Comfy-Org/Ovis-Image/tree/main/split_files/text_encoders
-
-## Examples
-
-```
-.\bin\Release\sd-cli.exe --diffusion-model  ovis_image-Q4_0.gguf --vae ..\..\ComfyUI\models\vae\ae.sft  --llm ..\..\ComfyUI\models\text_encoders\ovis_2.5.safetensors -p "a lovely cat" --cfg-scale 5.0 -v --offload-to-cpu --diffusion-fa
-```
-
-<img alt="ovis image example" src="../assets/ovis_image/example.png" />
@@ -1,26 +0,0 @@
-## Use Flash Attention to save memory and improve speed.
-
-Enabling flash attention for the diffusion model reduces memory usage by varying amounts of MB.
-eg.:
- - flux 768x768 ~600mb
- - SD2 768x768 ~1400mb
-
-For most backends, it slows things down, but for cuda it generally speeds it up too.
-At the moment, it is only supported for some models and some backends (like cpu, cuda/rocm, metal).
-
-Run by adding `--diffusion-fa` to the arguments and watch for:
-```
-[INFO ] stable-diffusion.cpp:312  - Using flash attention in the diffusion model
-```
-and the compute buffer shrink in the debug log:
-```
-[DEBUG] ggml_extend.hpp:1004 - flux compute buffer size: 650.00 MB(VRAM)
-```
-
-## Offload weights to the CPU to save VRAM without reducing generation speed.
-
-Using `--offload-to-cpu` allows you to offload weights to the CPU, saving VRAM without reducing generation speed.
-
-## Use quantization to reduce memory usage.
-
-[quantization](./quantization_and_gguf.md)
@@ -27,7 +27,7 @@ If on low memory GPUs (<= 8GB), recommend running with ```--vae-on-cpu``` option
 Example:

 ```bash
-bin/sd-cli -m ../models/sdxlUnstableDiffusers_v11.safetensors  --vae ../models/sdxl_vae.safetensors --photo-maker ../models/photomaker-v1.safetensors --pm-id-images-dir ../assets/photomaker_examples/scarletthead_woman -p "a girl img, retro futurism, retro game art style but extremely beautiful, intricate details, masterpiece, best quality, space-themed, cosmic, celestial, stars, galaxies, nebulas, planets, science fiction, highly detailed" -n "realistic, photo-realistic, worst quality, greyscale, bad anatomy, bad hands, error, text" --cfg-scale 5.0  --sampling-method euler -H 1024 -W 1024 --pm-style-strength 10 --vae-on-cpu --steps 50
+bin/sd -m ../models/sdxlUnstableDiffusers_v11.safetensors  --vae ../models/sdxl_vae.safetensors --photo-maker ../models/photomaker-v1.safetensors --pm-id-images-dir ../assets/photomaker_examples/scarletthead_woman -p "a girl img, retro futurism, retro game art style but extremely beautiful, intricate details, masterpiece, best quality, space-themed, cosmic, celestial, stars, galaxies, nebulas, planets, science fiction, highly detailed" -n "realistic, photo-realistic, worst quality, greyscale, bad anatomy, bad hands, error, text" --cfg-scale 5.0  --sampling-method euler -H 1024 -W 1024 --pm-style-strength 10 --vae-on-cpu --steps 50
 ```

 ## PhotoMaker Version 2
@@ -40,7 +40,7 @@ Running PMV2 is now a two-step process:
 ```
 python face_detect.py input_image_dir
 ```
-An ```id_embeds.bin``` file will be generated in ```input_images_dir```
+An ```id_embeds.safetensors``` file will be generated in ```input_images_dir```

 **Note: this step is only needed to run once; the same ```id_embeds``` can be reused**

@@ -48,6 +48,6 @@ An ```id_embeds.bin``` file will be generated in ```input_images_dir```

  You can download ```photomaker-v2.safetensors``` from [here](https://huggingface.co/bssrdf/PhotoMakerV2)

- All the command line parameters from Version 1 remain the same for Version 2 plus one extra pointing to a valid ```id_embeds``` file:  --pm-id-embed-path [path_to__id_embeds.bin] 
+- All the command line parameters from Version 1 remain the same for Version 2


@@ -23,5 +23,5 @@ You can also convert weights in the formats `ckpt/safetensors/diffusers` to gguf
 For example:

 ```sh
-./bin/sd-cli -M convert -m ../models/v1-5-pruned-emaonly.safetensors -o  ../models/v1-5-pruned-emaonly.q8_0.gguf -v --type q8_0
+./bin/sd -M convert -m ../models/v1-5-pruned-emaonly.safetensors -o  ../models/v1-5-pruned-emaonly.q8_0.gguf -v --type q8_0
 ```
@@ -1,23 +0,0 @@
-# How to Use
-
-## Download weights
-
- Download Qwen Image
-    - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main/split_files/diffusion_models
-    - gguf: https://huggingface.co/QuantStack/Qwen-Image-GGUF/tree/main
- Download vae
-    - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main/split_files/vae
- 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
-
-## Examples
-
-```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\qwen-image-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct-Q8_0.gguf  -p '一个穿着"QWEN"标志的T恤的中国美女正拿着黑色的马克笔面相镜头微笑。她身后的玻璃板上手写体写着 “一、Qwen-Image的技术路线： 探索视觉生成基础模型的极限，开创理解与生成一体化的未来。二、Qwen-Image的模型特色：1、复杂文字渲染。支持中英渲染、自动布局； 2、精准图像编辑。支持文字编辑、物体增减、风格变换。三、Qwen-Image的未来愿景：赋能专业内容创作、助力生成式AI发展。”' --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu -H 1024 -W 1024 --diffusion-fa --flow-shift 3
-```
-
-<img alt="qwen example" src="../assets/qwen/example.png" />
-
-
-
@@ -1,48 +0,0 @@
-# How to Use
-
-## Download weights
-
- Download Qwen Image
-    - Qwen Image Edit
-        - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image-Edit_ComfyUI/tree/main/split_files/diffusion_models
-        - gguf: https://huggingface.co/QuantStack/Qwen-Image-Edit-GGUF/tree/main
-    - Qwen Image Edit 2509
-        - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image-Edit_ComfyUI/tree/main/split_files/diffusion_models
-        - gguf: https://huggingface.co/QuantStack/Qwen-Image-Edit-2509-GGUF/tree/main
-    - Qwen Image Edit 2511
-        - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image-Edit_ComfyUI/tree/main/split_files/diffusion_models
-        - gguf: https://huggingface.co/unsloth/Qwen-Image-Edit-2511-GGUF/tree/main
- Download vae
-    - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main/split_files/vae
- 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
-
-## Examples
-
-### Qwen Image Edit
-
-```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Qwen_Image_Edit-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_2.5_vl_7b.safetensors --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu --diffusion-fa --flow-shift 3 -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'edit.cpp'" --seed 1118877715456453
-```
-
-<img alt="qwen_image_edit" src="../assets/qwen/qwen_image_edit.png" />
-
-
-### Qwen Image Edit 2509
-
-```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Qwen-Image-Edit-2509-Q4_K_S.gguf --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct-Q8_0.gguf --llm_vision ..\..\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct.mmproj-Q8_0.gguf --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu --diffusion-fa --flow-shift 3 -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'Qwen Image Edit 2509'"
-```
-
-<img alt="qwen_image_edit_2509" src="../assets/qwen/qwen_image_edit_2509.png" />
-
-### Qwen Image Edit 2511
-
-To use the new Qwen Image Edit 2511 mode, the  `--qwen-image-zero-cond-t` flag must be enabled; otherwise, image editing quality will degrade significantly.
-
-```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\qwen-image-edit-2511-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_2.5_vl_7b.safetensors --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu --diffusion-fa --flow-shift 3 -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'edit.cpp'"  --qwen-image-zero-cond-t
-```
-
-<img alt="qwen_image_edit_2509" src="../assets/qwen/qwen_image_edit_2511.png" />
@@ -1,37 +0,0 @@
-## Download weights
-
- download original weights(.ckpt or .safetensors). For example
-    - Stable Diffusion v1.4 from https://huggingface.co/CompVis/stable-diffusion-v-1-4-original
-    - Stable Diffusion v1.5 from https://huggingface.co/runwayml/stable-diffusion-v1-5
-    - Stable Diffuison v2.1 from https://huggingface.co/stabilityai/stable-diffusion-2-1
-    - Stable Diffusion 3 2B from https://huggingface.co/stabilityai/stable-diffusion-3-medium
-
-### txt2img example
-
-```sh
-./bin/sd-cli -m ../models/sd-v1-4.ckpt -p "a lovely cat"
-# ./bin/sd-cli -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat"
-# ./bin/sd-cli -m ../models/sd_xl_base_1.0.safetensors --vae ../models/sdxl_vae-fp16-fix.safetensors -H 1024 -W 1024 -p "a lovely cat" -v
-# ./bin/sd-cli -m ../models/sd3_medium_incl_clips_t5xxlfp16.safetensors -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable Diffusion CPP\"' --cfg-scale 4.5 --sampling-method euler -v --clip-on-cpu
-# ./bin/sd-cli --diffusion-model  ../models/flux1-dev-q3_k.gguf --vae ../models/ae.sft --clip_l ../models/clip_l.safetensors --t5xxl ../models/t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v --clip-on-cpu
-# ./bin/sd-cli -m  ..\models\sd3.5_large.safetensors --clip_l ..\models\clip_l.safetensors --clip_g ..\models\clip_g.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable diffusion 3.5 Large\"' --cfg-scale 4.5 --sampling-method euler -v --clip-on-cpu
-```
-
-Using formats of different precisions will yield results of varying quality.
-
-| f32  | f16  |q8_0  |q5_0  |q5_1  |q4_0  |q4_1  |
-| ----  |----  |----  |----  |----  |----  |----  |
-| ![](../assets/f32.png) |![](../assets/f16.png) |![](../assets/q8_0.png) |![](../assets/q5_0.png) |![](../assets/q5_1.png) |![](../assets/q4_0.png) |![](../assets/q4_1.png) |
-
-### img2img example
-
- `./output.png` is the image generated from the above txt2img pipeline
-
-
-```
-./bin/sd-cli -m ../models/sd-v1-4.ckpt -p "cat with blue eyes" -i ./output.png -o ./img2img_output.png --strength 0.4
-```
-
-<p align="center">
-  <img src="../assets/img2img_output.png" width="256x">
-</p>
@@ -14,7 +14,7 @@
 For example:

 ```
-.\bin\Release\sd-cli.exe -m  ..\models\sd3.5_large.safetensors --clip_l ..\models\clip_l.safetensors --clip_g ..\models\clip_g.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable diffusion 3.5 Large\"' --cfg-scale 4.5 --sampling-method euler -v --clip-on-cpu
+.\bin\Release\sd.exe -m  ..\models\sd3.5_large.safetensors --clip_l ..\models\clip_l.safetensors --clip_g ..\models\clip_g.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable diffusion 3.5 Large\"' --cfg-scale 4.5 --sampling-method euler -v --clip-on-cpu
 ```

 ![](../assets/sd3.5_large.png)
@@ -7,33 +7,11 @@ You can use TAESD to accelerate the decoding of latent images by following these
 Or curl

 ```bash
-curl -L -O https://huggingface.co/madebyollin/taesd/resolve/main/diffusion_pytorch_model.safetensors
+curl -L -O https://huggingface.co/madebyollin/taesd/blob/main/diffusion_pytorch_model.safetensors
 ```

 - Specify the model path using the `--taesd PATH` parameter. example:

 ```bash
-sd-cli -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat" --taesd ../models/diffusion_pytorch_model.safetensors
-```
-
-### Qwen-Image and wan (TAEHV)
-
-sd.cpp also supports [TAEHV](https://github.com/madebyollin/taehv) (#937), which can be used for Qwen-Image and wan.
-
- For **Qwen-Image and wan2.1 and wan2.2-A14B**, download the wan2.1 tae [safetensors weights](https://github.com/madebyollin/taehv/blob/main/safetensors/taew2_1.safetensors)
-  
-  Or curl
-  
-  ```bash
-  curl -L -O https://github.com/madebyollin/taehv/raw/refs/heads/main/safetensors/taew2_1.safetensors
-  ```
-
- For **wan2.2-TI2V-5B**, use the wan2.2 tae [safetensors weights](https://github.com/madebyollin/taehv/blob/main/safetensors/taew2_2.safetensors)
-  
-  Or curl
-  
-  ```bash
-  curl -L -O https://github.com/madebyollin/taehv/raw/refs/heads/main/safetensors/taew2_2.safetensors
-  ```
-
-Then simply replace the `--vae xxx.safetensors` with `--tae xxx.safetensors` in the commands. If it still out of VRAM, add `--vae-conv-direct` to your command though might be slower.
+sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat" --taesd ../models/diffusion_pytorch_model.safetensors
+```
@@ -39,9 +39,6 @@
        - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/vae/wan_2.1_vae.safetensors
    - wan_2.2_vae (for Wan2.2 TI2V 5B only)
        - safetensors: https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/blob/main/split_files/vae/wan2.2_vae.safetensors
-
-    > Wan models vae requires really much VRAM! If you do not have enough VRAM, please try tae instead, though the results may be poorer. For tae usage, please refer to [taesd](taesd.md)
-
 - Download umt5_xxl
    - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/text_encoders/umt5_xxl_fp16.safetensors
    - gguf: https://huggingface.co/city96/umt5-xxl-encoder-gguf/tree/main
@@ -55,7 +52,7 @@
 ### Wan2.1 T2V 1.3B

 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1_t2v_1.3B_fp16.safetensors --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 33 --flow-shift 3.0
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1_t2v_1.3B_fp16.safetensors --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 33 --flow-shift 3.0
 ```

 <video src=../assets/wan/Wan2.1_1.3B_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -63,7 +60,7 @@
 ### Wan2.1 T2V 14B

 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-t2v-14b-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa  --offload-to-cpu --video-frames 33 --flow-shift 3.0
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-t2v-14b-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa  --offload-to-cpu --video-frames 33 --flow-shift 3.0
 ```

 <video src=../assets/wan/Wan2.1_14B_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -73,7 +70,7 @@
 ### Wan2.1 I2V 14B

 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-i2v-14b-480p-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf --clip_vision ..\..\ComfyUI\models\clip_vision\clip_vision_h.safetensors -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --video-frames 33 --offload-to-cpu -i ..\assets\cat_with_sd_cpp_42.png --flow-shift 3.0
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-i2v-14b-480p-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf --clip_vision ..\..\ComfyUI\models\clip_vision\clip_vision_h.safetensors -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --video-frames 33 --offload-to-cpu -i ..\assets\cat_with_sd_cpp_42.png --flow-shift 3.0
 ```

 <video src=../assets/wan/Wan2.1_14B_i2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -81,7 +78,7 @@
 ### Wan2.2 T2V A14B

 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --video-frames 33 --flow-shift 3.0
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --video-frames 33 --flow-shift 3.0
 ```

 <video src=../assets/wan/Wan2.2_14B_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -89,7 +86,7 @@
 ### Wan2.2 I2V A14B

 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-I2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-I2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --video-frames 33 --offload-to-cpu -i ..\assets\cat_with_sd_cpp_42.png --flow-shift 3.0
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-I2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-I2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --video-frames 33 --offload-to-cpu -i ..\assets\cat_with_sd_cpp_42.png --flow-shift 3.0
 ```

 <video src=../assets/wan/Wan2.2_14B_i2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -97,7 +94,7 @@
 ### Wan2.2 T2V A14B T2I

 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --flow-shift 3.0
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --flow-shift 3.0
 ```

 <img width="832" height="480" alt="Wan2 2_14B_t2i" src="../assets/wan/Wan2.2_14B_t2i.png" />
@@ -105,7 +102,7 @@
 ### Wan2.2 T2V 14B with Lora

 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat<lora:wan2.2_t2v_lightx2v_4steps_lora_v1.1_low_noise:1><lora:|high_noise|wan2.2_t2v_lightx2v_4steps_lora_v1.1_high_noise:1>" --cfg-scale 3.5 --sampling-method euler --steps 4 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 4 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --lora-model-dir ..\..\ComfyUI\models\loras --video-frames 33 --flow-shift 3.0
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat<lora:wan2.2_t2v_lightx2v_4steps_lora_v1.1_low_noise:1><lora:|high_noise|wan2.2_t2v_lightx2v_4steps_lora_v1.1_high_noise:1>" --cfg-scale 3.5 --sampling-method euler --steps 4 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 4 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --lora-model-dir ..\..\ComfyUI\models\loras --video-frames 33 --flow-shift 3.0
 ```

 <video src=../assets/wan/Wan2.2_14B_t2v_lora.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -117,7 +114,7 @@
 #### T2V

 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.2_ti2v_5B_fp16.safetensors --vae ..\..\ComfyUI\models\vae\wan2.2_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --offload-to-cpu --video-frames 33 --flow-shift 3.0
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.2_ti2v_5B_fp16.safetensors --vae ..\..\ComfyUI\models\vae\wan2.2_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --offload-to-cpu --video-frames 33 --flow-shift 3.0
 ```

 <video src=../assets/wan/Wan2.2_5B_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -125,7 +122,7 @@
 #### I2V

 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.2_ti2v_5B_fp16.safetensors --vae ..\..\ComfyUI\models\vae\wan2.2_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --offload-to-cpu --video-frames 33 -i ..\assets\cat_with_sd_cpp_42.png --flow-shift 3.0
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.2_ti2v_5B_fp16.safetensors --vae ..\..\ComfyUI\models\vae\wan2.2_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --offload-to-cpu --video-frames 33 -i ..\assets\cat_with_sd_cpp_42.png --flow-shift 3.0
 ```

 <video src=../assets/wan/Wan2.2_5B_i2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -133,7 +130,7 @@
 ### Wan2.1 FLF2V 14B

 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-flf2v-14b-720p-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf --clip_vision ..\..\ComfyUI\models\clip_vision\clip_vision_h.safetensors -p "glass flower blossom" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --video-frames 33 --offload-to-cpu --init-img ..\..\ComfyUI\input\start_image.png --end-img ..\..\ComfyUI\input\end_image.png --flow-shift 3.0
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-flf2v-14b-720p-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf --clip_vision ..\..\ComfyUI\models\clip_vision\clip_vision_h.safetensors -p "glass flower blossom" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --video-frames 33 --offload-to-cpu --init-img ..\..\ComfyUI\input\start_image.png --end-img ..\..\ComfyUI\input\end_image.png --flow-shift 3.0
 ```


@@ -142,7 +139,7 @@
 ### Wan2.2 FLF2V 14B

 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-I2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-I2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -p "glass flower blossom" -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --video-frames 33 --offload-to-cpu --init-img ..\..\ComfyUI\input\start_image.png --end-img ..\..\ComfyUI\input\end_image.png --flow-shift 3.0
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-I2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-I2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -p "glass flower blossom" -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --video-frames 33 --offload-to-cpu --init-img ..\..\ComfyUI\input\start_image.png --end-img ..\..\ComfyUI\input\end_image.png --flow-shift 3.0
 ```

 <video src=../assets/wan/Wan2.2_14B_flf2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -152,7 +149,7 @@
 #### T2V

 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 1 --offload-to-cpu
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 1 --offload-to-cpu
 ```

 <video src=../assets/wan/Wan2.1_1.3B_vace_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -161,7 +158,7 @@
 #### R2V

 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa -i ..\assets\cat_with_sd_cpp_42.png --video-frames 33 --offload-to-cpu
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa -i ..\assets\cat_with_sd_cpp_42.png --video-frames 33 --offload-to-cpu
 ```

 <video src=../assets/wan/Wan2.1_1.3B_vace_r2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -172,7 +169,7 @@
 ```
 mkdir post+depth
 ffmpeg -i ..\..\ComfyUI\input\post+depth.mp4 -qscale:v 1 -vf fps=8 post+depth\frame_%04d.jpg
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "The girl is dancing in a sea of flowers, slowly moving her hands. There is a close - up shot of her upper body. The character is surrounded by other transparent glass flowers in the style of Nicoletta Ceccoli, creating a beautiful, surreal, and emotionally expressive movie scene with a white. transparent feel and a dreamyl atmosphere." --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa -i ..\..\ComfyUI\input\dance_girl.jpg --control-video ./post+depth --video-frames 33 --offload-to-cpu
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "The girl is dancing in a sea of flowers, slowly moving her hands. There is a close - up shot of her upper body. The character is surrounded by other transparent glass flowers in the style of Nicoletta Ceccoli, creating a beautiful, surreal, and emotionally expressive movie scene with a white. transparent feel and a dreamyl atmosphere." --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa -i ..\..\ComfyUI\input\dance_girl.jpg --control-video ./post+depth --video-frames 33 --offload-to-cpu
 ```

 <video src=../assets/wan/Wan2.1_1.3B_vace_v2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -182,7 +179,7 @@ ffmpeg -i ..\..\ComfyUI\input\post+depth.mp4 -qscale:v 1 -vf fps=8 post+depth\fr
 #### T2V

 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 33 --offload-to-cpu
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 33 --offload-to-cpu
 ```

 <video src=../assets/wan/Wan2.1_14B_vace_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -191,7 +188,7 @@ ffmpeg -i ..\..\ComfyUI\input\post+depth.mp4 -qscale:v 1 -vf fps=8 post+depth\fr
 #### R2V

 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa -i ..\assets\cat_with_sd_cpp_42.png --video-frames 33 --offload-to-cpu
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa -i ..\assets\cat_with_sd_cpp_42.png --video-frames 33 --offload-to-cpu
 ```

 <video src=../assets/wan/Wan2.1_14B_vace_r2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -201,7 +198,7 @@ ffmpeg -i ..\..\ComfyUI\input\post+depth.mp4 -qscale:v 1 -vf fps=8 post+depth\fr
 #### V2V

 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "The girl is dancing in a sea of flowers, slowly moving her hands. There is a close - up shot of her upper body. The character is surrounded by other transparent glass flowers in the style of Nicoletta Ceccoli, creating a beautiful, surreal, and emotionally expressive movie scene with a white. transparent feel and a dreamyl atmosphere." --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa -i ..\..\ComfyUI\input\dance_girl.jpg --control-video ./post+depth --video-frames 33 --offload-to-cpu
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "The girl is dancing in a sea of flowers, slowly moving her hands. There is a close - up shot of her upper body. The character is surrounded by other transparent glass flowers in the style of Nicoletta Ceccoli, creating a beautiful, surreal, and emotionally expressive movie scene with a white. transparent feel and a dreamyl atmosphere." --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa -i ..\..\ComfyUI\input\dance_girl.jpg --control-video ./post+depth --video-frames 33 --offload-to-cpu
 ```

 <video src=../assets/wan/Wan2.1_14B_vace_v2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -1,41 +0,0 @@
-# How to Use
-
-You can run Z-Image with stable-diffusion.cpp on GPUs with 4GB of VRAM — or even less.
-
-## Download weights
-
- Download Z-Image-Turbo
-    - safetensors: https://huggingface.co/Comfy-Org/z_image_turbo/tree/main/split_files/diffusion_models
-    - gguf: https://huggingface.co/leejet/Z-Image-Turbo-GGUF/tree/main
- Download Z-Image
-    - safetensors: https://huggingface.co/Comfy-Org/z_image/tree/main/split_files/diffusion_models
-    - gguf: https://huggingface.co/unsloth/Z-Image-GGUF/tree/main
- Download vae
-    - safetensors: https://huggingface.co/black-forest-labs/FLUX.1-schnell/tree/main
- Download Qwen3 4b
-    - safetensors: https://huggingface.co/Comfy-Org/z_image_turbo/tree/main/split_files/text_encoders
-    - gguf: https://huggingface.co/unsloth/Qwen3-4B-Instruct-2507-GGUF/tree/main
-
-## Examples
-
-### 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 --steps 8
-```
-
-<img width="256" alt="z-image example" src="../assets/z_image/q3_K.png" />
-
-### Z-Image-Base
-
-```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\z_image_bf16.safetensors --vae ..\..\ComfyUI\models\vae\ae.sft  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_4b.safetensors -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 5.0 -v --offload-to-cpu --diffusion-fa -H 1024 -W 512
-```
-
-<img width="256" alt="z-image example" src="../assets/z_image/base_bf16.png" />
-
-## Comparison of Different Quantization Types
-
-| bf16 | q8_0 | q6_K | q5_0 | q4_K | q4_0 | q3_K | q2_K|
-|---|---|---|---|---|---|---|---|
-| <img width="256" alt="bf16" src="../assets/z_image/bf16.png" /> | <img width="256" alt="q8_0" src="../assets/z_image/q8_0.png" /> | <img width="256" alt="q6_K" src="../assets/z_image/q6_K.png" /> | <img width="256" alt="q5_0" src="../assets/z_image/q5_0.png" />  | <img width="256" alt="q4_K" src="../assets/z_image/q4_K.png" /> | <img width="256" alt="q4_0" src="../assets/z_image/q4_0.png" /> | <img width="256" alt="q3_K" src="../assets/z_image/q3_K.png" /> | <img width="256" alt="q2_K" src="../assets/z_image/q2_K.png" /> |
@@ -0,0 +1,210 @@
+#ifndef __ESRGAN_HPP__
+#define __ESRGAN_HPP__
+
+#include "ggml_extend.hpp"
+#include "model.h"
+
+/*
+    ===================================    ESRGAN  ===================================
+    References:
+    https://github.com/xinntao/Real-ESRGAN/blob/master/inference_realesrgan.py
+    https://github.com/XPixelGroup/BasicSR/blob/v1.4.2/basicsr/archs/rrdbnet_arch.py
+
+*/
+
+class ResidualDenseBlock : public GGMLBlock {
+protected:
+    int num_feat;
+    int num_grow_ch;
+
+public:
+    ResidualDenseBlock(int num_feat = 64, int num_grow_ch = 32)
+        : num_feat(num_feat), num_grow_ch(num_grow_ch) {
+        blocks["conv1"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat, num_grow_ch, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv2"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat + num_grow_ch, num_grow_ch, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv3"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat + 2 * num_grow_ch, num_grow_ch, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv4"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat + 3 * num_grow_ch, num_grow_ch, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv5"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat + 4 * num_grow_ch, num_feat, {3, 3}, {1, 1}, {1, 1}));
+    }
+
+    struct ggml_tensor* lrelu(struct ggml_context* ctx, struct ggml_tensor* x) {
+        return ggml_leaky_relu(ctx, x, 0.2f, true);
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [n, num_feat, h, w]
+        // return: [n, num_feat, h, w]
+
+        auto conv1 = std::dynamic_pointer_cast<Conv2d>(blocks["conv1"]);
+        auto conv2 = std::dynamic_pointer_cast<Conv2d>(blocks["conv2"]);
+        auto conv3 = std::dynamic_pointer_cast<Conv2d>(blocks["conv3"]);
+        auto conv4 = std::dynamic_pointer_cast<Conv2d>(blocks["conv4"]);
+        auto conv5 = std::dynamic_pointer_cast<Conv2d>(blocks["conv5"]);
+
+        auto x1    = lrelu(ctx, conv1->forward(ctx, x));
+        auto x_cat = ggml_concat(ctx, x, x1, 2);
+        auto x2    = lrelu(ctx, conv2->forward(ctx, x_cat));
+        x_cat      = ggml_concat(ctx, x_cat, x2, 2);
+        auto x3    = lrelu(ctx, conv3->forward(ctx, x_cat));
+        x_cat      = ggml_concat(ctx, x_cat, x3, 2);
+        auto x4    = lrelu(ctx, conv4->forward(ctx, x_cat));
+        x_cat      = ggml_concat(ctx, x_cat, x4, 2);
+        auto x5    = conv5->forward(ctx, x_cat);
+
+        x5 = ggml_add(ctx, ggml_scale(ctx, x5, 0.2f), x);
+        return x5;
+    }
+};
+
+class RRDB : public GGMLBlock {
+public:
+    RRDB(int num_feat, int num_grow_ch = 32) {
+        blocks["rdb1"] = std::shared_ptr<GGMLBlock>(new ResidualDenseBlock(num_feat, num_grow_ch));
+        blocks["rdb2"] = std::shared_ptr<GGMLBlock>(new ResidualDenseBlock(num_feat, num_grow_ch));
+        blocks["rdb3"] = std::shared_ptr<GGMLBlock>(new ResidualDenseBlock(num_feat, num_grow_ch));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [n, num_feat, h, w]
+        // return: [n, num_feat, h, w]
+
+        auto rdb1 = std::dynamic_pointer_cast<ResidualDenseBlock>(blocks["rdb1"]);
+        auto rdb2 = std::dynamic_pointer_cast<ResidualDenseBlock>(blocks["rdb2"]);
+        auto rdb3 = std::dynamic_pointer_cast<ResidualDenseBlock>(blocks["rdb3"]);
+
+        auto out = rdb1->forward(ctx, x);
+        out      = rdb2->forward(ctx, out);
+        out      = rdb3->forward(ctx, out);
+
+        out = ggml_add(ctx, ggml_scale(ctx, out, 0.2f), x);
+        return out;
+    }
+};
+
+class RRDBNet : public GGMLBlock {
+protected:
+    int scale       = 4;  // default RealESRGAN_x4plus_anime_6B
+    int num_block   = 6;  // default RealESRGAN_x4plus_anime_6B
+    int num_in_ch   = 3;
+    int num_out_ch  = 3;
+    int num_feat    = 64;  // default RealESRGAN_x4plus_anime_6B
+    int num_grow_ch = 32;  // default RealESRGAN_x4plus_anime_6B
+
+public:
+    RRDBNet() {
+        blocks["conv_first"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_in_ch, num_feat, {3, 3}, {1, 1}, {1, 1}));
+        for (int i = 0; i < num_block; i++) {
+            std::string name = "body." + std::to_string(i);
+            blocks[name]     = std::shared_ptr<GGMLBlock>(new RRDB(num_feat, num_grow_ch));
+        }
+        blocks["conv_body"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat, num_feat, {3, 3}, {1, 1}, {1, 1}));
+        // upsample
+        blocks["conv_up1"]  = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat, num_feat, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv_up2"]  = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat, num_feat, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv_hr"]   = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat, num_feat, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv_last"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat, num_out_ch, {3, 3}, {1, 1}, {1, 1}));
+    }
+
+    struct ggml_tensor* lrelu(struct ggml_context* ctx, struct ggml_tensor* x) {
+        return ggml_leaky_relu(ctx, x, 0.2f, true);
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [n, num_in_ch, h, w]
+        // return: [n, num_out_ch, h*4, w*4]
+        auto conv_first = std::dynamic_pointer_cast<Conv2d>(blocks["conv_first"]);
+        auto conv_body  = std::dynamic_pointer_cast<Conv2d>(blocks["conv_body"]);
+        auto conv_up1   = std::dynamic_pointer_cast<Conv2d>(blocks["conv_up1"]);
+        auto conv_up2   = std::dynamic_pointer_cast<Conv2d>(blocks["conv_up2"]);
+        auto conv_hr    = std::dynamic_pointer_cast<Conv2d>(blocks["conv_hr"]);
+        auto conv_last  = std::dynamic_pointer_cast<Conv2d>(blocks["conv_last"]);
+
+        auto feat      = conv_first->forward(ctx, x);
+        auto body_feat = feat;
+        for (int i = 0; i < num_block; i++) {
+            std::string name = "body." + std::to_string(i);
+            auto block       = std::dynamic_pointer_cast<RRDB>(blocks[name]);
+
+            body_feat = block->forward(ctx, body_feat);
+        }
+        body_feat = conv_body->forward(ctx, body_feat);
+        feat      = ggml_add(ctx, feat, body_feat);
+        // upsample
+        feat     = lrelu(ctx, conv_up1->forward(ctx, ggml_upscale(ctx, feat, 2, GGML_SCALE_MODE_NEAREST)));
+        feat     = lrelu(ctx, conv_up2->forward(ctx, ggml_upscale(ctx, feat, 2, GGML_SCALE_MODE_NEAREST)));
+        auto out = conv_last->forward(ctx, lrelu(ctx, conv_hr->forward(ctx, feat)));
+        return out;
+    }
+};
+
+struct ESRGAN : public GGMLRunner {
+    RRDBNet rrdb_net;
+    int scale     = 4;
+    int tile_size = 128;  // avoid cuda OOM for 4gb VRAM
+
+    ESRGAN(ggml_backend_t backend,
+           bool offload_params_to_cpu,
+           const String2GGMLType& tensor_types = {})
+        : GGMLRunner(backend, offload_params_to_cpu) {
+        rrdb_net.init(params_ctx, tensor_types, "");
+    }
+
+    void enable_conv2d_direct() {
+        std::vector<GGMLBlock*> blocks;
+        rrdb_net.get_all_blocks(blocks);
+        for (auto block : blocks) {
+            if (block->get_desc() == "Conv2d") {
+                auto conv_block = (Conv2d*)block;
+                conv_block->enable_direct();
+            }
+        }
+    }
+
+    std::string get_desc() {
+        return "esrgan";
+    }
+
+    bool load_from_file(const std::string& file_path) {
+        LOG_INFO("loading esrgan from '%s'", file_path.c_str());
+
+        alloc_params_buffer();
+        std::map<std::string, ggml_tensor*> esrgan_tensors;
+        rrdb_net.get_param_tensors(esrgan_tensors);
+
+        ModelLoader model_loader;
+        if (!model_loader.init_from_file(file_path)) {
+            LOG_ERROR("init esrgan model loader from file failed: '%s'", file_path.c_str());
+            return false;
+        }
+
+        bool success = model_loader.load_tensors(esrgan_tensors);
+
+        if (!success) {
+            LOG_ERROR("load esrgan tensors from model loader failed");
+            return false;
+        }
+
+        LOG_INFO("esrgan model loaded");
+        return success;
+    }
+
+    struct ggml_cgraph* build_graph(struct ggml_tensor* x) {
+        struct ggml_cgraph* gf  = ggml_new_graph(compute_ctx);
+        x                       = to_backend(x);
+        struct ggml_tensor* out = rrdb_net.forward(compute_ctx, x);
+        ggml_build_forward_expand(gf, out);
+        return gf;
+    }
+
+    void compute(const int n_threads,
+                 struct ggml_tensor* x,
+                 ggml_tensor** output,
+                 ggml_context* output_ctx = NULL) {
+        auto get_graph = [&]() -> struct ggml_cgraph* {
+            return build_graph(x);
+        };
+        GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
+    }
+};
+
+#endif  // __ESRGAN_HPP__
@@ -1,4 +1,3 @@
 include_directories(${CMAKE_CURRENT_SOURCE_DIR})

-add_subdirectory(cli)
-add_subdirectory(server)
+add_subdirectory(cli)
@@ -1,24 +1,6 @@
-set(TARGET sd-cli)
+set(TARGET sd)

-add_executable(${TARGET}
-    ../common/common.cpp
-    ../common/log.cpp
-    ../common/media_io.cpp
-    image_metadata.cpp
-    main.cpp
-)
-target_include_directories(${TARGET} PRIVATE
-    "${CMAKE_CURRENT_SOURCE_DIR}/.."
-    "${PROJECT_SOURCE_DIR}/src"
-)
+add_executable(${TARGET} main.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE stable-diffusion zip ${CMAKE_THREAD_LIBS_INIT})
-if(SD_WEBP)
-    target_compile_definitions(${TARGET} PRIVATE SD_USE_WEBP)
-    target_link_libraries(${TARGET} PRIVATE webp libwebpmux)
-endif()
-if(SD_WEBM)
-    target_compile_definitions(${TARGET} PRIVATE SD_USE_WEBM)
-    target_link_libraries(${TARGET} PRIVATE webm)
-endif()
-target_compile_features(${TARGET} PUBLIC c_std_11 cxx_std_17)
+target_link_libraries(${TARGET} PRIVATE stable-diffusion ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PUBLIC c_std_11 cxx_std_17)
@@ -1,203 +0,0 @@
-# Run
-
-```
-usage: ./bin/sd-cli  [options]
-
-CLI Options:
-  -o, --output <string>         path to write result image to. you can use printf-style %d format specifiers for image
-                                sequences (default: ./output.png) (eg. output_%03d.png). Single-file video outputs
-                                support .avi, .webm, and animated .webp
-  --image <string>              path to the image to inspect (for metadata mode)
-  --metadata-format <string>    metadata output format, one of [text, json] (default: text)
-  --preview-path <string>       path to write preview image to (default: ./preview.png). Multi-frame previews support
-                                .avi, .webm, and animated .webp
-  --preview-interval <int>      interval in denoising steps between consecutive updates of the image preview file
-                                (default is 1, meaning updating at every step)
-  --output-begin-idx <int>      starting index for output image sequence, must be non-negative (default 0 if specified
-                                %d in output path, 1 otherwise)
-  --canny                       apply canny preprocessor (edge detection)
-  --convert-name                convert tensor name (for convert mode)
-  -v, --verbose                 print extra info
-  --color                       colors the logging tags according to level
-  --taesd-preview-only          prevents usage of taesd for decoding the final image. (for use with --preview tae)
-  --preview-noisy               enables previewing noisy inputs of the models rather than the denoised outputs
-  --metadata-raw                include raw hex previews for unparsed metadata payloads
-  --metadata-brief              truncate long metadata text values in text output
-  --metadata-all                include structural/container entries such as IHDR, IDAT, and non-metadata JPEG segments
-  -M, --mode                    run mode, one of [img_gen, vid_gen, upscale, convert, metadata], default: img_gen
-  --preview                     preview method. must be one of the following [none, proj, tae, vae] (default is none)
-  -h, --help                    show this help message and exit
-
-Context Options:
-  -m, --model <string>                     path to full model
-  --clip_l <string>                        path to the clip-l text encoder
-  --clip_g <string>                        path to the clip-g text encoder
-  --clip_vision <string>                   path to the clip-vision encoder
-  --t5xxl <string>                         path to the t5xxl text encoder
-  --llm <string>                           path to the llm text encoder. For example: (qwenvl2.5 for qwen-image,
-                                           mistral-small3.2 for flux2, ...)
-  --llm_vision <string>                    path to the llm vit
-  --qwen2vl <string>                       alias of --llm. Deprecated.
-  --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
-  --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
-  --control-net <string>                   path to control net model
-  --embd-dir <string>                      embeddings directory
-  --lora-model-dir <string>                lora model directory
-  --hires-upscalers-dir <string>           highres fix upscaler model directory
-  --tensor-type-rules <string>             weight type per tensor pattern (example: "^vae\.=f16,model\.=q8_0")
-  --photo-maker <string>                   path to PHOTOMAKER model
-  --upscale-model <string>                 path to esrgan model.
-  -t, --threads <int>                      number of threads to use during computation (default: -1). If threads <= 0,
-                                           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; -1 auto-detects free VRAM minus 1 GiB
-  --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
-  --mmap                                   whether to memory-map model
-  --control-net-cpu                        keep controlnet in cpu (for low vram)
-  --clip-on-cpu                            keep clip in cpu (for low vram)
-  --vae-on-cpu                             keep vae in cpu (for low vram)
-  --fa                                     use flash attention
-  --diffusion-fa                           use flash attention in the diffusion model only
-  --diffusion-conv-direct                  use ggml_conv2d_direct in the diffusion model
-  --vae-conv-direct                        use ggml_conv2d_direct in the vae model
-  --circular                               enable circular padding for convolutions
-  --circularx                              enable circular RoPE wrapping on x-axis (width) only
-  --circulary                              enable circular RoPE wrapping on y-axis (height) only
-  --chroma-disable-dit-mask                disable dit mask for chroma
-  --qwen-image-zero-cond-t                 enable zero_cond_t for qwen image
-  --chroma-enable-t5-mask                  enable t5 mask for chroma
-  --type                                   weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K,
-                                           q4_K). If not specified, the default is the type of the weight file
-  --rng                                    RNG, one of [std_default, cuda, cpu], default: cuda(sd-webui), cpu(comfyui)
-  --sampler-rng                            sampler RNG, one of [std_default, cuda, cpu]. If not specified, use --rng
-  --prediction                             prediction type override, one of [eps, v, edm_v, sd3_flow, flux_flow,
-                                           flux2_flow]
-  --lora-apply-mode                        the way to apply LoRA, one of [auto, immediately, at_runtime], default is
-                                           auto. In auto mode, if the model weights contain any quantized parameters,
-                                           the at_runtime mode will be used; otherwise, immediately will be used.The
-                                           immediately mode may have precision and compatibility issues with quantized
-                                           parameters, but it usually offers faster inference speed and, in some cases,
-                                           lower memory usage. The at_runtime mode, on the other hand, is exactly the
-                                           opposite.
-
-Generation Options:
-  -p, --prompt <string>                    the prompt to render
-  -n, --negative-prompt <string>           the negative prompt (default: "")
-  -i, --init-img <string>                  path to the init image
-  --end-img <string>                       path to the end image, required by flf2v
-  --mask <string>                          path to the mask image
-  --control-image <string>                 path to control image, control net
-  --control-video <string>                 path to control video frames, It must be a directory path. The video frames
-                                           inside should be stored as images in lexicographical (character) order. For
-                                           example, if the control video path is `frames`, the directory contain images
-                                           such as 00.png, 01.png, ... etc.
-  --pm-id-images-dir <string>              path to PHOTOMAKER input id images dir
-  --pm-id-embed-path <string>              path to PHOTOMAKER v2 id embed
-  --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 args, key=value list. lcm supports noise_clip_std,
-                                           noise_scale_start, noise_scale_end; ltx2 supports max_shift, base_shift,
-                                           stretch, terminal; euler_ge supports gamma
-  -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)
-  --high-noise-steps <int>                 (high noise) number of sample steps (default: -1 = auto)
-  --clip-skip <int>                        ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer
-                                           (default: -1). <= 0 represents unspecified, will be 1 for SD1.x, 2 for SD2.x
-  -b, --batch-count <int>                  batch count
-  --video-frames <int>                     video frames (default: 1)
-  --fps <int>                              fps (default: 24)
-  --timestep-shift <int>                   shift timestep for NitroFusion models (default: 0). recommended N for
-                                           NitroSD-Realism around 250 and 500 for NitroSD-Vibrant
-  --upscale-repeats <int>                  Run the ESRGAN upscaler this many times (default: 1)
-  --upscale-tile-size <int>                tile size for ESRGAN upscaling (default: 128)
-  --hires-width <int>                      highres fix target width, 0 to use --hires-scale (default: 0)
-  --hires-height <int>                     highres fix target height, 0 to use --hires-scale (default: 0)
-  --hires-steps <int>                      highres fix second pass sample steps, 0 to reuse --steps (default: 0)
-  --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)
-  --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
-  --skip-layer-start <float>               SLG enabling point (default: 0.01)
-  --skip-layer-end <float>                 SLG disabling point (default: 0.2)
-  --eta <float>                            noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and
-                                           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-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:
-                                           0)
-  --high-noise-skip-layer-start <float>    (high noise) SLG enabling point (default: 0.01)
-  --high-noise-skip-layer-end <float>      (high noise) SLG disabling point (default: 0.2)
-  --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>              
-  --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
-                                           `--high-noise-steps` is set to -1
-  --vace-strength <float>                  wan vace strength
-  --vae-tile-overlap <float>               tile overlap for vae tiling, in fraction of tile size (default: 0.5)
-  --hires-scale <float>                    highres fix scale when target size is not set (default: 2.0)
-  --hires-denoising-strength <float>       highres fix second pass denoising strength (default: 0.7)
-  --increase-ref-index                     automatically increase the indices of references images based on the order
-                                           they are listed (starting with 1).
-  --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, 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, 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, ltx2], default:
-                                           model-specific
-  --sigmas                                 custom sigma values for the sampler, comma-separated (e.g.,
-                                           "14.61,7.8,3.5,0.0").
-  --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)
-  --cache-mode                             caching method: 'easycache' (DiT), 'ucache' (UNET),
-                                           'dbcache'/'taylorseer'/'cache-dit' (DiT block-level), 'spectrum' (UNET/DiT
-                                           Chebyshev+Taylor forecasting)
-  --cache-option                           named cache params (key=value format, comma-separated). easycache/ucache:
-                                           threshold=,start=,end=,decay=,relative=,reset=; dbcache/taylorseer/cache-dit:
-                                           Fn=,Bn=,threshold=,warmup=; spectrum: w=,m=,lam=,window=,flex=,warmup=,stop=.
-                                           Examples: "threshold=0.25" or "threshold=1.5,reset=0"
-  --scm-mask                               SCM steps mask for cache-dit: comma-separated 0/1 (e.g.,
-                                           "1,1,1,0,0,1,0,0,1,0") - 1=compute, 0=can cache
-  --scm-policy                             SCM policy: 'dynamic' (default) or 'static'
-  --vae-tile-size                          tile size for vae tiling, format [X]x[Y] (default: 32x32)
-  --vae-relative-tile-size                 relative tile size for vae tiling, format [X]x[Y], in fraction of image size
-                                           if < 1, in number of tiles per dim if >=1 (overrides --vae-tile-size)
-```
-
-Metadata mode inspects PNG/JPEG container metadata without loading any model:
-
-```bash
-./bin/sd-cli -M metadata --image ./output.png
-./bin/sd-cli -M metadata --image ./output.jpg --metadata-format json
-./bin/sd-cli -M metadata --image ./output.png --metadata-raw
-./bin/sd-cli -M metadata --image ./output.png --metadata-all
-```
@@ -0,0 +1,217 @@
+#ifndef __AVI_WRITER_H__
+#define __AVI_WRITER_H__
+
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "stable-diffusion.h"
+
+#ifndef INCLUDE_STB_IMAGE_WRITE_H
+#include "stb_image_write.h"
+#endif
+
+typedef struct {
+    uint32_t offset;
+    uint32_t size;
+} avi_index_entry;
+
+// Write 32-bit little-endian integer
+void write_u32_le(FILE* f, uint32_t val) {
+    fwrite(&val, 4, 1, f);
+}
+
+// Write 16-bit little-endian integer
+void write_u16_le(FILE* f, uint16_t val) {
+    fwrite(&val, 2, 1, f);
+}
+
+/**
+ * Create an MJPG AVI file from an array of sd_image_t images.
+ * Images are encoded to JPEG using stb_image_write.
+ *
+ * @param filename Output AVI file name.
+ * @param images Array of input images.
+ * @param num_images Number of images in the array.
+ * @param fps Frames per second for the video.
+ * @param quality JPEG quality (0-100).
+ * @return 0 on success, -1 on failure.
+ */
+int create_mjpg_avi_from_sd_images(const char* filename, sd_image_t* images, int num_images, int fps, int quality = 90) {
+    if (num_images == 0) {
+        fprintf(stderr, "Error: Image array is empty.\n");
+        return -1;
+    }
+
+    FILE* f = fopen(filename, "wb");
+    if (!f) {
+        perror("Error opening file for writing");
+        return -1;
+    }
+
+    uint32_t width    = images[0].width;
+    uint32_t height   = images[0].height;
+    uint32_t channels = images[0].channel;
+    if (channels != 3 && channels != 4) {
+        fprintf(stderr, "Error: Unsupported channel count: %u\n", channels);
+        fclose(f);
+        return -1;
+    }
+
+    // --- RIFF AVI Header ---
+    fwrite("RIFF", 4, 1, f);
+    long riff_size_pos = ftell(f);
+    write_u32_le(f, 0);  // Placeholder for file size
+    fwrite("AVI ", 4, 1, f);
+
+    // 'hdrl' LIST (header list)
+    fwrite("LIST", 4, 1, f);
+    write_u32_le(f, 4 + 8 + 56 + 8 + 4 + 8 + 56 + 8 + 40);
+    fwrite("hdrl", 4, 1, f);
+
+    // 'avih' chunk (AVI main header)
+    fwrite("avih", 4, 1, f);
+    write_u32_le(f, 56);
+    write_u32_le(f, 1000000 / fps);       // Microseconds per frame
+    write_u32_le(f, 0);                   // Max bytes per second
+    write_u32_le(f, 0);                   // Padding granularity
+    write_u32_le(f, 0x110);               // Flags (HASINDEX | ISINTERLEAVED)
+    write_u32_le(f, num_images);          // Total frames
+    write_u32_le(f, 0);                   // Initial frames
+    write_u32_le(f, 1);                   // Number of streams
+    write_u32_le(f, width * height * 3);  // Suggested buffer size
+    write_u32_le(f, width);
+    write_u32_le(f, height);
+    write_u32_le(f, 0);  // Reserved
+    write_u32_le(f, 0);  // Reserved
+    write_u32_le(f, 0);  // Reserved
+    write_u32_le(f, 0);  // Reserved
+
+    // 'strl' LIST (stream list)
+    fwrite("LIST", 4, 1, f);
+    write_u32_le(f, 4 + 8 + 56 + 8 + 40);
+    fwrite("strl", 4, 1, f);
+
+    // 'strh' chunk (stream header)
+    fwrite("strh", 4, 1, f);
+    write_u32_le(f, 56);
+    fwrite("vids", 4, 1, f);              // Stream type: video
+    fwrite("MJPG", 4, 1, f);              // Codec: Motion JPEG
+    write_u32_le(f, 0);                   // Flags
+    write_u16_le(f, 0);                   // Priority
+    write_u16_le(f, 0);                   // Language
+    write_u32_le(f, 0);                   // Initial frames
+    write_u32_le(f, 1);                   // Scale
+    write_u32_le(f, fps);                 // Rate
+    write_u32_le(f, 0);                   // Start
+    write_u32_le(f, num_images);          // Length
+    write_u32_le(f, width * height * 3);  // Suggested buffer size
+    write_u32_le(f, (uint32_t)-1);        // Quality
+    write_u32_le(f, 0);                   // Sample size
+    write_u16_le(f, 0);                   // rcFrame.left
+    write_u16_le(f, 0);                   // rcFrame.top
+    write_u16_le(f, 0);                   // rcFrame.right
+    write_u16_le(f, 0);                   // rcFrame.bottom
+
+    // 'strf' chunk (stream format: BITMAPINFOHEADER)
+    fwrite("strf", 4, 1, f);
+    write_u32_le(f, 40);
+    write_u32_le(f, 40);  // biSize
+    write_u32_le(f, width);
+    write_u32_le(f, height);
+    write_u16_le(f, 1);                   // biPlanes
+    write_u16_le(f, 24);                  // biBitCount
+    fwrite("MJPG", 4, 1, f);              // biCompression (FOURCC)
+    write_u32_le(f, width * height * 3);  // biSizeImage
+    write_u32_le(f, 0);                   // XPelsPerMeter
+    write_u32_le(f, 0);                   // YPelsPerMeter
+    write_u32_le(f, 0);                   // Colors used
+    write_u32_le(f, 0);                   // Colors important
+
+    // 'movi' LIST (video frames)
+    long movi_list_pos = ftell(f);
+    fwrite("LIST", 4, 1, f);
+    long movi_size_pos = ftell(f);
+    write_u32_le(f, 0);  // Placeholder for movi size
+    fwrite("movi", 4, 1, f);
+
+    avi_index_entry* index = (avi_index_entry*)malloc(sizeof(avi_index_entry) * num_images);
+    if (!index) {
+        fclose(f);
+        return -1;
+    }
+
+    // Encode and write each frame as JPEG
+    struct {
+        uint8_t* buf;
+        size_t size;
+    } jpeg_data;
+
+    for (int i = 0; i < num_images; i++) {
+        jpeg_data.buf  = NULL;
+        jpeg_data.size = 0;
+
+        // Callback function to collect JPEG data into memory
+        auto write_to_buf = [](void* context, void* data, int size) {
+            auto jd = (decltype(jpeg_data)*)context;
+            jd->buf = (uint8_t*)realloc(jd->buf, jd->size + size);
+            memcpy(jd->buf + jd->size, data, size);
+            jd->size += size;
+        };
+
+        // Encode to JPEG in memory
+        stbi_write_jpg_to_func(
+            write_to_buf,
+            &jpeg_data,
+            images[i].width,
+            images[i].height,
+            channels,
+            images[i].data,
+            quality);
+
+        // Write '00dc' chunk (video frame)
+        fwrite("00dc", 4, 1, f);
+        write_u32_le(f, jpeg_data.size);
+        index[i].offset = ftell(f) - 8;
+        index[i].size   = jpeg_data.size;
+        fwrite(jpeg_data.buf, 1, jpeg_data.size, f);
+
+        // Align to even byte size
+        if (jpeg_data.size % 2)
+            fputc(0, f);
+
+        free(jpeg_data.buf);
+    }
+
+    // Finalize 'movi' size
+    long cur_pos   = ftell(f);
+    long movi_size = cur_pos - movi_size_pos - 4;
+    fseek(f, movi_size_pos, SEEK_SET);
+    write_u32_le(f, movi_size);
+    fseek(f, cur_pos, SEEK_SET);
+
+    // Write 'idx1' index
+    fwrite("idx1", 4, 1, f);
+    write_u32_le(f, num_images * 16);
+    for (int i = 0; i < num_images; i++) {
+        fwrite("00dc", 4, 1, f);
+        write_u32_le(f, 0x10);
+        write_u32_le(f, index[i].offset);
+        write_u32_le(f, index[i].size);
+    }
+
+    // Finalize RIFF size
+    cur_pos        = ftell(f);
+    long file_size = cur_pos - riff_size_pos - 4;
+    fseek(f, riff_size_pos, SEEK_SET);
+    write_u32_le(f, file_size);
+    fseek(f, cur_pos, SEEK_SET);
+
+    fclose(f);
+    free(index);
+
+    return 0;
+}
+
+#endif  // __AVI_WRITER_H__
@@ -1,21 +0,0 @@
-#pragma once
-
-#include <iosfwd>
-#include <string>
-
-enum class MetadataOutputFormat {
-    TEXT,
-    JSON,
-};
-
-struct MetadataReadOptions {
-    MetadataOutputFormat output_format = MetadataOutputFormat::TEXT;
-    bool include_raw                   = false;
-    bool brief                         = false;
-    bool include_structural            = false;
-};
-
-bool print_image_metadata(const std::string& image_path,
-                          const MetadataReadOptions& options,
-                          std::ostream& out,
-                          std::string& error);
@@ -1,268 +0,0 @@
-#ifndef __EXAMPLES_COMMON_COMMON_H__
-#define __EXAMPLES_COMMON_COMMON_H__
-
-#include <cmath>
-#include <cstdint>
-#include <functional>
-#include <map>
-#include <string>
-#include <vector>
-
-#include "log.h"
-#include "resource_owners.hpp"
-#include "stable-diffusion.h"
-
-#define SAFE_STR(s) ((s) ? (s) : "")
-#define BOOL_STR(b) ((b) ? "true" : "false")
-
-extern const char* const modes_str[];
-#define SD_ALL_MODES_STR "img_gen, vid_gen, convert, upscale, metadata"
-
-enum SDMode {
-    IMG_GEN,
-    VID_GEN,
-    CONVERT,
-    UPSCALE,
-    METADATA,
-    MODE_COUNT
-};
-
-struct StringOption {
-    std::string short_name;
-    std::string long_name;
-    std::string desc;
-    std::string* target;
-};
-
-struct IntOption {
-    std::string short_name;
-    std::string long_name;
-    std::string desc;
-    int* target;
-};
-
-struct FloatOption {
-    std::string short_name;
-    std::string long_name;
-    std::string desc;
-    float* target;
-};
-
-struct BoolOption {
-    std::string short_name;
-    std::string long_name;
-    std::string desc;
-    bool keep_true;
-    bool* target;
-};
-
-struct ManualOption {
-    std::string short_name;
-    std::string long_name;
-    std::string desc;
-    std::function<int(int argc, const char** argv, int index)> cb;
-};
-
-struct ArgOptions {
-    std::vector<StringOption> string_options;
-    std::vector<IntOption> int_options;
-    std::vector<FloatOption> float_options;
-    std::vector<BoolOption> bool_options;
-    std::vector<ManualOption> manual_options;
-
-    static std::string wrap_text(const std::string& text, size_t width, size_t indent);
-    void print() const;
-};
-
-bool parse_options(int argc, const char** argv, const std::vector<ArgOptions>& options_list);
-bool decode_base64_image(const std::string& encoded_input,
-                         int target_channels,
-                         int expected_width,
-                         int expected_height,
-                         SDImageOwner& out_image);
-
-struct SDContextParams {
-    int n_threads = -1;
-    std::string model_path;
-    std::string clip_l_path;
-    std::string clip_g_path;
-    std::string clip_vision_path;
-    std::string t5xxl_path;
-    std::string llm_path;
-    std::string llm_vision_path;
-    std::string diffusion_model_path;
-    std::string high_noise_diffusion_model_path;
-    std::string embeddings_connectors_path;
-    std::string vae_path;
-    std::string audio_vae_path;
-    std::string taesd_path;
-    std::string esrgan_path;
-    std::string control_net_path;
-    std::string embedding_dir;
-    std::string photo_maker_path;
-    sd_type_t wtype = SD_TYPE_COUNT;
-    std::string tensor_type_rules;
-    std::string lora_model_dir = ".";
-    std::string hires_upscalers_dir;
-
-    std::map<std::string, std::string> embedding_map;
-    std::vector<sd_embedding_t> embedding_vec;
-
-    rng_type_t rng_type         = CUDA_RNG;
-    rng_type_t sampler_rng_type = RNG_TYPE_COUNT;
-    bool offload_params_to_cpu  = false;
-    float max_vram              = 0.f;
-    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;
-    bool circular_y = false;
-
-    bool chroma_use_dit_mask = true;
-    bool chroma_use_t5_mask  = false;
-    int chroma_t5_mask_pad   = 1;
-
-    bool qwen_image_zero_cond_t = false;
-
-    prediction_t prediction           = PREDICTION_COUNT;
-    lora_apply_mode_t lora_apply_mode = LORA_APPLY_AUTO;
-
-    bool force_sdxl_vae_conv_scale = false;
-
-    float flow_shift = INFINITY;
-    ArgOptions get_options();
-    void build_embedding_map();
-    bool resolve(SDMode mode);
-    bool validate(SDMode mode);
-    bool resolve_and_validate(SDMode mode);
-    std::string to_string() const;
-    sd_ctx_params_t to_sd_ctx_params_t(bool vae_decode_only, bool free_params_immediately, bool taesd_preview);
-};
-
-struct SDGenerationParams {
-    // User-facing input fields.
-    std::string prompt;
-    std::string negative_prompt;
-    int clip_skip              = -1;  // <= 0 represents unspecified
-    int width                  = -1;
-    int height                 = -1;
-    int batch_count            = 1;
-    int64_t seed               = 42;
-    float strength             = 0.75f;
-    float control_strength     = 0.9f;
-    bool auto_resize_ref_image = true;
-    bool increase_ref_index    = false;
-    bool embed_image_metadata  = true;
-
-    std::string init_image_path;
-    std::string end_image_path;
-    std::string mask_image_path;
-    std::string control_image_path;
-    std::vector<std::string> ref_image_paths;
-    std::string control_video_path;
-
-    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};
-
-    std::vector<float> custom_sigmas;
-
-    std::string cache_mode;
-    std::string cache_option;
-    std::string scm_mask;
-    bool scm_policy_dynamic = true;
-    sd_cache_params_t cache_params{};
-
-    float moe_boundary                   = 0.875f;
-    int video_frames                     = 1;
-    int fps                              = 16;
-    float vace_strength                  = 1.f;
-    sd_tiling_params_t vae_tiling_params = {false, false, 0, 0, 0.5f, 0.0f, 0.0f};
-
-    std::string pm_id_images_dir;
-    std::string pm_id_embed_path;
-    float pm_style_strength = 20.f;
-
-    int upscale_repeats   = 1;
-    int upscale_tile_size = 128;
-
-    bool hires_enabled         = false;
-    std::string hires_upscaler = "Latent";
-    std::string hires_upscaler_model_path;
-    float hires_scale              = 2.f;
-    int hires_width                = 0;
-    int hires_height               = 0;
-    int hires_steps                = 0;
-    float hires_denoising_strength = 0.7f;
-    int hires_upscale_tile_size    = 128;
-
-    std::map<std::string, float> lora_map;
-    std::map<std::string, float> high_noise_lora_map;
-
-    // Derived and normalized fields.
-    std::string prompt_with_lora;  // for metadata record only
-    std::vector<sd_lora_t> lora_vec;
-    sd_hires_upscaler_t resolved_hires_upscaler;
-
-    // Owned execution payload.
-    SDImageOwner init_image;
-    SDImageOwner end_image;
-    std::vector<SDImageOwner> ref_images;
-    SDImageOwner mask_image;
-    SDImageOwner control_image;
-    std::vector<SDImageOwner> pm_id_images;
-    std::vector<SDImageOwner> control_frames;
-
-    // Backing storage for sd_img_gen_params_t view fields.
-    std::vector<sd_image_t> ref_image_views;
-    std::vector<sd_image_t> pm_id_image_views;
-    std::vector<sd_image_t> control_frame_views;
-
-    SDGenerationParams();
-    SDGenerationParams(const SDGenerationParams& other)                = default;
-    SDGenerationParams& operator=(const SDGenerationParams& other)     = default;
-    SDGenerationParams(SDGenerationParams&& other) noexcept            = default;
-    SDGenerationParams& operator=(SDGenerationParams&& other) noexcept = default;
-    ArgOptions get_options();
-    bool from_json_str(const std::string& json_str,
-                       const std::function<std::string(const std::string&)>& lora_path_resolver = {});
-    bool initialize_cache_params();
-    void extract_and_remove_lora(const std::string& lora_model_dir);
-    bool width_and_height_are_set() const;
-    void set_width_and_height_if_unset(int w, int h);
-    int get_resolved_width() const;
-    int get_resolved_height() const;
-    bool resolve(const std::string& lora_model_dir, const std::string& hires_upscalers_dir, bool strict = false);
-    bool validate(SDMode mode);
-    bool resolve_and_validate(SDMode mode,
-                              const std::string& lora_model_dir,
-                              const std::string& hires_upscalers_dir,
-                              bool strict = false);
-    sd_img_gen_params_t to_sd_img_gen_params_t();
-    sd_vid_gen_params_t to_sd_vid_gen_params_t();
-    std::string to_string() const;
-};
-
-std::string version_string();
-std::string build_sdcpp_image_metadata_json(const SDContextParams& ctx_params,
-                                            const SDGenerationParams& gen_params,
-                                            int64_t seed,
-                                            SDMode mode = IMG_GEN);
-std::string get_image_params(const SDContextParams& ctx_params,
-                             const SDGenerationParams& gen_params,
-                             int64_t seed,
-                             SDMode mode = IMG_GEN);
-
-#endif  // __EXAMPLES_COMMON_COMMON_H__
@@ -1,115 +0,0 @@
-#include "log.h"
-
-#include <vector>
-
-bool log_verbose = false;
-bool log_color   = false;
-
-std::string sd_basename(const std::string& path) {
-    size_t pos = path.find_last_of('/');
-    if (pos != std::string::npos) {
-        return path.substr(pos + 1);
-    }
-    pos = path.find_last_of('\\');
-    if (pos != std::string::npos) {
-        return path.substr(pos + 1);
-    }
-    return path;
-}
-
-void print_utf8(FILE* stream, const char* utf8) {
-    if (!utf8) {
-        return;
-    }
-
-#ifdef _WIN32
-    HANDLE h = (stream == stderr)
-                   ? GetStdHandle(STD_ERROR_HANDLE)
-                   : GetStdHandle(STD_OUTPUT_HANDLE);
-
-    DWORD mode;
-    BOOL is_console = GetConsoleMode(h, &mode);
-
-    if (is_console) {
-        int wlen = MultiByteToWideChar(CP_UTF8, 0, utf8, -1, NULL, 0);
-        if (wlen <= 0) {
-            return;
-        }
-
-        std::vector<wchar_t> wbuf(static_cast<size_t>(wlen));
-
-        MultiByteToWideChar(CP_UTF8, 0, utf8, -1, wbuf.data(), wlen);
-
-        DWORD written;
-        WriteConsoleW(h, wbuf.data(), wlen - 1, &written, NULL);
-    } else {
-        DWORD written;
-        WriteFile(h, utf8, (DWORD)strlen(utf8), &written, NULL);
-    }
-#else
-    fputs(utf8, stream);
-#endif
-}
-
-void log_print(enum sd_log_level_t level, const char* log, bool verbose, bool color) {
-    int tag_color;
-    const char* level_str;
-    FILE* out_stream = (level == SD_LOG_ERROR) ? stderr : stdout;
-
-    if (!log || (!verbose && level <= SD_LOG_DEBUG)) {
-        return;
-    }
-
-    switch (level) {
-        case SD_LOG_DEBUG:
-            tag_color = 37;
-            level_str = "DEBUG";
-            break;
-        case SD_LOG_INFO:
-            tag_color = 34;
-            level_str = "INFO";
-            break;
-        case SD_LOG_WARN:
-            tag_color = 35;
-            level_str = "WARN";
-            break;
-        case SD_LOG_ERROR:
-            tag_color = 31;
-            level_str = "ERROR";
-            break;
-        default:
-            tag_color = 33;
-            level_str = "?????";
-            break;
-    }
-
-    if (color) {
-        fprintf(out_stream, "\033[%d;1m[%-5s]\033[0m ", tag_color, level_str);
-    } else {
-        fprintf(out_stream, "[%-5s] ", level_str);
-    }
-    print_utf8(out_stream, log);
-    fflush(out_stream);
-}
-
-void example_log_printf(sd_log_level_t level, const char* file, int line, const char* format, ...) {
-    constexpr size_t LOG_BUFFER_SIZE = 4096;
-
-    va_list args;
-    va_start(args, format);
-
-    static char log_buffer[LOG_BUFFER_SIZE + 1];
-    int written = snprintf(log_buffer, LOG_BUFFER_SIZE, "%s:%-4d - ", sd_basename(file).c_str(), line);
-
-    if (written >= 0 && written < static_cast<int>(LOG_BUFFER_SIZE)) {
-        vsnprintf(log_buffer + written, LOG_BUFFER_SIZE - written, format, args);
-    }
-    size_t len = strlen(log_buffer);
-    if (len == 0 || log_buffer[len - 1] != '\n') {
-        strncat(log_buffer, "\n", LOG_BUFFER_SIZE - len);
-    }
-
-    log_print(level, log_buffer, log_verbose, log_color);
-
-    va_end(args);
-}
@@ -1,32 +0,0 @@
-#ifndef __EXAMPLE_LOG_H__
-#define __EXAMPLE_LOG_H__
-
-#include <cstdarg>
-#include <cstdio>
-#include <cstdlib>
-#include <cstring>
-#include <string>
-
-#if defined(_WIN32)
-#ifndef NOMINMAX
-#define NOMINMAX
-#endif
-#include <windows.h>
-#endif  // _WIN32
-
-#include "stable-diffusion.h"
-
-extern bool log_verbose;
-extern bool log_color;
-
-std::string sd_basename(const std::string& path);
-void print_utf8(FILE* stream, const char* utf8);
-void log_print(sd_log_level_t level, const char* log, bool verbose, bool color);
-void example_log_printf(sd_log_level_t level, const char* file, int line, const char* format, ...);
-
-#define LOG_DEBUG(format, ...) example_log_printf(SD_LOG_DEBUG, __FILE__, __LINE__, format, ##__VA_ARGS__)
-#define LOG_INFO(format, ...) example_log_printf(SD_LOG_INFO, __FILE__, __LINE__, format, ##__VA_ARGS__)
-#define LOG_WARN(format, ...) example_log_printf(SD_LOG_WARN, __FILE__, __LINE__, format, ##__VA_ARGS__)
-#define LOG_ERROR(format, ...) example_log_printf(SD_LOG_ERROR, __FILE__, __LINE__, format, ##__VA_ARGS__)
-
-#endif  // __EXAMPLE_LOG_H__
@@ -1,113 +0,0 @@
-#ifndef __MEDIA_IO_H__
-#define __MEDIA_IO_H__
-
-#include <cstdint>
-#include <string>
-#include <vector>
-
-#include "stable-diffusion.h"
-
-enum class EncodedImageFormat {
-    JPEG,
-    PNG,
-    WEBP,
-    UNKNOWN,
-};
-
-EncodedImageFormat encoded_image_format_from_path(const std::string& path);
-
-std::vector<uint8_t> encode_image_to_vector(EncodedImageFormat format,
-                                            const uint8_t* image,
-                                            int width,
-                                            int height,
-                                            int channels,
-                                            const std::string& parameters = "",
-                                            int quality                   = 90);
-
-bool write_image_to_file(const std::string& path,
-                         const uint8_t* image,
-                         int width,
-                         int height,
-                         int channels,
-                         const std::string& parameters = "",
-                         int quality                   = 90);
-
-uint8_t* load_image_from_file(const char* image_path,
-                              int& width,
-                              int& height,
-                              int expected_width   = 0,
-                              int expected_height  = 0,
-                              int expected_channel = 3);
-
-bool load_sd_image_from_file(sd_image_t* image,
-                             const char* image_path,
-                             int expected_width   = 0,
-                             int expected_height  = 0,
-                             int expected_channel = 3);
-
-uint8_t* load_image_from_memory(const char* image_bytes,
-                                int len,
-                                int& width,
-                                int& height,
-                                int expected_width   = 0,
-                                int expected_height  = 0,
-                                int expected_channel = 3);
-
-int create_mjpg_avi_from_sd_images(const char* filename,
-                                   sd_image_t* images,
-                                   int num_images,
-                                   int fps,
-                                   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,
-                                                              const sd_audio_t* audio = nullptr);
-
-#ifdef SD_USE_WEBP
-int create_animated_webp_from_sd_images(const char* filename,
-                                        sd_image_t* images,
-                                        int num_images,
-                                        int fps,
-                                        int quality = 90);
-std::vector<uint8_t> create_animated_webp_from_sd_images_to_vector(sd_image_t* images,
-                                                                   int num_images,
-                                                                   int fps,
-                                                                   int quality = 90);
-#endif
-
-#ifdef SD_USE_WEBM
-int create_webm_from_sd_images(const char* filename,
-                               sd_image_t* images,
-                               int num_images,
-                               int fps,
-                               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,
-                                                          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,
-                                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,
-                                                           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__
@@ -1,236 +0,0 @@
-#ifndef __EXAMPLE_RESOURCE_OWNERS_H__
-#define __EXAMPLE_RESOURCE_OWNERS_H__
-
-#include <cstdio>
-#include <cstdlib>
-#include <cstring>
-#include <memory>
-#include <utility>
-#include <vector>
-
-#include "stable-diffusion.h"
-
-struct FreeDeleter {
-    void operator()(void* ptr) const {
-        free(ptr);
-    }
-};
-
-struct FileCloser {
-    void operator()(FILE* file) const {
-        if (file != nullptr) {
-            fclose(file);
-        }
-    }
-};
-
-struct SDCtxDeleter {
-    void operator()(sd_ctx_t* ctx) const {
-        if (ctx != nullptr) {
-            free_sd_ctx(ctx);
-        }
-    }
-};
-
-struct UpscalerCtxDeleter {
-    void operator()(upscaler_ctx_t* ctx) const {
-        if (ctx != nullptr) {
-            free_upscaler_ctx(ctx);
-        }
-    }
-};
-
-template <typename T>
-using FreeUniquePtr = std::unique_ptr<T, FreeDeleter>;
-
-using FilePtr        = std::unique_ptr<FILE, FileCloser>;
-using SDCtxPtr       = std::unique_ptr<sd_ctx_t, SDCtxDeleter>;
-using UpscalerCtxPtr = std::unique_ptr<upscaler_ctx_t, UpscalerCtxDeleter>;
-
-class SDImageOwner {
-private:
-    static sd_image_t copy_image(const sd_image_t& image) {
-        if (image.data == nullptr) {
-            return {image.width, image.height, image.channel, nullptr};
-        }
-
-        const size_t byte_count = static_cast<size_t>(image.width) * image.height * image.channel;
-        uint8_t* raw_copy       = static_cast<uint8_t*>(malloc(byte_count));
-        if (raw_copy == nullptr) {
-            return {0, 0, 0, nullptr};
-        }
-
-        std::memcpy(raw_copy, image.data, byte_count);
-        return {image.width, image.height, image.channel, raw_copy};
-    }
-
-    sd_image_t image_ = {0, 0, 0, nullptr};
-
-public:
-    SDImageOwner() = default;
-    explicit SDImageOwner(sd_image_t image)
-        : image_(image) {
-    }
-
-    SDImageOwner(const SDImageOwner& other)
-        : image_(copy_image(other.image_)) {
-    }
-
-    SDImageOwner& operator=(const SDImageOwner& other) {
-        if (this != &other) {
-            reset(copy_image(other.image_));
-        }
-        return *this;
-    }
-
-    SDImageOwner(SDImageOwner&& other) noexcept
-        : image_(other.release()) {
-    }
-
-    SDImageOwner& operator=(SDImageOwner&& other) noexcept {
-        if (this != &other) {
-            reset();
-            image_ = other.release();
-        }
-        return *this;
-    }
-
-    ~SDImageOwner() {
-        reset();
-    }
-
-    sd_image_t* put() {
-        if (image_.data != nullptr) {
-            free(image_.data);
-            image_.data = nullptr;
-        }
-        image_.width   = 0;
-        image_.height  = 0;
-        image_.channel = 0;
-        return &image_;
-    }
-
-    sd_image_t& get() {
-        return image_;
-    }
-
-    const sd_image_t& get() const {
-        return image_;
-    }
-
-    sd_image_t release() {
-        sd_image_t image = image_;
-        image_           = {0, 0, 0, nullptr};
-        return image;
-    }
-
-    void reset(sd_image_t image = {0, 0, 0, nullptr}) {
-        if (image_.data != nullptr) {
-            free(image_.data);
-        }
-        image_ = image;
-    }
-};
-
-class SDImageVec {
-private:
-    std::vector<sd_image_t> images_;
-
-public:
-    SDImageVec() = default;
-
-    SDImageVec(const SDImageVec&)            = delete;
-    SDImageVec& operator=(const SDImageVec&) = delete;
-
-    SDImageVec(SDImageVec&& other) noexcept
-        : images_(std::move(other.images_)) {
-    }
-
-    SDImageVec& operator=(SDImageVec&& other) noexcept {
-        if (this != &other) {
-            clear();
-            images_ = std::move(other.images_);
-        }
-        return *this;
-    }
-
-    ~SDImageVec() {
-        clear();
-    }
-
-    void push_back(sd_image_t image) {
-        images_.push_back(image);
-    }
-
-    void push_back(SDImageOwner&& image) {
-        images_.push_back(image.release());
-    }
-
-    void reserve(size_t count) {
-        images_.reserve(count);
-    }
-
-    void adopt(sd_image_t* images, int count) {
-        clear();
-        if (images == nullptr || count <= 0) {
-            free(images);
-            return;
-        }
-
-        images_.reserve(static_cast<size_t>(count));
-        for (int i = 0; i < count; ++i) {
-            images_.push_back(images[i]);
-        }
-        free(images);
-    }
-
-    size_t size() const {
-        return images_.size();
-    }
-
-    bool empty() const {
-        return images_.empty();
-    }
-
-    int count() const {
-        return static_cast<int>(images_.size());
-    }
-
-    explicit operator bool() const {
-        return !images_.empty();
-    }
-
-    sd_image_t* data() {
-        return images_.data();
-    }
-
-    const sd_image_t* data() const {
-        return images_.data();
-    }
-
-    sd_image_t& operator[](size_t index) {
-        return images_[index];
-    }
-
-    const sd_image_t& operator[](size_t index) const {
-        return images_[index];
-    }
-
-    std::vector<sd_image_t>& raw() {
-        return images_;
-    }
-
-    const std::vector<sd_image_t>& raw() const {
-        return images_;
-    }
-
-    void clear() {
-        for (sd_image_t& image : images_) {
-            free(image.data);
-            image.data = nullptr;
-        }
-        images_.clear();
-    }
-};
-
-#endif  // __EXAMPLE_RESOURCE_OWNERS_H__
@@ -1,104 +0,0 @@
-set(TARGET sd-server)
-
-option(SD_SERVER_BUILD_FRONTEND "Build server frontend with pnpm" ON)
-
-set(FRONTEND_DIR "${CMAKE_CURRENT_SOURCE_DIR}/frontend")
-set(GENERATED_HTML_HEADER "${FRONTEND_DIR}/dist/gen_index_html.h")
-
-set(HAVE_FRONTEND_BUILD OFF)
-
-if(SD_SERVER_BUILD_FRONTEND AND EXISTS "${FRONTEND_DIR}")
-    if(WIN32)
-        find_program(PNPM_EXECUTABLE NAMES pnpm.cmd pnpm)
-    else()
-        find_program(PNPM_EXECUTABLE NAMES pnpm)
-    endif()
-
-    if(PNPM_EXECUTABLE)
-        message(STATUS "Frontend dir found: ${FRONTEND_DIR}")
-        message(STATUS "pnpm found: ${PNPM_EXECUTABLE}")
-
-        set(HAVE_FRONTEND_BUILD ON)
-
-        add_custom_target(${TARGET}_frontend_install
-            COMMAND "${PNPM_EXECUTABLE}" -C "${FRONTEND_DIR}" install
-            WORKING_DIRECTORY "${FRONTEND_DIR}"
-            COMMENT "Installing frontend dependencies"
-            VERBATIM
-        )
-
-        add_custom_target(${TARGET}_frontend_build
-            COMMAND "${PNPM_EXECUTABLE}" -C "${FRONTEND_DIR}" run build
-            WORKING_DIRECTORY "${FRONTEND_DIR}"
-            COMMENT "Building frontend"
-            VERBATIM
-        )
-
-        add_custom_target(${TARGET}_frontend_header
-            COMMAND "${PNPM_EXECUTABLE}" -C "${FRONTEND_DIR}" run build:header
-            WORKING_DIRECTORY "${FRONTEND_DIR}"
-            COMMENT "Generating gen_index_html.h"
-            VERBATIM
-        )
-
-        add_dependencies(${TARGET}_frontend_build ${TARGET}_frontend_install)
-        add_dependencies(${TARGET}_frontend_header ${TARGET}_frontend_build)
-
-        add_custom_target(${TARGET}_frontend
-            DEPENDS ${TARGET}_frontend_header
-        )
-
-        set_source_files_properties("${GENERATED_HTML_HEADER}" PROPERTIES GENERATED TRUE)
-    else()
-        if(EXISTS "${GENERATED_HTML_HEADER}")
-            message(STATUS "pnpm not found; using pre-built frontend header detected at ${GENERATED_HTML_HEADER}")
-            set(HAVE_FRONTEND_BUILD ON)
-            add_custom_target(${TARGET}_frontend)
-        else()
-            message(WARNING "pnpm not found; frontend build disabled.")
-        endif()
-    endif()
-else()
-    message(STATUS "Frontend disabled or directory not found: ${FRONTEND_DIR}")
-endif()
-
-add_executable(${TARGET}
-    ../common/common.cpp
-    ../common/log.cpp
-    ../common/media_io.cpp
-    main.cpp
-    runtime.cpp
-    async_jobs.cpp
-    routes_index.cpp
-    routes_openai.cpp
-    routes_sdapi.cpp
-    routes_sdcpp.cpp
-)
-
-if(HAVE_FRONTEND_BUILD)
-    add_dependencies(${TARGET} ${TARGET}_frontend)
-    target_sources(${TARGET} PRIVATE "${GENERATED_HTML_HEADER}")
-    target_include_directories(${TARGET} PRIVATE "${FRONTEND_DIR}/dist")
-    target_compile_definitions(${TARGET} PRIVATE HAVE_INDEX_HTML)
-    message(STATUS "HAVE_INDEX_HTML enabled")
-else()
-    message(STATUS "HAVE_INDEX_HTML disabled")
-endif()
-
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE stable-diffusion ${CMAKE_THREAD_LIBS_INIT})
-if(SD_WEBP)
-    target_compile_definitions(${TARGET} PRIVATE SD_USE_WEBP)
-    target_link_libraries(${TARGET} PRIVATE webp libwebpmux)
-endif()
-if(SD_WEBM)
-    target_compile_definitions(${TARGET} PRIVATE SD_USE_WEBM)
-    target_link_libraries(${TARGET} PRIVATE webm)
-endif()
-
-# due to httplib; it contains a pragma for MSVC, but other things need explicit flags
-if(WIN32 AND NOT MSVC)
-    target_link_libraries(${TARGET} PRIVATE ws2_32)
-endif()
-
-target_compile_features(${TARGET} PUBLIC c_std_11 cxx_std_17)
@@ -1,295 +0,0 @@
-# Example
-
-The following example starts `sd-server` with a standalone diffusion model, VAE, and LLM text encoder:
-
-```
-.\bin\Release\sd-server.exe --diffusion-model  ..\models\diffusion_models\z_image_turbo_bf16.safetensors --vae ..\models\vae\ae.sft  --llm ..\models\text_encoders\qwen_3_4b.safetensors --diffusion-fa --offload-to-cpu -v --cfg-scale 1.0
-```
-
-What this example does:
-
-* `--diffusion-model` selects the standalone diffusion model
-* `--vae` selects the VAE decoder
-* `--llm` selects the text encoder / language model used by this pipeline
-* `--diffusion-fa` enables flash attention in the diffusion model
-* `--offload-to-cpu` reduces VRAM pressure by keeping weights in RAM when possible
-* `-v` enables verbose logging
-* `--cfg-scale 1.0` sets the default CFG scale for generation
-
-After the server starts successfully:
-
-* the web UI is available at `http://127.0.0.1:1234/`
-* the native async API is available under `/sdcpp/v1/...`
-* the compatibility APIs are available under `/v1/...` and `/sdapi/v1/...`
-
-If you want to use a different host or port, pass:
-
-```bash
--listen-ip <ip> --listen-port <port>
-```
-
-# Frontend
-
-## Build with Frontend
-
-The server can optionally build the web frontend and embed it into the binary as `gen_index_html.h`.
-
-### Requirements
-
-Install the following tools:
-
-* **Node.js** ≥ 20
-  https://nodejs.org/
-
-* **pnpm** ≥ 10
-  Install via npm:
-
-```bash
-npm install -g pnpm
-```
-
-Verify installation:
-
-```bash
-node -v
-pnpm -v
-```
-
-### Install frontend dependencies
-
-Go to the frontend directory and install dependencies:
-
-```bash
-cd examples/server/frontend
-pnpm install
-```
-
-### Build the server with CMake
-
-Enable the frontend build option when configuring CMake:
-
-```bash
-cmake -B build -DSD_SERVER_BUILD_FRONTEND=ON
-cmake --build build --config Release
-```
-
-If `pnpm` is available, the build system will automatically run:
-
-```
-pnpm run build
-pnpm run build:header
-```
-
-and embed the generated frontend into the server binary.
-
-## Frontend Repository
-
-The web frontend is maintained in a **separate repository**, https://github.com/leejet/sdcpp-webui.
-
-If you want to modify the UI or frontend logic, please submit pull requests to the **frontend repository**.
-
-This repository (`stable-diffusion.cpp`) only vendors the frontend periodically. Changes from the frontend repo are synchronized:
-
-* approximately **every 1–2 weeks**, or
-* when there are **major frontend updates**
-
-Because of this, frontend changes will **not appear here immediately** after being merged upstream.
-
-## Using an external frontend
-
-By default, the server uses the **embedded frontend** generated during the build (`gen_index_html.h`).
-
-You can also serve a custom frontend file instead of the embedded one by using:
-
-```bash
--serve-html-path <path-to-index.html>
-```
-
-For example:
-
-```bash
-sd-server --serve-html-path ./index.html
-```
-
-In this case, the server will load and serve the specified `index.html` file instead of the embedded frontend. This is useful when:
-
-* developing or testing frontend changes
-* using a custom UI
-* avoiding rebuilding the binary after frontend modifications
-
-# Run
-
-```
-usage: ./bin/sd-server  [options]
-
-Svr Options:
-  -l, --listen-ip <string>      server listen ip (default: 127.0.0.1)
-  --serve-html-path <string>    path to HTML file to serve at root (optional)
-  --listen-port <int>           server listen port (default: 1234)
-  -v, --verbose                 print extra info
-  --color                       colors the logging tags according to level
-  -h, --help                    show this help message and exit
-
-Context Options:
-  -m, --model <string>                     path to full model
-  --clip_l <string>                        path to the clip-l text encoder
-  --clip_g <string>                        path to the clip-g text encoder
-  --clip_vision <string>                   path to the clip-vision encoder
-  --t5xxl <string>                         path to the t5xxl text encoder
-  --llm <string>                           path to the llm text encoder. For example: (qwenvl2.5 for qwen-image,
-                                           mistral-small3.2 for flux2, ...)
-  --llm_vision <string>                    path to the llm vit
-  --qwen2vl <string>                       alias of --llm. Deprecated.
-  --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
-  --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
-  --control-net <string>                   path to control net model
-  --embd-dir <string>                      embeddings directory
-  --lora-model-dir <string>                lora model directory
-  --hires-upscalers-dir <string>           highres fix upscaler model directory
-  --tensor-type-rules <string>             weight type per tensor pattern (example: "^vae\.=f16,model\.=q8_0")
-  --photo-maker <string>                   path to PHOTOMAKER model
-  --upscale-model <string>                 path to esrgan model.
-  -t, --threads <int>                      number of threads to use during computation (default: -1). If threads <= 0,
-                                           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; -1 auto-detects free VRAM minus 1 GiB
-  --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
-  --mmap                                   whether to memory-map model
-  --control-net-cpu                        keep controlnet in cpu (for low vram)
-  --clip-on-cpu                            keep clip in cpu (for low vram)
-  --vae-on-cpu                             keep vae in cpu (for low vram)
-  --fa                                     use flash attention
-  --diffusion-fa                           use flash attention in the diffusion model only
-  --diffusion-conv-direct                  use ggml_conv2d_direct in the diffusion model
-  --vae-conv-direct                        use ggml_conv2d_direct in the vae model
-  --circular                               enable circular padding for convolutions
-  --circularx                              enable circular RoPE wrapping on x-axis (width) only
-  --circulary                              enable circular RoPE wrapping on y-axis (height) only
-  --chroma-disable-dit-mask                disable dit mask for chroma
-  --qwen-image-zero-cond-t                 enable zero_cond_t for qwen image
-  --chroma-enable-t5-mask                  enable t5 mask for chroma
-  --type                                   weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K,
-                                           q4_K). If not specified, the default is the type of the weight file
-  --rng                                    RNG, one of [std_default, cuda, cpu], default: cuda(sd-webui), cpu(comfyui)
-  --sampler-rng                            sampler RNG, one of [std_default, cuda, cpu]. If not specified, use --rng
-  --prediction                             prediction type override, one of [eps, v, edm_v, sd3_flow, flux_flow,
-                                           flux2_flow]
-  --lora-apply-mode                        the way to apply LoRA, one of [auto, immediately, at_runtime], default is
-                                           auto. In auto mode, if the model weights contain any quantized parameters,
-                                           the at_runtime mode will be used; otherwise, immediately will be used.The
-                                           immediately mode may have precision and compatibility issues with quantized
-                                           parameters, but it usually offers faster inference speed and, in some cases,
-                                           lower memory usage. The at_runtime mode, on the other hand, is exactly the
-                                           opposite.
-
-Default Generation Options:
-  -p, --prompt <string>                    the prompt to render
-  -n, --negative-prompt <string>           the negative prompt (default: "")
-  -i, --init-img <string>                  path to the init image
-  --end-img <string>                       path to the end image, required by flf2v
-  --mask <string>                          path to the mask image
-  --control-image <string>                 path to control image, control net
-  --control-video <string>                 path to control video frames, It must be a directory path. The video frames
-                                           inside should be stored as images in lexicographical (character) order. For
-                                           example, if the control video path is `frames`, the directory contain images
-                                           such as 00.png, 01.png, ... etc.
-  --pm-id-images-dir <string>              path to PHOTOMAKER input id images dir
-  --pm-id-embed-path <string>              path to PHOTOMAKER v2 id embed
-  --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 args, key=value list. lcm supports noise_clip_std,
-                                           noise_scale_start, noise_scale_end; ltx2 supports max_shift, base_shift,
-                                           stretch, terminal; euler_ge supports gamma
-  -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)
-  --high-noise-steps <int>                 (high noise) number of sample steps (default: -1 = auto)
-  --clip-skip <int>                        ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer
-                                           (default: -1). <= 0 represents unspecified, will be 1 for SD1.x, 2 for SD2.x
-  -b, --batch-count <int>                  batch count
-  --video-frames <int>                     video frames (default: 1)
-  --fps <int>                              fps (default: 24)
-  --timestep-shift <int>                   shift timestep for NitroFusion models (default: 0). recommended N for
-                                           NitroSD-Realism around 250 and 500 for NitroSD-Vibrant
-  --upscale-repeats <int>                  Run the ESRGAN upscaler this many times (default: 1)
-  --upscale-tile-size <int>                tile size for ESRGAN upscaling (default: 128)
-  --hires-width <int>                      highres fix target width, 0 to use --hires-scale (default: 0)
-  --hires-height <int>                     highres fix target height, 0 to use --hires-scale (default: 0)
-  --hires-steps <int>                      highres fix second pass sample steps, 0 to reuse --steps (default: 0)
-  --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)
-  --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
-  --skip-layer-start <float>               SLG enabling point (default: 0.01)
-  --skip-layer-end <float>                 SLG disabling point (default: 0.2)
-  --eta <float>                            noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and
-                                           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-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:
-                                           0)
-  --high-noise-skip-layer-start <float>    (high noise) SLG enabling point (default: 0.01)
-  --high-noise-skip-layer-end <float>      (high noise) SLG disabling point (default: 0.2)
-  --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>              
-  --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
-                                           `--high-noise-steps` is set to -1
-  --vace-strength <float>                  wan vace strength
-  --vae-tile-overlap <float>               tile overlap for vae tiling, in fraction of tile size (default: 0.5)
-  --hires-scale <float>                    highres fix scale when target size is not set (default: 2.0)
-  --hires-denoising-strength <float>       highres fix second pass denoising strength (default: 0.7)
-  --increase-ref-index                     automatically increase the indices of references images based on the order
-                                           they are listed (starting with 1).
-  --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
-  --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, 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, 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, ltx2], default:
-                                           model-specific
-  --sigmas                                 custom sigma values for the sampler, comma-separated (e.g.,
-                                           "14.61,7.8,3.5,0.0").
-  --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)
-  --cache-mode                             caching method: 'easycache' (DiT), 'ucache' (UNET),
-                                           'dbcache'/'taylorseer'/'cache-dit' (DiT block-level), 'spectrum' (UNET/DiT
-                                           Chebyshev+Taylor forecasting)
-  --cache-option                           named cache params (key=value format, comma-separated). easycache/ucache:
-                                           threshold=,start=,end=,decay=,relative=,reset=; dbcache/taylorseer/cache-dit:
-                                           Fn=,Bn=,threshold=,warmup=; spectrum: w=,m=,lam=,window=,flex=,warmup=,stop=.
-                                           Examples: "threshold=0.25" or "threshold=1.5,reset=0"
-  --scm-mask                               SCM steps mask for cache-dit: comma-separated 0/1 (e.g.,
-                                           "1,1,1,0,0,1,0,0,1,0") - 1=compute, 0=can cache
-  --scm-policy                             SCM policy: 'dynamic' (default) or 'static'
-  --vae-tile-size                          tile size for vae tiling, format [X]x[Y] (default: 32x32)
-  --vae-relative-tile-size                 relative tile size for vae tiling, format [X]x[Y], in fraction of image size
-                                           if < 1, in number of tiles per dim if >=1 (overrides --vae-tile-size)
-```
@@ -1,356 +0,0 @@
-// Extracted from main.cpp during server refactor.
-
-#include "async_jobs.h"
-
-#include <iomanip>
-#include <sstream>
-
-#include "common/log.h"
-#include "common/media_io.h"
-#include "common/resource_owners.hpp"
-
-const char* async_job_kind_name(AsyncJobKind kind) {
-    switch (kind) {
-        case AsyncJobKind::ImgGen:
-            return "img_gen";
-        case AsyncJobKind::VidGen:
-            return "vid_gen";
-        default:
-            return "img_gen";
-    }
-}
-
-const char* async_job_status_name(AsyncJobStatus status) {
-    switch (status) {
-        case AsyncJobStatus::Queued:
-            return "queued";
-        case AsyncJobStatus::Generating:
-            return "generating";
-        case AsyncJobStatus::Completed:
-            return "completed";
-        case AsyncJobStatus::Failed:
-            return "failed";
-        case AsyncJobStatus::Cancelled:
-            return "cancelled";
-        default:
-            return "failed";
-    }
-}
-
-void purge_expired_jobs(AsyncJobManager& manager) {
-    const int64_t now = unix_timestamp_now();
-
-    for (auto it = manager.expired_jobs.begin(); it != manager.expired_jobs.end();) {
-        if (it->second <= now) {
-            it = manager.expired_jobs.erase(it);
-        } else {
-            ++it;
-        }
-    }
-
-    for (auto it = manager.jobs.begin(); it != manager.jobs.end();) {
-        const auto& job = it->second;
-        if (job->completed_at == 0) {
-            ++it;
-            continue;
-        }
-
-        int64_t ttl_seconds = job->status == AsyncJobStatus::Completed
-                                  ? manager.completed_ttl_seconds
-                                  : manager.failed_ttl_seconds;
-        if (now - job->completed_at >= ttl_seconds) {
-            manager.expired_jobs[job->id] = now + std::max<int64_t>(ttl_seconds, 60);
-            it                            = manager.jobs.erase(it);
-        } else {
-            ++it;
-        }
-    }
-}
-
-size_t count_pending_jobs(const AsyncJobManager& manager) {
-    size_t pending = 0;
-    for (const auto& entry : manager.jobs) {
-        if (entry.second->status == AsyncJobStatus::Queued ||
-            entry.second->status == AsyncJobStatus::Generating) {
-            ++pending;
-        }
-    }
-    return pending;
-}
-
-std::string make_async_job_id(AsyncJobManager& manager) {
-    std::ostringstream oss;
-    oss << "job_" << std::hex << unix_timestamp_now() << "_" << std::setw(8)
-        << std::setfill('0') << manager.next_id++;
-    return oss.str();
-}
-
-bool cancel_queued_job(AsyncJobManager& manager, AsyncGenerationJob& job) {
-    auto new_end = std::remove(manager.queue.begin(), manager.queue.end(), job.id);
-    if (new_end == manager.queue.end()) {
-        return false;
-    }
-
-    manager.queue.erase(new_end, manager.queue.end());
-    job.status       = AsyncJobStatus::Cancelled;
-    job.completed_at = unix_timestamp_now();
-    job.result_images_b64.clear();
-    job.result_media_b64.clear();
-    job.result_media_mime_type.clear();
-    job.result_frame_count = 0;
-    job.result_fps         = 0;
-    job.error_code         = "cancelled";
-    job.error_message      = "job cancelled by client";
-    return true;
-}
-
-json make_async_job_json(const AsyncJobManager& manager, const AsyncGenerationJob& job) {
-    json result;
-    result["id"]             = job.id;
-    result["kind"]           = async_job_kind_name(job.kind);
-    result["status"]         = async_job_status_name(job.status);
-    result["created"]        = job.created_at;
-    result["started"]        = job.started_at == 0 ? json(nullptr) : json(job.started_at);
-    result["completed"]      = job.completed_at == 0 ? json(nullptr) : json(job.completed_at);
-    result["queue_position"] = 0;
-
-    if (job.status == AsyncJobStatus::Queued) {
-        size_t position = 1;
-        for (const auto& queued_id : manager.queue) {
-            if (queued_id == job.id) {
-                result["queue_position"] = position;
-                break;
-            }
-            ++position;
-        }
-    }
-
-    if (job.status == AsyncJobStatus::Completed) {
-        if (job.kind == AsyncJobKind::VidGen) {
-            result["result"] = {
-                {"output_format", job.vid_gen.output_format},
-                {"mime_type", job.result_media_mime_type},
-                {"fps", job.result_fps},
-                {"frame_count", job.result_frame_count},
-                {"b64_json", job.result_media_b64},
-            };
-        } else {
-            json images = json::array();
-            for (size_t i = 0; i < job.result_images_b64.size(); ++i) {
-                images.push_back({{"index", i}, {"b64_json", job.result_images_b64[i]}});
-            }
-            result["result"] = {
-                {"output_format", job.img_gen.output_format},
-                {"images", images},
-            };
-        }
-        result["error"] = nullptr;
-    } else if (job.status == AsyncJobStatus::Failed ||
-               job.status == AsyncJobStatus::Cancelled) {
-        result["result"] = nullptr;
-        result["error"]  = {
-             {"code",
-             job.error_code.empty()
-                  ? (job.status == AsyncJobStatus::Cancelled ? "cancelled" : "generation_failed")
-                  : job.error_code},
-             {"message", job.error_message},
-        };
-    } else {
-        result["result"] = nullptr;
-        result["error"]  = nullptr;
-    }
-
-    return result;
-}
-
-bool execute_img_gen_job(ServerRuntime& runtime,
-                         AsyncGenerationJob& job,
-                         std::vector<std::string>& output_images,
-                         std::string& error_message) {
-    sd_img_gen_params_t params = job.img_gen.to_sd_img_gen_params_t();
-
-    SDImageVec results;
-
-    {
-        std::lock_guard<std::mutex> lock(*runtime.sd_ctx_mutex);
-        sd_image_t* raw_results = generate_image(runtime.sd_ctx, &params);
-        results.adopt(raw_results, params.batch_count);
-    }
-
-    const int num_results = results.count();
-    if (num_results <= 0) {
-        error_message = "generate_image returned no results";
-        return false;
-    }
-
-    EncodedImageFormat encoded_format = EncodedImageFormat::PNG;
-    if (job.img_gen.output_format == "jpeg") {
-        encoded_format = EncodedImageFormat::JPEG;
-    } else if (job.img_gen.output_format == "webp") {
-        encoded_format = EncodedImageFormat::WEBP;
-    }
-
-    for (int i = 0; i < num_results; ++i) {
-        if (results[i].data == nullptr) {
-            continue;
-        }
-
-        const std::string metadata = job.img_gen.gen_params.embed_image_metadata
-                                         ? get_image_params(*runtime.ctx_params,
-                                                            job.img_gen.gen_params,
-                                                            job.img_gen.gen_params.seed + i)
-                                         : "";
-        auto image_bytes           = encode_image_to_vector(encoded_format,
-                                                            results[i].data,
-                                                            results[i].width,
-                                                            results[i].height,
-                                                            results[i].channel,
-                                                            metadata,
-                                                            job.img_gen.output_compression);
-        if (image_bytes.empty()) {
-            continue;
-        }
-        output_images.push_back(base64_encode(image_bytes));
-    }
-
-    if (output_images.empty()) {
-        error_message = "generate_image returned empty encoded outputs";
-        return false;
-    }
-
-    return true;
-}
-
-bool execute_vid_gen_job(ServerRuntime& runtime,
-                         AsyncGenerationJob& job,
-                         std::string& output_media_b64,
-                         std::string& output_media_mime_type,
-                         int& output_frame_count,
-                         int& output_fps,
-                         std::string& error_message) {
-    sd_vid_gen_params_t params = job.vid_gen.to_sd_vid_gen_params_t();
-
-    SDImageVec results;
-    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 = 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;
-    }
-
-    std::vector<uint8_t> video_bytes = create_video_from_sd_images_to_vector(job.vid_gen.output_format,
-                                                                             results.data(),
-                                                                             num_results,
-                                                                             job.vid_gen.gen_params.fps,
-                                                                             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;
-    }
-
-    output_media_b64       = base64_encode(video_bytes);
-    output_media_mime_type = video_mime_type(job.vid_gen.output_format);
-    output_frame_count     = num_results;
-    output_fps             = job.vid_gen.gen_params.fps;
-    return true;
-}
-
-void async_job_worker(ServerRuntime& runtime) {
-    AsyncJobManager& manager = *runtime.async_job_manager;
-
-    while (true) {
-        std::shared_ptr<AsyncGenerationJob> job;
-        {
-            std::unique_lock<std::mutex> lock(manager.mutex);
-            manager.cv.wait(lock, [&]() { return manager.stop || !manager.queue.empty(); });
-
-            if (manager.stop && manager.queue.empty()) {
-                break;
-            }
-
-            purge_expired_jobs(manager);
-            if (manager.queue.empty()) {
-                continue;
-            }
-
-            const std::string job_id = manager.queue.front();
-            manager.queue.pop_front();
-
-            auto it = manager.jobs.find(job_id);
-            if (it == manager.jobs.end()) {
-                continue;
-            }
-
-            job             = it->second;
-            job->status     = AsyncJobStatus::Generating;
-            job->started_at = unix_timestamp_now();
-        }
-
-        std::vector<std::string> output_images;
-        std::string output_media_b64;
-        std::string output_media_mime_type;
-        int output_frame_count = 0;
-        int output_fps         = 0;
-        std::string error_message;
-        bool ok = false;
-
-        if (job->kind == AsyncJobKind::ImgGen) {
-            ok = execute_img_gen_job(runtime, *job, output_images, error_message);
-        } else if (job->kind == AsyncJobKind::VidGen) {
-            ok = execute_vid_gen_job(runtime,
-                                     *job,
-                                     output_media_b64,
-                                     output_media_mime_type,
-                                     output_frame_count,
-                                     output_fps,
-                                     error_message);
-        } else {
-            error_message = "unsupported job kind";
-        }
-
-        {
-            std::lock_guard<std::mutex> lock(manager.mutex);
-            auto it = manager.jobs.find(job->id);
-            if (it == manager.jobs.end()) {
-                continue;
-            }
-
-            job->completed_at = unix_timestamp_now();
-            if (ok) {
-                job->status                 = AsyncJobStatus::Completed;
-                job->result_images_b64      = std::move(output_images);
-                job->result_media_b64       = std::move(output_media_b64);
-                job->result_media_mime_type = std::move(output_media_mime_type);
-                job->result_frame_count     = output_frame_count;
-                job->result_fps             = output_fps;
-                job->error_code.clear();
-                job->error_message.clear();
-            } else {
-                job->status        = AsyncJobStatus::Failed;
-                job->error_code    = "generation_failed";
-                job->error_message = error_message.empty() ? "unknown generation error" : error_message;
-                job->result_images_b64.clear();
-                job->result_media_b64.clear();
-                job->result_media_mime_type.clear();
-                job->result_frame_count = 0;
-                job->result_fps         = 0;
-            }
-
-            purge_expired_jobs(manager);
-        }
-    }
-}
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
leejet	ee620760a2	fix	2025-09-17 21:42:42 +08:00
leejet	aad5d345bb	disable CUDA for quick verification	2025-09-17 21:34:05 +08:00
leejet	61895e4f78	fix	2025-09-17 00:15:05 +08:00
leejet	7d652f317b	fix workflow	2025-09-17 00:12:31 +08:00