feat: add TAEQI2.1 support for Qwen Image 2.1 (#2123)

Co-authored-by: leejet <leejet714@gmail.com>

.github/workflows/build.yml

@@ -451,8 +451,8 @@ jobs:
     runs-on: windows-2022
 
     env:
-      ROCM_VERSION: "7.14.0"
-      GPU_TARGETS: "gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1103;gfx1150;gfx1151;gfx1152;gfx1153;gfx1200;gfx1201"
+      ROCM_VERSION: "10.1.0"
+      gpu_targets: "gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1103;gfx1150;gfx1151;gfx1152;gfx1153;gfx1200;gfx1201"
 
     steps:
       - uses: actions/checkout@v3
@@ -498,7 +498,7 @@ jobs:
 
           write-host "Installing ROCm wheels for multi-arch support"
           # Install ROCm wheels for multi-arch support (this may take several minutes)
-          python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "rocm[libraries,devel]==${{env.ROCM_VERSION}}"
+          python -m pip install --index-url https://stable.repo.amd.com/rocm/whl-next/ "rocm[libraries,devel]==${{env.ROCM_VERSION}}"
 
           # Pre-expand the devel tree so it is included in the cache
           write-host "Initializing ROCm devel tree"
@@ -594,9 +594,9 @@ jobs:
     strategy:
       matrix:
         include:
-          - ROCM_VERSION: "7.14.0"
+          - ROCM_VERSION: "10.1.0"
             gpu_targets: "gfx900;gfx906;gfx908;gfx90a;gfx90c;gfx942;gfx950;gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1103;gfx1150;gfx1151;gfx1152;gfx1153;gfx1200;gfx1201"
-            build: x64
+            build: 'x64'
 
     steps:
       - name: Clone
@@ -610,6 +610,16 @@ jobs:
         with:
           key: ubuntu-rocm-cmake-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
           evict-old-files: 1d
+          max-size: "2G"
+
+      - name: Tune ccache for reinstalled ROCm toolchain
+        run: |
+          # ROCm is pip-installed fresh each run, so the clang binary's mtime
+          # changes every time. With the default compiler_check=mtime that
+          # invalidates the cache; hash compiler contents instead so warm
+          # builds hit.
+          ccache --set-config=compiler_check=content
+          ccache --set-config=sloppiness=time_macros,include_file_mtime,include_file_ctime
 
       - name: Dependencies
         id: depends
@@ -646,7 +656,7 @@ jobs:
           # libraries = HIP runtime and CMake configs needed for linking
           # devel = compilers, headers, static libs
           python -m pip install --upgrade pip
-          python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "rocm[libraries,devel]==${{matrix.ROCM_VERSION}}"
+          python -m pip install --index-url https://stable.repo.amd.com/rocm/whl-next/ "rocm[libraries,devel]==${{matrix.ROCM_VERSION}}"
 
           # Get ROCm installation paths using the rocm-sdk CLI tool
           ROCM_PATH=$(rocm-sdk path --root)

CONTRIBUTING.md

@@ -29,6 +29,10 @@ PRs should include:
 - Linked issue/PR context when applicable.
 - Screenshots or sample outputs for UI/visual behavior changes.
 
+## README UI Listings
+
+The [UIs list in the README](README.md#uis) only includes open-source projects that use `stable-diffusion.cpp` as a backend. Do not submit PRs to add closed-source UIs.
+
 ## Code Style
 
 Format code according to the repository style before submitting changes.

README.md

@@ -67,6 +67,7 @@ API and command-line option may change frequently.***
     - [LLaDA-Image](./docs/llada_image.md)
     - [Ming-Image Design](./docs/ming_image.md)
     - [PixArt](./docs/pixart.md)
+    - [Iris-3B](./docs/iris.md)
   - [Image Edit Models](./docs/edit.md)
     - [FLUX.1-Kontext-dev](./docs/kontext.md)
     - [Qwen Image Edit series](./docs/qwen_image_edit.md)
@@ -182,7 +183,6 @@ These projects wrap `stable-diffusion.cpp` for easier use in other languages/fra
 These projects use `stable-diffusion.cpp` as a backend for their image generation.
 
 - [GIMP Plugins](https://github.com/themanyone/gimp-plugins)
-- [Jellybox](https://jellybox.com)
 - [Stable Diffusion GUI](https://github.com/fszontagh/sd.cpp.gui.wx)
 - [Stable Diffusion CLI-GUI](https://github.com/piallai/stable-diffusion.cpp)
 - [Local Diffusion](https://github.com/rmatif/Local-Diffusion)

docs/iris.md

@@ -0,0 +1,24 @@
+# How to Use
+
+Iris-3B generates images directly in pixel space and uses Qwen3-VL-4B-Instruct as the text encoder. No VAE is required.
+
+## Download weights
+
+- Download Iris-3B
+    - safetensors: https://huggingface.co/speridlabs/iris-3b/tree/main (`model.safetensors` in the root directory)
+- Download Qwen3-VL-4B-Instruct
+    - safetensors: https://huggingface.co/Comfy-Org/Krea-2/tree/main/text_encoders
+    - gguf: https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct-GGUF/tree/main
+
+## Examples
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model ..\models\diffusion_models\iris-3b.safetensors --llm ..\models\text_encoders\Qwen3-VL-4B-Instruct-Q4_K_M.gguf -p "a lovely cat" --cfg-scale 3 -H 1024 -W 1024 --diffusion-fa
+```
+
+<img width="256" alt="iris-3b example" src="../assets/iris/example.png" />
+
+## Notes
+
+- Width and height must be multiples of 16. Do not pass `--vae`.
+- Captions are limited to 300 tokens including the assistant-turn suffix. Positive and negative prompts use the same template.

docs/performance.md

@@ -21,6 +21,16 @@ CPU fallback. It excludes weights and cache buffers. Within a runner lifecycle,
 the summary is printed only on the first graph or when backend capacities or the
 segment count change.
 
+## Run conditional and unconditional CFG in one batched UNet forward.
+
+For UNet models, the conditional and unconditional guidance branches are
+concatenated into a single batch of two and run through one UNet forward per
+step instead of two separate forwards. This is enabled by default whenever the
+run qualifies for it.
+
+Use `--batched-cfg off` to force separate conditional and unconditional
+forwards.
+
 ## Use VAE tiling to reduce encode and decode memory usage.
 
 `--vae-tiling` enables spatial tiling for both VAE encoding and decoding. The

docs/qwen_image_2.1.md

@@ -16,6 +16,8 @@ Qwen Image 2.1 supports text-to-image generation and image editing, using Qwen3-
 
 Use `qwen_image_2.1_vae_bf16.safetensors` with this model. The earlier Qwen Image and Wan 2.2 VAE weights are not interchangeable with the Qwen Image 2.1 VAE weights.
 
+For faster encoding, decoding, or previews, see [TAEQI2.1](taesd.md#qwen-image-21-taeqi21).
+
 ## Examples
 
 Run the following commands from the build directory. Use image dimensions divisible by 32. The resolution-dependent flow schedule is selected automatically.

docs/taesd.md

@@ -37,3 +37,15 @@ sd.cpp also supports [TAEHV](https://github.com/madebyollin/taehv) (#937), which
   ```
 
 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.
+
+### Qwen Image 2.1 (TAEQI2.1)
+
+For Qwen Image 2.1, use [taeqi2_1](https://github.com/madebyollin/taesd), which supports 64-channel latents, 16x spatial scaling, and RGBA images.
+
+Download the official [safetensors weights](https://huggingface.co/madebyollin/taeqi2_1/blob/main/taeqi2_1.safetensors) directly; no conversion is needed:
+
+```bash
+curl -L -o taeqi2_1.safetensors https://huggingface.co/madebyollin/taeqi2_1/resolve/main/taeqi2_1.safetensors
+```
+
+Replace `--vae PATH` with `--taesd taeqi2_1.safetensors` in the [Qwen Image 2.1 examples](qwen_image_2.1.md) to use it for encoding and decoding. For previews only, keep `--vae PATH` and add `--taesd taeqi2_1.safetensors --taesd-preview-only --preview tae`.

examples/common/common.cpp

@@ -378,6 +378,23 @@ static int parse_scale_override(int argc, const char** argv, int index, float& s
     return 1;
 }
 
+static int parse_on_off_arg(int argc, const char** argv, int index, const char* option, bool& value) {
+    if (++index >= argc) {
+        LOG_ERROR("%s requires 'on' or 'off'", option);
+        return -1;
+    }
+    const std::string arg = argv[index];
+    if (arg == "on") {
+        value = true;
+    } else if (arg == "off") {
+        value = false;
+    } else {
+        LOG_ERROR("invalid %s value '%s'; expected 'on' or 'off'", option, argv[index]);
+        return -1;
+    }
+    return 1;
+}
+
 ArgOptions SDContextParams::get_options() {
     ArgOptions options;
     options.string_options = {
@@ -637,23 +654,6 @@ ArgOptions SDContextParams::get_options() {
          true, &vae_conv_direct},
     };
 
-    auto on_auto_fit_arg = [&](int argc, const char** argv, int index) {
-        if (++index >= argc) {
-            LOG_ERROR("--auto-fit requires 'on' or 'off'");
-            return -1;
-        }
-        const std::string arg = argv[index];
-        if (arg == "on") {
-            auto_fit = true;
-        } else if (arg == "off") {
-            auto_fit = false;
-        } else {
-            LOG_ERROR("invalid --auto-fit value '%s'; expected 'on' or 'off'", argv[index]);
-            return -1;
-        }
-        return 1;
-    };
-
     auto on_type_arg = [&](int argc, const char** argv, int index) {
         if (++index >= argc) {
             return -1;
@@ -742,7 +742,15 @@ ArgOptions SDContextParams::get_options() {
          "on|off (default: on). Preserve --backend (otherwise select one GPU) and place weights on the compute GPU, "
          "RAM, another GPU, or disk in that order, according to available memory (--max-vram limits GPU budgets). "
          "Disabled by explicit --params-backend; uses automatic graph segmentation when needed",
-         on_auto_fit_arg},
+         [this](int argc, const char** argv, int index) {
+             return parse_on_off_arg(argc, argv, index, "--auto-fit", auto_fit);
+         }},
+        {"",
+         "--batched-cfg",
+         "on|off (default: on). Run the conditional and unconditional CFG branches in one batched UNet forward when supported",
+         [this](int argc, const char** argv, int index) {
+             return parse_on_off_arg(argc, argv, index, "--batched-cfg", batched_cfg);
+         }},
         {"",
          "--type",
          "weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K, q4_K). "
@@ -940,6 +948,7 @@ std::string SDContextParams::to_string() const {
         << "  max_vram: \"" << max_vram << "\",\n"
         << "  disable_prefetch: " << (disable_prefetch ? "true" : "false") << ",\n"
         << "  disable_segmented_compute: " << (disable_segmented_compute ? "true" : "false") << ",\n"
+        << "  batched_cfg: " << (batched_cfg ? "true" : "false") << ",\n"
         << "  eager_load: " << (eager_load ? "true" : "false") << ",\n"
         << "  backend: \"" << backend << "\",\n"
         << "  params_backend: \"" << params_backend << "\",\n"
@@ -1022,6 +1031,7 @@ sd_ctx_params_t SDContextParams::to_sd_ctx_params_t(bool taesd_preview) {
     sd_ctx_params.max_vram                        = max_vram.c_str();
     sd_ctx_params.disable_prefetch                = disable_prefetch;
     sd_ctx_params.disable_segmented_compute       = disable_segmented_compute;
+    sd_ctx_params.batched_cfg                     = batched_cfg;
     sd_ctx_params.eager_load                      = eager_load;
     sd_ctx_params.backend                         = effective_backend.c_str();
     sd_ctx_params.params_backend                  = effective_params_backend.c_str();

examples/common/common.h

@@ -157,6 +157,7 @@ struct SDContextParams {
     bool disable_prefetch          = false;
     bool disable_segmented_compute = false;
     bool eager_load                = false;
+    bool batched_cfg               = true;
     std::string backend;
     std::string params_backend;
     std::string split_mode;

examples/common/log.cpp

@@ -91,34 +91,34 @@ void log_print(enum sd_log_level_t level, const char* log, sd_log_level_t min_le
     switch (level) {
         case SD_LOG_DEBUG:
             tag_color = 37;
-            level_str = "DEBUG";
+            level_str = "D";
             break;
         case SD_LOG_VERBOSE:
             tag_color = 37;
-            level_str = "VERBOSE";
+            level_str = "V";
             break;
         case SD_LOG_INFO:
             tag_color = 34;
-            level_str = "INFO";
+            level_str = "I";
             break;
         case SD_LOG_WARN:
             tag_color = 35;
-            level_str = "WARN";
+            level_str = "W";
             break;
         case SD_LOG_ERROR:
             tag_color = 31;
-            level_str = "ERROR";
+            level_str = "E";
             break;
         default:
             tag_color = 33;
-            level_str = "?????";
+            level_str = "?";
             break;
     }
 
     if (color) {
-        fprintf(out_stream, "\033[%d;1m[%-7s]\033[0m ", tag_color, level_str);
+        fprintf(out_stream, "\033[%d;1m[%s]\033[0m ", tag_color, level_str);
     } else {
-        fprintf(out_stream, "[%-7s] ", level_str);
+        fprintf(out_stream, "[%s] ", level_str);
     }
     fflush(out_stream);
     print_utf8(out_stream, log);
@@ -131,18 +131,15 @@ void example_log_printf(sd_log_level_t level, const char* file, int line, const
     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_level, log_color);
-
+    char log_buffer[LOG_BUFFER_SIZE + 1] = {};
+    vsnprintf(log_buffer, sizeof(log_buffer), format, args);
     va_end(args);
+
+    std::string message = log_buffer;
+    while (!message.empty() && (message.back() == '\n' || message.back() == '\r')) {
+        message.pop_back();
+    }
+    message += " --- " + sd_basename(file) + ":" + std::to_string(line) + '\n';
+
+    log_print(level, message.c_str(), log_level, log_color);
 }

examples/server/api.md

@@ -37,6 +37,7 @@ Current generation-related endpoints include:
 
 - `POST /sdapi/v1/txt2img`
 - `POST /sdapi/v1/img2img`
+- `GET /sdapi/v1/progress`
 - `GET /sdapi/v1/loras`
 - `GET /sdapi/v1/upscalers`
 - `GET /sdapi/v1/latent-upscale-modes`
@@ -278,6 +279,26 @@ Response fields:
 | `parameters` | `object` | Echo of the parsed outer request body |
 | `info` | `string` | Currently empty string |
 
+#### `GET /sdapi/v1/progress`
+
+Poll this endpoint while a synchronous SDAPI generation request is running.
+An optional `id_task` query parameter selects the task with the same `id_task`
+provided in the generation request (default: `"sdapi"`). Set
+`skip_current_image=true` or `1` to omit the preview image.
+
+`current_image` contains the latest base64 JPEG latent projection, or `null`
+when unavailable. `state.sampling_step` is the positive logical step and
+`state.sampling_steps` is the actual step count for the current sampling pass,
+including schedule and img2img strength adjustments. `state.job_no` is the
+zero-based pass index; `state.job_count` includes one pass per batch image and
+an additional pass per image when highres fix is enabled. `progress` weights
+these passes equally. Sampling progress can reach `1` before final decoding
+finishes; `eta_relative` is currently always `0`.
+
+When no SDAPI task is active, or `id_task` does not match, the endpoint returns
+`progress=0`, `state.job=""`, zero job/step counts, and `current_image=null`.
+Completed and failed requests do not retain an active preview.
+
 #### Discovery / Compatibility Endpoints
 
 Currently exposed:
@@ -410,6 +431,20 @@ Field types:
 | `queue_position` | `integer` |
 | `result` | `object \| null` |
 | `error` | `object \| null` |
+| `preview` | `object \| null` |
+
+`preview` sub-fields:
+
+| Field | Type | Notes |
+| --- | --- | --- |
+| `preview.pass` | `integer` | Sampling pass number, starting at 1 within the job |
+| `preview.step` | `integer` | Positive logical sample step within this pass |
+| `preview.total_steps` | `integer` | Actual sample steps in this pass, after schedule and strength adjustments |
+| `preview.b64_json` | `string` | Base64-encoded PNG of the current preview frame |
+
+The preview is updated in place as sampling progresses; poll `GET /sdcpp/v1/jobs/{id}` to retrieve the latest frame. Only the most recent preview is retained.
+
+`step / total_steps` describes the current sampling pass, not overall job completion. A new batch item, high/low-noise stage, or highres pass increments `pass` and restarts `step`. Samplers with multiple denoiser evaluations per logical step may update a preview more than once at the same step. The frame and its pass/step metadata are published together.
 
 ### Endpoints
 
@@ -437,6 +472,7 @@ Top-level fields:
 | `loras` | `array<object>` | Available LoRA entries |
 | `upscalers` | `array<object>` | Available highres upscalers, built-in and model-backed |
 | `upscale` | `boolean` | Whether a compatible RGB ESRGAN model is available for `POST /sdcpp/v1/upscale` |
+| `preview_modes` | `array<string>` | Available preview modes, e.g. `["none", "proj", "tae", "vae"]` |
 | `limits` | `object` | Shared queue and size limits |
 
 `model`
@@ -598,6 +634,7 @@ Fields returned in `features_by_mode.img_gen`:
 - `cache`
 - `cancel_queued`
 - `cancel_generating`
+- `preview`
 
 Fields returned in `features_by_mode.vid_gen`:
 
@@ -610,6 +647,7 @@ Fields returned in `features_by_mode.vid_gen`:
 - `cache`
 - `cancel_queued`
 - `cancel_generating`
+- `preview`
 
 #### `POST /sdcpp/v1/img_gen`
 
@@ -782,6 +820,9 @@ Example:
   "scm_mask": "",
   "scm_policy_dynamic": true,
 
+  "preview": "none",
+  "preview_interval": 1,
+
   "output_format": "png",
   "output_compression": 100
 }
@@ -847,6 +888,8 @@ Top-level scalar fields:
 | `control_strength` | `number` |
 | `ip_adapter_strength` | `number` |
 | `embed_image_metadata` | `boolean` |
+| `preview` | `string` |
+| `preview_interval` | `integer` |
 
 Image fields:
 
@@ -934,6 +977,10 @@ When omitted, backend defaults apply to these fields:
 - `sample_params.flow_shift`
 - `sample_params.guidance.img_cfg`
 
+### Preview Interval Semantics
+
+`preview_interval` controls the period (in sample steps) at which preview frames are generated. The default is `1` (every step). Any non-positive values will be clamped to `1`. Note that if `preview` is set to `"none"`, no previews are produced regardless of the interval.
+
 ### Completion Result
 
 Example completed job:
@@ -1048,7 +1095,8 @@ Response fields:
 Compared with `img_gen`, the `vid_gen` request body:
 
 - `vid_gen` is a single video sequence job, so `batch_count` is not part of the request schema
-- `ref_images`, `mask_image`, `control_image`, `control_strength`, `ip_adapter_image`, `ip_adapter_strength`, and `embed_image_metadata` are not part of the request schema
+- `mask_image`, `control_image`, `control_strength`, `ip_adapter_image`, `ip_adapter_strength`, and `embed_image_metadata` are not part of the request schema
+- `ref_images` is accepted for MiniMax-H3 Ref2VA conditioning; other video model families currently ignore it
 - `vid_gen` adds `end_image`, `control_frames`, `high_noise_sample_params`, `video_frames`, `fps`, `moe_boundary`, and `vace_strength`
 
 Example:
@@ -1069,6 +1117,7 @@ Example:
 
   "init_image": null,
   "end_image": null,
+  "ref_images": [],
   "control_frames": [],
 
   "sample_params": {
@@ -1130,11 +1179,25 @@ Example:
   "scm_mask": "",
   "scm_policy_dynamic": true,
 
+  "preview": "none",
+  "preview_interval": 1,
+
   "output_format": "webm",
   "output_compression": 100
 }
 ```
 
+### Reference Image Rules
+
+- `ref_images` contains reference images for MiniMax-H3 Ref2VA conditioning.
+- Images retain request order and correspond to `<Picture 1>`, `<Picture 2>`, and so on in the prompt.
+- For MiniMax-H3, non-empty `ref_images` cannot be combined with `init_image` or `end_image`. A conflicting API request fails during generation; the WebUI checks this before submission.
+- MiniMax-H3 does not support `control_frames`; leave that array empty.
+- Other video model families currently ignore `ref_images`.
+- `features_by_mode.vid_gen` does not currently advertise `ref_images`; its absence is not an indication that MiniMax-H3 reference images are unsupported.
+
+See [MiniMax-H3 reference conditioning](../../docs/minimax_h3.md#reference-to-audio-video-conditioning) for model requirements and prompt examples.
+
 ### LoRA Rules
 
 - The server only accepts explicit LoRA entries from the `lora` field.
@@ -1152,6 +1215,7 @@ Channel expectations:
 
 - `init_image`: 3 channels
 - `end_image`: 3 channels
+- `ref_images[]`: decoded with native channels, then converted to RGB by MiniMax-H3
 - `control_frames[]`: 3 channels
 
 Frame ordering rules:
@@ -1181,6 +1245,8 @@ Top-level scalar fields:
 | `fps` | `integer` |
 | `moe_boundary` | `number` |
 | `vace_strength` | `number` |
+| `preview` | `string` |
+| `preview_interval` | `integer` |
 
 Image and frame fields:
 
@@ -1188,6 +1254,7 @@ Image and frame fields:
 | --- | --- |
 | `init_image` | `string \| null` |
 | `end_image` | `string \| null` |
+| `ref_images` | `array<string>` |
 | `control_frames` | `array<string>` |
 
 LoRA fields:

examples/server/async_jobs.cpp

@@ -10,6 +10,55 @@
 #include "common/media_io.h"
 #include "common/resource_owners.hpp"
 
+static void preview_callback(int step, int frame_count, sd_image_t* frames, bool is_noisy, void* data) {
+    (void)is_noisy;
+    if (frame_count <= 0 || frames == nullptr || frames[0].data == nullptr) {
+        return;
+    }
+
+    AsyncGenerationJob* job      = static_cast<AsyncGenerationJob*>(data);
+    const sd_image_t& frame      = frames[0];
+    const sd_preview_info_t info = sd_get_preview_info();
+
+    auto image_bytes = encode_image_to_vector(EncodedImageFormat::PNG,
+                                              frame.data,
+                                              frame.width,
+                                              frame.height,
+                                              frame.channel,
+                                              "",
+                                              80);
+    if (image_bytes.empty()) {
+        return;
+    }
+
+    std::string b64 = base64_encode(image_bytes);
+    std::lock_guard<std::mutex> lock(job->preview_mutex);
+    job->preview_b64         = std::move(b64);
+    job->preview_step        = step < 0 ? -step : step;
+    job->preview_pass        = info.sample_pass;
+    job->preview_total_steps = info.total_steps;
+}
+
+static void clear_preview_callback() {
+    sd_set_preview_callback(nullptr, PREVIEW_NONE, 1, false, false, nullptr);
+}
+
+static void set_preview_callback(AsyncGenerationJob& job) {
+    preview_t mode = str_to_preview(job.kind == AsyncJobKind::ImgGen
+                                        ? job.img_gen.preview_mode.c_str()
+                                        : job.vid_gen.preview_mode.c_str());
+    if (mode == PREVIEW_COUNT || mode == PREVIEW_NONE) {
+        return;
+    }
+    int interval = job.kind == AsyncJobKind::ImgGen
+                       ? job.img_gen.preview_interval
+                       : job.vid_gen.preview_interval;
+    if (interval <= 0) {
+        interval = 1;
+    }
+    sd_set_preview_callback(preview_callback, mode, interval, true, false, &job);
+}
+
 const char* async_job_kind_name(AsyncJobKind kind) {
     switch (kind) {
         case AsyncJobKind::ImgGen:
@@ -161,6 +210,18 @@ json make_async_job_json(const AsyncJobManager& manager, const AsyncGenerationJo
         result["error"]  = nullptr;
     }
 
+    if (job.status == AsyncJobStatus::Generating) {
+        std::lock_guard<std::mutex> lock(job.preview_mutex);
+        if (!job.preview_b64.empty()) {
+            result["preview"] = {
+                {"pass", job.preview_pass},
+                {"step", job.preview_step},
+                {"total_steps", job.preview_total_steps},
+                {"b64_json", job.preview_b64},
+            };
+        }
+    }
+
     return result;
 }
 
@@ -174,12 +235,14 @@ bool execute_img_gen_job(ServerRuntime& runtime,
 
     {
         std::lock_guard<std::mutex> lock(*runtime.sd_ctx_mutex);
+        set_preview_callback(job);
         sd_image_t* raw_results = nullptr;
         int num_results         = 0;
         if (!generate_image(runtime.sd_ctx, &params, &raw_results, &num_results)) {
             raw_results = nullptr;
             num_results = 0;
         }
+        clear_preview_callback();
         results.adopt(raw_results, num_results);
     }
 
@@ -247,10 +310,12 @@ bool execute_vid_gen_job(ServerRuntime& runtime,
 
     {
         std::lock_guard<std::mutex> lock(*runtime.sd_ctx_mutex);
+        set_preview_callback(job);
         sd_image_t* raw_results = nullptr;
         if (!generate_video(runtime.sd_ctx, &params, &raw_results, &num_results, &generated_audio, &output_fps)) {
             raw_results = nullptr;
         }
+        clear_preview_callback();
         results.adopt(raw_results, num_results);
     }
 

examples/server/async_jobs.h

@@ -44,6 +44,11 @@ struct AsyncGenerationJob {
     int result_fps         = 0;
     std::string error_code;
     std::string error_message;
+    mutable std::mutex preview_mutex;
+    std::string preview_b64;
+    int preview_pass          = 0;
+    int preview_step          = 0;
+    int preview_total_steps   = 0;
 };
 
 struct AsyncJobManager {

examples/server/main.cpp

@@ -75,6 +75,8 @@ int main(int argc, const char** argv) {
     }
     SDSvrParams svr_params;
     SDContextParams ctx_params;
+    // Preserve the CLI default while disabling cwd scans for sd-server.
+    ctx_params.lora_model_dir.clear();
     SDGenerationParams default_gen_params;
 
     sd_set_log_callback(sd_log_cb, (void*)&svr_params);

examples/server/routes_sdapi.cpp

@@ -301,6 +301,76 @@ static nlohmann::json prepare_info_field(const SDContextParams& ctx_params,
     return jsoninfo;
 }
 
+struct SDAPIPreviewState {
+    std::mutex mutex;
+    std::string b64;
+    std::string id_task;
+    std::string job_type;
+    int step          = 0;
+    int total         = 0;
+    bool active       = false;
+    int64_t job_count = 0;
+    int job_no        = 0;
+};
+
+static SDAPIPreviewState g_preview;
+
+static void preview_callback(int step, int frame_count, sd_image_t* frames, bool is_noisy, void* data) {
+    (void)is_noisy;
+    (void)data;
+    if (frame_count <= 0 || frames == nullptr || frames[0].data == nullptr) {
+        return;
+    }
+
+    const sd_image_t& frame      = frames[0];
+    const sd_preview_info_t info = sd_get_preview_info();
+
+    auto image_bytes = encode_image_to_vector(EncodedImageFormat::JPEG,
+                                              frame.data,
+                                              frame.width,
+                                              frame.height,
+                                              frame.channel,
+                                              "",
+                                              80);
+    if (image_bytes.empty()) {
+        return;
+    }
+
+    std::string b64 = base64_encode(image_bytes);
+    std::lock_guard<std::mutex> lock(g_preview.mutex);
+    g_preview.b64    = std::move(b64);
+    g_preview.step   = static_cast<int>(std::min<int64_t>(step < 0 ? -static_cast<int64_t>(step) : step,
+                                                        std::max(0, info.total_steps)));
+    g_preview.total  = std::max(0, info.total_steps);
+    g_preview.job_no = std::max(0, info.sample_pass - 1);
+}
+
+struct SDAPIPreviewGuard {
+    SDAPIPreviewGuard(const std::string& id_task, const sd_img_gen_params_t& params, const std::string& job_type) {
+        std::lock_guard<std::mutex> lock(g_preview.mutex);
+        g_preview.id_task  = id_task;
+        g_preview.total    = 0;
+        g_preview.job_type = job_type;
+        g_preview.b64.clear();
+        g_preview.step   = 0;
+        g_preview.active = true;
+        // Each batch item has a base pass and, when enabled, a hires pass.
+        g_preview.job_count = static_cast<int64_t>(std::max(1, params.batch_count)) * (params.hires.enabled ? 2 : 1);
+        g_preview.job_no    = 0;
+        sd_set_preview_callback(preview_callback, PREVIEW_PROJ, 1, true, false, nullptr);
+    }
+
+    ~SDAPIPreviewGuard() {
+        sd_set_preview_callback(nullptr, PREVIEW_NONE, 1, false, false, nullptr);
+        std::lock_guard<std::mutex> lock(g_preview.mutex);
+        g_preview.active = false;
+        g_preview.b64.clear();
+    }
+
+    SDAPIPreviewGuard(const SDAPIPreviewGuard&)            = delete;
+    SDAPIPreviewGuard& operator=(const SDAPIPreviewGuard&) = delete;
+};
+
 void register_sdapi_endpoints(httplib::Server& svr, ServerRuntime& rt) {
     ServerRuntime* runtime = &rt;
 
@@ -317,7 +387,9 @@ void register_sdapi_endpoints(httplib::Server& svr, ServerRuntime& rt) {
                 return;
             }
 
-            json j = json::parse(req.body);
+            std::string id_task = "sdapi";
+            json j              = json::parse(req.body);
+            id_task             = j.value("id_task", id_task);
             ImgGenJobRequest request;
             std::string error_message;
             if (!build_sdapi_img_gen_request(j, *runtime, img2img, request, error_message)) {
@@ -334,6 +406,7 @@ void register_sdapi_endpoints(httplib::Server& svr, ServerRuntime& rt) {
 
             {
                 std::lock_guard<std::mutex> lock(*runtime->sd_ctx_mutex);
+                SDAPIPreviewGuard preview_guard(id_task, img_gen_params, img2img ? "img2img" : "txt2img");
                 sd_image_t* raw_results = nullptr;
                 if (!generate_image(runtime->sd_ctx, &img_gen_params, &raw_results, &num_results)) {
                     raw_results = nullptr;
@@ -528,4 +601,62 @@ void register_sdapi_endpoints(httplib::Server& svr, ServerRuntime& rt) {
         r["sd_model_checkpoint"] = model_path.stem();
         res.set_content(r.dump(), "application/json");
     });
+
+    svr.Get("/sdapi/v1/progress", [](const httplib::Request& req, httplib::Response& res) {
+        bool skip_current_image = false;
+        bool has_id_param       = req.has_param("id_task");
+        std::string req_id_task = has_id_param ? req.get_param_value("id_task") : "";
+
+        if (req.has_param("skip_current_image")) {
+            std::string skip_str = req.get_param_value("skip_current_image");
+            skip_current_image   = (skip_str == "true" || skip_str == "1");
+        }
+
+        json r;
+        r["eta_relative"] = 0.0f;
+        json state;
+        state["skipped"]       = false;
+        state["interrupted"]   = false;
+        state["job_timestamp"] = "0";
+        r["textinfo"]          = nullptr;
+
+        {
+            std::lock_guard<std::mutex> lock(g_preview.mutex);
+
+            bool is_matching_task = g_preview.active && (!has_id_param || req_id_task == g_preview.id_task);
+
+            int step             = is_matching_task ? g_preview.step : 0;
+            int steps            = is_matching_task ? g_preview.total : 0;
+            std::string b64      = is_matching_task ? g_preview.b64 : "";
+            std::string job_type = is_matching_task ? g_preview.job_type : "";
+            int64_t job_count    = is_matching_task ? g_preview.job_count : 0;
+            int job_no           = is_matching_task ? g_preview.job_no : 0;
+
+            float batch_progress = (steps > 0) ? (static_cast<float>(step) / static_cast<float>(steps)) : 0.0f;
+            if (batch_progress > 1.0f)
+                batch_progress = 1.0f;
+
+            float overall_progress = job_count > 0
+                                         ? (static_cast<float>(job_no) + batch_progress) / static_cast<float>(job_count)
+                                         : 0.0f;
+            if (overall_progress > 1.0f)
+                overall_progress = 1.0f;
+
+            r["progress"]           = overall_progress;
+            state["job"]            = job_type;
+            state["sampling_step"]  = step;
+            state["sampling_steps"] = steps;
+            state["job_count"]      = job_count;
+            state["job_no"]         = job_no;
+            r["state"]              = state;
+
+            if (!skip_current_image && !b64.empty()) {
+                r["current_image"] = b64;
+            } else {
+                r["current_image"] = nullptr;
+            }
+        }
+
+        res.set_content(r.dump(), "application/json");
+    });
 }

examples/server/routes_sdcpp.cpp

@@ -205,6 +205,7 @@ static json make_img_gen_features_json() {
         {"cache", true},
         {"cancel_queued", true},
         {"cancel_generating", false},
+        {"preview", true},
     };
 }
 
@@ -219,9 +220,18 @@ static json make_vid_gen_features_json() {
         {"cache", true},
         {"cancel_queued", true},
         {"cancel_generating", false},
+        {"preview", true},
     };
 }
 
+static json make_preview_modes_json() {
+    json modes = json::array();
+    for (int i = 0; i < PREVIEW_COUNT; ++i) {
+        modes.push_back(sd_preview_name((preview_t)i));
+    }
+    return modes;
+}
+
 static json make_capabilities_json(ServerRuntime& runtime) {
     refresh_lora_cache(runtime);
     refresh_upscaler_cache(runtime);
@@ -396,6 +406,7 @@ static json make_capabilities_json(ServerRuntime& runtime) {
     result["loras"]                  = available_loras;
     result["upscalers"]              = available_upscalers;
     result["upscale"]                = have_upscaler_models;
+    result["preview_modes"]          = make_preview_modes_json();
     return result;
 }
 
@@ -418,6 +429,8 @@ static bool parse_img_gen_request(const json& body,
     if (!assign_output_options(request, output_format, output_compression, true, error_message)) {
         return false;
     }
+    request.preview_mode     = body.value("preview", std::string("none"));
+    request.preview_interval = body.value("preview_interval", 1);
     // Intentionally disable prompt-embedded LoRA tag parsing for server APIs.
     if (!request.gen_params.resolve_and_validate(IMG_GEN, "", runtime.ctx_params->hires_upscalers_dir, true)) {
         error_message = "invalid generation parameters";
@@ -445,6 +458,8 @@ static bool parse_vid_gen_request(const json& body,
     if (!assign_output_options(request, output_format, output_compression, error_message)) {
         return false;
     }
+    request.preview_mode     = body.value("preview", std::string("none"));
+    request.preview_interval = body.value("preview_interval", 1);
     // Intentionally disable prompt-embedded LoRA tag parsing for server APIs.
     if (!request.gen_params.resolve_and_validate(VID_GEN, "", runtime.ctx_params->hires_upscalers_dir, true)) {
         error_message = "invalid generation parameters";

examples/server/runtime.cpp

@@ -254,6 +254,12 @@ std::string SDSvrParams::to_string() const {
 void refresh_lora_cache(ServerRuntime& rt) {
     std::vector<LoraEntry> new_cache;
 
+    if (rt.ctx_params->lora_model_dir.empty()) {
+        std::lock_guard<std::mutex> lock(*rt.lora_mutex);
+        rt.lora_cache->clear();
+        return;
+    }
+
     fs::path lora_dir = rt.ctx_params->lora_model_dir;
     std::error_code ec;
     if (fs::exists(lora_dir, ec) && !ec && fs::is_directory(lora_dir, ec) && !ec) {

examples/server/runtime.h

@@ -66,6 +66,8 @@ struct ImgGenJobRequest {
     SDGenerationParams gen_params;
     std::string output_format = "png";
     int output_compression    = 100;
+    std::string preview_mode  = "none";
+    int preview_interval      = 1;
 
     sd_img_gen_params_t to_sd_img_gen_params_t() {
         return gen_params.to_sd_img_gen_params_t();
@@ -76,6 +78,8 @@ struct VidGenJobRequest {
     SDGenerationParams gen_params;
     std::string output_format = "webm";
     int output_compression    = 100;
+    std::string preview_mode  = "none";
+    int preview_interval      = 1;
 
     sd_vid_gen_params_t to_sd_vid_gen_params_t() {
         return gen_params.to_sd_vid_gen_params_t();

include/stable-diffusion.h

@@ -245,6 +245,7 @@ typedef struct {
     const char* rpc_servers;
     const char* model_args;
     bool disable_segmented_compute;  // Force monolithic graph execution even when automatic graph cutting would fit memory better
+    bool batched_cfg;                // Run the conditional and unconditional CFG branches in one batched UNet forward when supported
     float linear_scale;              // Override linear input scaling; 0 keeps the model default
     float attn_scale;                // Override flash-attention K/V scaling; 0 keeps the model default
     const char* tokenizer;           // tokenizer.json path or main=FILE,clip-l=FILE,clip-g=FILE assignments; required for PiD and Lens
@@ -461,6 +462,11 @@ struct ggml_tensor;
 typedef void (*sd_log_cb_t)(enum sd_log_level_t level, const char* text, void* data);
 typedef void (*sd_progress_cb_t)(int step, int steps, float time, void* data);
 typedef void (*sd_preview_cb_t)(int step, int frame_count, sd_image_t* frames, bool is_noisy, void* data);
+
+typedef struct {
+    int sample_pass;
+    int total_steps;
+} sd_preview_info_t;
 typedef bool (*sd_graph_eval_callback_t)(struct ggml_tensor* t, bool ask, void* user_data);
 
 SD_API void sd_set_log_callback(sd_log_cb_t sd_log_cb, void* data);
@@ -469,6 +475,10 @@ SD_API void sd_set_progress_callback(sd_progress_cb_t cb, void* data);
 // negative interval previews only completed logical step -interval. Zero previews the final
 // completed step of the first sampling pass (base-resolution or high-noise).
 SD_API void sd_set_preview_callback(sd_preview_cb_t cb, enum preview_t mode, int interval, bool denoised, bool noisy, void* data);
+// Query from a preview callback. Passes are numbered from 1 since the last
+// sd_set_preview_callback call; total_steps is the actual count in that pass.
+// Both fields are zero before sampling starts.
+SD_API sd_preview_info_t sd_get_preview_info();
 SD_API void sd_set_backend_eval_callback(sd_graph_eval_callback_t cb, void* data);
 SD_API int32_t sd_get_num_physical_cores();
 SD_API const char* sd_get_system_info();

src/conditioning/conditioner.hpp

@@ -399,7 +399,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
                 ss << "['" << item.first << "', " << item.second << "], ";
             }
             ss << "]";
-            LOG_VERBOSE("parse '%s' to %s", text.c_str(), ss.str().c_str());
+            LOG_VERBOSE("parse '%s' to %s", escape_newlines(text).c_str(), escape_newlines(ss.str()).c_str());
         }
 
         auto on_new_token_cb = [&](std::string& str, std::vector<int32_t>& bpe_tokens) -> bool {
@@ -817,7 +817,7 @@ struct SD3CLIPEmbedder : public Conditioner {
                 ss << "['" << item.first << "', " << item.second << "], ";
             }
             ss << "]";
-            LOG_VERBOSE("parse '%s' to %s", text.c_str(), ss.str().c_str());
+            LOG_VERBOSE("parse '%s' to %s", escape_newlines(text).c_str(), escape_newlines(ss.str()).c_str());
         }
 
         auto on_new_token_cb = [&](std::string& str, std::vector<int32_t>& bpe_tokens) -> bool {
@@ -909,6 +909,31 @@ struct SD3CLIPEmbedder : public Conditioner {
 
         size_t chunk_count = std::max(std::max(clip_l_tokens.size(), clip_g_tokens.size()), t5_tokens.size()) / chunk_len;
 
+        // Each active encoder must cover every chunk of the longest token sequence.
+        Tokenizer* tokenizers[] = {
+            clip_l ? clip_l_tokenizer.get() : nullptr,
+            clip_g ? clip_g_tokenizer.get() : nullptr,
+            t5 ? &t5_tokenizer : nullptr,
+        };
+        const size_t target_length = chunk_count * chunk_len;
+
+        for (size_t encoder_idx = 0; encoder_idx < 3; encoder_idx++) {
+            auto& tokens  = token_and_weights[encoder_idx].first;
+            auto& weights = token_and_weights[encoder_idx].second;
+            if (!tokenizers[encoder_idx] || tokens.size() >= target_length) {
+                continue;
+            }
+
+            std::vector<int> empty_tokens;
+            std::vector<float> empty_weights;
+            tokenizers[encoder_idx]->pad_tokens(empty_tokens, &empty_weights, nullptr, chunk_len, chunk_len, false);
+
+            while (tokens.size() < target_length) {
+                tokens.insert(tokens.end(), empty_tokens.begin(), empty_tokens.end());
+                weights.insert(weights.end(), empty_weights.begin(), empty_weights.end());
+            }
+        }
+
         for (int chunk_idx = 0; chunk_idx < chunk_count; chunk_idx++) {
             // clip_l
             sd::Tensor<float> chunk_hidden_states_l;
@@ -1207,7 +1232,7 @@ struct FluxCLIPEmbedder : public Conditioner {
                 ss << "['" << item.first << "', " << item.second << "], ";
             }
             ss << "]";
-            LOG_VERBOSE("parse '%s' to %s", text.c_str(), ss.str().c_str());
+            LOG_VERBOSE("parse '%s' to %s", escape_newlines(text).c_str(), escape_newlines(ss.str()).c_str());
         }
 
         auto on_new_token_cb = [&](std::string& str, std::vector<int32_t>& bpe_tokens) -> bool {
@@ -1484,7 +1509,7 @@ struct T5CLIPEmbedder : public Conditioner {
                 ss << "['" << item.first << "', " << item.second << "], ";
             }
             ss << "]";
-            LOG_VERBOSE("parse '%s' to %s", text.c_str(), ss.str().c_str());
+            LOG_VERBOSE("parse '%s' to %s", escape_newlines(text).c_str(), escape_newlines(ss.str()).c_str());
         }
 
         auto on_new_token_cb = [&](std::string& str, std::vector<int32_t>& bpe_tokens) -> bool {
@@ -1873,7 +1898,7 @@ struct AnimaConditioner : public Conditioner {
                 ss << "['" << item.first << "', " << item.second << "], ";
             }
             ss << "]";
-            LOG_VERBOSE("parse '%s' to %s", text.c_str(), ss.str().c_str());
+            LOG_VERBOSE("parse '%s' to %s", escape_newlines(text).c_str(), escape_newlines(ss.str()).c_str());
         }
 
         std::vector<int> qwen_tokens;
@@ -1987,7 +2012,8 @@ struct LLMEmbedder : public Conditioner {
                    sd_version_is_sefi_image(version) ||
                    sd_version_is_krea2(version) ||
                    sd_version_is_minimax_h3(version) ||
-                   sd_version_is_mage_flow(version)) {
+                   sd_version_is_mage_flow(version) ||
+                   version == VERSION_IRIS) {
             arch = LLM::LLMArch::QWEN3_VL;
         } else if (sd_version_is_z_image(version) || sd_version_is_z_image_l2p(version) || version == VERSION_OVIS_IMAGE || version == VERSION_FLUX2_KLEIN) {
             arch = LLM::LLMArch::QWEN3;
@@ -2113,9 +2139,10 @@ struct LLMEmbedder : public Conditioner {
 
     std::tuple<std::vector<int>, std::vector<float>, std::vector<float>> tokenize(std::string text,
                                                                                   const std::pair<int, int>& attn_range,
-                                                                                  size_t min_length = 0,
-                                                                                  size_t max_length = 100000000,
-                                                                                  bool spell_quotes = false) {
+                                                                                  size_t min_length         = 0,
+                                                                                  size_t max_length         = 100000000,
+                                                                                  bool spell_quotes         = false,
+                                                                                  const std::string& suffix = "") {
         std::vector<std::pair<std::string, float>> parsed_attention;
         if (attn_range.first >= 0 && attn_range.second > 0) {
             if (attn_range.first > 0) {
@@ -2144,7 +2171,7 @@ struct LLMEmbedder : public Conditioner {
                 ss << "['" << item.first << "', " << item.second << "], ";
             }
             ss << "]";
-            LOG_VERBOSE("parse '%s' to %s", text.c_str(), ss.str().c_str());
+            LOG_VERBOSE("parse '%s' to %s", escape_newlines(text).c_str(), escape_newlines(ss.str()).c_str());
         }
 
         std::vector<int> tokens;
@@ -2160,6 +2187,20 @@ struct LLMEmbedder : public Conditioner {
             weights.insert(weights.end(), curr_tokens.size(), curr_weight);
         }
 
+        if (!suffix.empty()) {
+            std::vector<int> suffix_tokens;
+            if (!tokenizer->encode(suffix, suffix_tokens) || suffix_tokens.size() >= max_length) {
+                return {};
+            }
+            // Reserve the assistant-turn marker before truncating the caption.
+            if (tokens.size() > max_length - suffix_tokens.size()) {
+                tokens.resize(max_length - suffix_tokens.size());
+                weights.resize(tokens.size());
+            }
+            tokens.insert(tokens.end(), suffix_tokens.begin(), suffix_tokens.end());
+            weights.insert(weights.end(), suffix_tokens.size(), 1.f);
+        }
+
         std::vector<float> mask;
         tokenizer->pad_tokens(tokens, &weights, &mask, min_length, max_length);
 
@@ -2182,8 +2223,10 @@ struct LLMEmbedder : public Conditioner {
                                     bool spell_quotes                                       = false,
                                     int max_length                                          = 100000000,
                                     const LLM::DeepStackImageEmbeds& deepstack_image_embeds = {},
-                                    const std::vector<LLM::ImageGrid>& image_grids          = {}) {
-        auto tokens_weights_mask = tokenize(prompt, prompt_attn_range, min_length, max_length, spell_quotes);
+                                    const std::vector<LLM::ImageGrid>& image_grids          = {},
+                                    const std::string& prompt_suffix                        = "",
+                                    sd::Tensor<float>* output_mask                          = nullptr) {
+        auto tokens_weights_mask = tokenize(prompt, prompt_attn_range, min_length, max_length, spell_quotes, prompt_suffix);
         auto& tokens             = std::get<0>(tokens_weights_mask);
         auto& weights            = std::get<1>(tokens_weights_mask);
         auto& mask               = std::get<2>(tokens_weights_mask);
@@ -2246,6 +2289,12 @@ struct LLMEmbedder : public Conditioner {
                                                 1);
         }
 
+        if (output_mask != nullptr) {
+            *output_mask = sd::Tensor<float>::zeros({new_hidden_states.shape()[1]});
+            for (size_t i = prompt_template_encode_start_idx; i < mask.size(); ++i) {
+                output_mask->data()[i - prompt_template_encode_start_idx] = mask[i];
+            }
+        }
         return new_hidden_states;
     }
 
@@ -2312,6 +2361,8 @@ struct LLMEmbedder : public Conditioner {
         int hidden_states_min_length         = 0;  // zero pad hidden_states
         bool spell_quotes                    = false;
         std::set<int> out_layers;
+        std::string prompt_suffix;
+        sd::Tensor<float> text_mask;
 
         int64_t t0                     = ggml_time_ms();
         RefImageResizeMode resize_mode = conditioner_params.ref_image_params.vlm_resize_mode;
@@ -3061,6 +3112,23 @@ struct LLMEmbedder : public Conditioner {
             SDCondition result;
             result.c_crossattn = std::move(hidden_states);
             return result;
+        } else if (version == VERSION_IRIS) {
+            prompt =
+                "<|im_start|>system\n"
+                "Describe the image by detailing the color, shape, size, texture, quantity, text, spatial "
+                "relationships of the objects and background:<|im_end|>\n<|im_start|>user\n";
+            auto prefix_tokens = std::get<0>(tokenize(prompt, {0, 0}));
+            if (prefix_tokens.empty()) {
+                return {};
+            }
+            prompt_template_encode_start_idx = static_cast<int>(prefix_tokens.size());
+            hidden_states_min_length         = 300;
+            max_length                       = prompt_template_encode_start_idx + hidden_states_min_length;
+            out_layers                       = {2, 5, 8, 11, 14, 17, 20, 23, 26, 29, 32, 35};
+            prompt_attn_range.first          = static_cast<int>(prompt.size());
+            prompt += conditioner_params.text;
+            prompt_attn_range.second = static_cast<int>(prompt.size());
+            prompt_suffix            = "<|im_end|>\n<|im_start|>assistant\n";
         } else {
             GGML_ABORT("unknown version %d", version);
         }
@@ -3076,7 +3144,9 @@ struct LLMEmbedder : public Conditioner {
                                            spell_quotes,
                                            max_length,
                                            deepstack_image_embeds,
-                                           image_grids);
+                                           image_grids,
+                                           prompt_suffix,
+                                           version == VERSION_IRIS ? &text_mask : nullptr);
         if (hidden_states.empty()) {
             return {};
         }
@@ -3141,6 +3211,7 @@ struct LLMEmbedder : public Conditioner {
         SDCondition result;
         result.c_crossattn        = std::move(hidden_states);
         result.extra_c_crossattns = std::move(extra_hidden_states_vec);
+        result.c_vector           = std::move(text_mask);
         if (version == VERSION_QWEN_IMAGE_2_1) {
             auto slots = sd::Tensor<int32_t>::zeros({result.c_crossattn.shape()[1]});
             for (size_t i = 0; i < image_embeds.size(); ++i) {

src/core/backend_fit.cpp

@@ -497,11 +497,7 @@ namespace sd::backend_fit {
                 return false;
             }
             auto smaller_tile = [&](int size) {
-                int next_size = size / 2;
-                if (!tiling_params.enabled) {
-                    next_size = std::min(next_size, 256 / scale_factor);
-                }
-                return std::min(size, std::max(4, next_size));
+                return std::min(size, std::max(4, size / 2));
             };
             const int tile_size_w = smaller_tile(latent_tile_size_w);
             const int tile_size_h = smaller_tile(latent_tile_size_h);

src/core/ggml_extend.cpp

@@ -223,13 +223,13 @@ ggml_tensor* ggml_ext_linear(ggml_context* ctx,
         x           = ggml_reshape_2d(ctx, x, x->ne[0], x->ne[1] * x->ne[2] * x->ne[3]);
         x           = ggml_mul_mat(ctx, w, x);
         if (force_prec_f32) {
-            ggml_mul_mat_set_prec(x, GGML_PREC_F32);
+            ggml_prec_set_acc(x, GGML_PREC_F32);
         }
         x = ggml_reshape_4d(ctx, x, x->ne[0], x->ne[1] / ne2 / ne3, ne2, ne3);
     } else {
         x = ggml_mul_mat(ctx, w, x);
         if (force_prec_f32) {
-            ggml_mul_mat_set_prec(x, GGML_PREC_F32);
+            ggml_prec_set_acc(x, GGML_PREC_F32);
         }
     }
     if (scale != 1.f) {
@@ -488,7 +488,7 @@ ggml_tensor* ggml_ext_conv_3d(ggml_context* ctx,
         x          = ggml_mul_mat(ctx,
                                   ggml_reshape_2d(ctx, im2col, im2col->ne[0], im2col->ne[3] * im2col->ne[2] * im2col->ne[1]),
                                   ggml_reshape_2d(ctx, w, w->ne[0] * w->ne[1] * w->ne[2] * IC, OC));
-        ggml_mul_mat_set_prec(x, GGML_PREC_F32);
+        ggml_prec_set_acc(x, GGML_PREC_F32);
 
         int64_t OD = im2col->ne[3] / N;
         x          = ggml_reshape_4d(ctx, x, im2col->ne[1] * im2col->ne[2], OD, N, OC);
@@ -683,8 +683,8 @@ ggml_tensor* ggml_ext_attention_ext(ggml_context* ctx,
                               v_in->type == GGML_TYPE_F32 && sd_backend_supports_cuda_mma(backend);
         if (pad_head) {
             // CUDA FA MMA starts at 64 channels; keep the original head's attention scale.
-            q_in = ggml_pad(ctx, q_in, 64 - d_head, 0, 0, 0);
-            k_in = ggml_pad(ctx, k_in, 64 - d_head, 0, 0, 0);
+            q_in = ggml_pad(ctx, q_in, static_cast<int>(64 - d_head), 0, 0, 0);
+            k_in = ggml_pad(ctx, k_in, static_cast<int>(64 - d_head), 0, 0, 0);
         }
         if (kv_scale != 1.0f) {
             k_in = ggml_ext_scale(ctx, k_in, kv_scale);
@@ -694,7 +694,7 @@ ggml_tensor* ggml_ext_attention_ext(ggml_context* ctx,
         v_in = ggml_ext_cont(ctx, ggml_permute(ctx, v_in, 0, 2, 1, 3));
         v_in = ggml_reshape_3d(ctx, v_in, d_head, L_k, n_kv_head * N);
         if (pad_head) {
-            v_in = ggml_pad(ctx, v_in, 64 - d_head, 0, 0, 0);
+            v_in = ggml_pad(ctx, v_in, static_cast<int>(64 - d_head), 0, 0, 0);
         }
         if (kv_scale != 1.0f) {
             v_in = ggml_ext_scale(ctx, v_in, kv_scale);
@@ -721,7 +721,7 @@ ggml_tensor* ggml_ext_attention_ext(ggml_context* ctx,
         if (!ggml_backend_supports_op(backend, out)) {
             return nullptr;
         }
-        ggml_flash_attn_ext_set_prec(out, GGML_PREC_F32);
+        ggml_prec_set_acc(out, GGML_PREC_F32);
         if (kv_scale != 1.0f) {
             out = ggml_ext_scale(ctx, out, 1.0f / kv_scale);
         }
@@ -742,9 +742,9 @@ ggml_tensor* ggml_ext_attention_ext(ggml_context* ctx,
         }
         if (padded_head != d_head) {
             // Keep the original head's softmax scale when padding for the CUDA kernel.
-            q_in = ggml_pad(ctx, q_in, padded_head - d_head, 0, 0, 0);
-            k_in = ggml_pad(ctx, k_in, padded_head - d_head, 0, 0, 0);
-            v_in = ggml_pad(ctx, v_in, padded_head - d_head, 0, 0, 0);
+            q_in = ggml_pad(ctx, q_in, static_cast<int>(padded_head - d_head), 0, 0, 0);
+            k_in = ggml_pad(ctx, k_in, static_cast<int>(padded_head - d_head), 0, 0, 0);
+            v_in = ggml_pad(ctx, v_in, static_cast<int>(padded_head - d_head), 0, 0, 0);
         }
         if (kv_scale != 1.0f) {
             k_in = ggml_ext_scale(ctx, k_in, kv_scale);
@@ -797,7 +797,7 @@ ggml_tensor* ggml_ext_attention_ext(ggml_context* ctx,
         v = ggml_reshape_3d(ctx, v, L_k, d_head, n_kv_head * N);   // [N * n_kv_head, d_head, L_k]
 
         auto kq = ggml_mul_mat(ctx, k, q);  // [N * n_head, L_q, L_k]
-        ggml_mul_mat_set_prec(kq, GGML_PREC_F32);
+        ggml_prec_set_acc(kq, GGML_PREC_F32);
         kq = ggml_scale_inplace(ctx, kq, scale);
         if (mask) {
             kq = ggml_add_inplace(ctx, kq, mask);

src/core/util.cpp

@@ -62,6 +62,21 @@ void replace_all_chars(std::string& str, char target, char replacement) {
     }
 }
 
+std::string escape_newlines(const std::string& text) {
+    std::string escaped;
+    escaped.reserve(text.size());
+    for (char ch : text) {
+        if (ch == '\n') {
+            escaped += "\\n";
+        } else if (ch == '\r') {
+            escaped += "\\r";
+        } else {
+            escaped += ch;
+        }
+    }
+    return escaped;
+}
+
 static std::string sd_vformat(const char* fmt, va_list ap) {
     char small[128];
     va_list ap2;
@@ -360,12 +375,13 @@ int32_t sd_get_num_physical_cores() {
 static sd_progress_cb_t sd_progress_cb = nullptr;
 void* sd_progress_cb_data              = nullptr;
 
-static sd_preview_cb_t sd_preview_cb = nullptr;
-static void* sd_preview_cb_data      = nullptr;
-preview_t sd_preview_mode            = PREVIEW_NONE;
-int sd_preview_interval              = 1;
-bool sd_preview_denoised             = true;
-bool sd_preview_noisy                = false;
+static sd_preview_info_t sd_preview_info = {};
+static sd_preview_cb_t sd_preview_cb     = nullptr;
+static void* sd_preview_cb_data          = nullptr;
+preview_t sd_preview_mode                = PREVIEW_NONE;
+int sd_preview_interval                  = 1;
+bool sd_preview_denoised                 = true;
+bool sd_preview_noisy                    = false;
 
 static sd_graph_eval_callback_t sd_backend_eval_cb = nullptr;
 static void* sd_backend_eval_cb_data               = nullptr;
@@ -644,9 +660,14 @@ void* sd_log_cb_data         = nullptr;
 static void sd_log_dispatch(sd_log_level_t level, const std::string& origin, const std::string& text) {
     if (sd_log_cb == nullptr)
         return;
-    std::string message = origin + " - " + text;
-    if (message.back() != '\n') {
-        message += '\n';
+    std::string message = text;
+    while (!message.empty() && (message.back() == '\n' || message.back() == '\r')) {
+        message.pop_back();
+    }
+    if (origin == "ggml") {
+        message = origin + ": " + message + '\n';
+    } else {
+        message += " --- " + origin + '\n';
     }
     sd_log_cb(level, message.c_str(), sd_log_cb_data);
 }
@@ -656,7 +677,7 @@ void log_printf(sd_log_level_t level, const char* file, int line, const char* fo
     va_start(args, format);
     std::string message = sd_vformat(format, args);
     va_end(args);
-    std::string origin = sd_format("%s:%-4d", sd_basename(file).c_str(), line);
+    std::string origin = sd_format("%s:%d", sd_basename(file).c_str(), line);
     sd_log_dispatch(level, origin, message);
 }
 
@@ -691,6 +712,7 @@ void sd_set_progress_callback(sd_progress_cb_t cb, void* data) {
     sd_progress_cb_data = data;
 }
 void sd_set_preview_callback(sd_preview_cb_t cb, preview_t mode, int interval, bool denoised, bool noisy, void* data) {
+    sd_preview_info     = {};
     sd_preview_cb       = cb;
     sd_preview_cb_data  = data;
     sd_preview_mode     = mode;
@@ -704,6 +726,17 @@ void sd_set_backend_eval_callback(sd_graph_eval_callback_t cb, void* data) {
     sd_backend_eval_cb_data = data;
 }
 
+void sd_begin_preview_pass(int total_steps) {
+    if (sd_preview_cb != nullptr) {
+        ++sd_preview_info.sample_pass;
+        sd_preview_info.total_steps = total_steps;
+    }
+}
+
+sd_preview_info_t sd_get_preview_info() {
+    return sd_preview_info;
+}
+
 sd_preview_cb_t sd_get_preview_callback() {
     return sd_preview_cb;
 }

src/core/util.h

@@ -29,6 +29,7 @@ bool contains(const std::string& str, const std::string& substr);
 std::string sd_format(const char* fmt, ...);
 
 void replace_all_chars(std::string& str, char target, char replacement);
+std::string escape_newlines(const std::string& text);
 
 int round_up_to(int value, int base);
 
@@ -101,6 +102,7 @@ std::vector<std::pair<std::string, float>> split_quotation_attention(
 sd_progress_cb_t sd_get_progress_callback();
 void* sd_get_progress_callback_data();
 
+void sd_begin_preview_pass(int total_steps);
 sd_preview_cb_t sd_get_preview_callback();
 void* sd_get_preview_callback_data();
 preview_t sd_get_preview_mode();

src/model.h

@@ -65,6 +65,7 @@ enum SDVersion {
     VERSION_PIXART,
     VERSION_MING_IMAGE,
     VERSION_Z_IMAGE_L2P,
+    VERSION_IRIS,
     VERSION_COUNT,
 };
 
@@ -334,7 +335,8 @@ static inline bool sd_version_is_dit(SDVersion version) {
         sd_version_is_krea2(version) ||
         sd_version_is_mage_flow(version) ||
         sd_version_is_sensenova_u1(version) ||
-        sd_version_is_pixart(version)) {
+        sd_version_is_pixart(version) ||
+        version == VERSION_IRIS) {
         return true;
     }
     return false;

src/model/adapter/lora_ops.cpp

@@ -11,7 +11,9 @@ ggml_tensor* ggml_ext_merge_lora(ggml_context* ctx,
                                  ggml_tensor* lora_mid) {
     ggml_tensor* updown;
     // flat lora tensors to multiply it
-    int64_t lora_up_rows  = lora_up->ne[ggml_n_dims(lora_up) - 1];
+    // ggml_n_dims drops the output dimension when a linear layer has one output.
+    int lora_up_n_dims    = ggml_n_dims(lora_up);
+    int64_t lora_up_rows  = lora_up_n_dims == 1 ? 1 : lora_up->ne[lora_up_n_dims - 1];
     lora_up               = ggml_reshape_2d(ctx, lora_up, ggml_nelements(lora_up) / lora_up_rows, lora_up_rows);
     auto lora_down_n_dims = ggml_n_dims(lora_down);
     // assume n_dims should always be a multiple of 2 (otherwise rank 1 doesn't work)

src/model/common/rope.hpp

@@ -1067,10 +1067,22 @@ namespace Rope {
         return result;
     }
 
-    __STATIC_INLINE__ ggml_tensor* apply_rope(ggml_context* ctx,
+    __STATIC_INLINE__ ggml_tensor* apply_rope(GGMLRunnerContext* runner_ctx,
                                               ggml_tensor* x,
                                               ggml_tensor* pe,
                                               bool rope_interleaved = true) {
+        auto ctx = runner_ctx->ggml_ctx;
+#ifndef SD_USE_UPSTREAM_GGML
+        auto backend = runner_ctx->backend;
+        if (backend != nullptr && x->type == GGML_TYPE_F32 && pe->type == GGML_TYPE_F32 &&
+            x->ne[0] % 2 == 0 && pe->ne[0] == 2 && pe->ne[1] == 2 &&
+            pe->ne[2] == x->ne[0] / 2 && pe->ne[3] == x->ne[2]) {
+            auto out = ggml_rope_apply(ctx, x, pe, rope_interleaved);
+            if (ggml_backend_supports_op(backend, out)) {
+                return out;
+            }
+        }
+#endif
         // x: [N, L, n_head, d_head]
         // pe: [L, d_head/2, 2, 2], [[cos, -sin], [sin, cos]]
         int64_t d_head = x->ne[0];
@@ -1126,8 +1138,8 @@ namespace Rope {
         // return: [N, L, n_head*d_head]
         int64_t n_head = q->ne[1];
 
-        q = apply_rope(ctx->ggml_ctx, q, pe, rope_interleaved);  // [N*n_head, L, d_head]
-        k = apply_rope(ctx->ggml_ctx, k, pe, rope_interleaved);  // [N*n_head, L, d_head]
+        q = apply_rope(ctx, q, pe, rope_interleaved);  // [N*n_head, L, d_head]
+        k = apply_rope(ctx, k, pe, rope_interleaved);  // [N*n_head, L, d_head]
 
         auto x = ggml_ext_attention_ext(ctx, q, k, v, n_head, mask, true, ctx->flash_attn_enabled, kv_scale);  // [N, L, n_head*d_head]
         return x;

src/model/diffusion/anima.hpp

@@ -235,8 +235,8 @@ namespace Anima {
                 if (pe_k == nullptr) {
                     pe_k = pe_q;
                 }
-                auto q_rope = Rope::apply_rope(ctx->ggml_ctx, q4, pe_q, false);
-                auto k_rope = Rope::apply_rope(ctx->ggml_ctx, k4, pe_k, false);
+                auto q_rope = Rope::apply_rope(ctx, q4, pe_q, false);
+                auto k_rope = Rope::apply_rope(ctx, k4, pe_k, false);
                 attn_out    = ggml_ext_attention_ext(ctx,
                                                      q_rope,
                                                      k_rope,

src/model/diffusion/iris.hpp

@@ -0,0 +1,450 @@
+#ifndef __SD_MODEL_DIFFUSION_IRIS_HPP__
+#define __SD_MODEL_DIFFUSION_IRIS_HPP__
+
+#include <array>
+#include <stdexcept>
+#include "model/diffusion/pid.hpp"
+
+namespace Iris {
+    struct IrisConfig {
+        int64_t hidden_size    = 2560;
+        int64_t depth          = 24;
+        int64_t dual_depth     = 8;
+        int64_t head_dim       = 128;
+        int64_t num_heads      = 20;
+        int64_t kv_heads       = 5;
+        int64_t mlp_dim        = 6826;
+        int64_t patch_size     = 16;
+        int64_t text_dim       = 2560;
+        int64_t text_len       = 300;
+        int64_t text_layers    = 12;
+        int64_t layer_heads    = 32;
+        int64_t layer_mlp_dim  = 3328;
+        int64_t pixel_dim      = 16;
+        int64_t pixel_attn_dim = 1280;
+        int64_t pixel_heads    = 10;
+        int64_t pixel_depth    = 4;
+
+        static IrisConfig detect_from_weights(const String2TensorStorage& weights, const std::string& prefix) {
+            IrisConfig config;
+            auto find = [&](const std::string& name) -> const TensorStorage& {
+                auto it = weights.find(prefix + "." + name);
+                if (it == weights.end()) {
+                    throw std::runtime_error("Iris-3B: missing weight " + prefix + "." + name);
+                }
+                return it->second;
+            };
+            config.hidden_size    = find("s_embedder.proj.weight").ne[1];
+            config.patch_size     = static_cast<int64_t>(std::sqrt(find("s_embedder.proj.weight").ne[0] / 3));
+            config.head_dim       = find("blocks.0.attn.q_norm_x.weight").ne[0];
+            config.num_heads      = config.hidden_size / config.head_dim;
+            config.kv_heads       = find("blocks.0.attn.k_proj_x.weight").ne[1] / config.head_dim;
+            config.mlp_dim        = find("blocks.0.mlp_x.w1.weight").ne[1];
+            config.text_dim       = find("y_embedder.refiner.proj.weight").ne[0];
+            config.text_len       = find("y_pos_embedding").ne[1];
+            config.text_layers    = find("y_embedder.layer_pool.weight").ne[0];
+            config.layer_mlp_dim  = find("y_embedder.layer_blocks.0.mlp.0.weight").ne[1];
+            config.pixel_dim      = find("pixel_embedder.proj.weight").ne[1];
+            config.pixel_attn_dim = find("pixel_blocks.0.compress.weight").ne[1];
+            config.pixel_heads    = config.pixel_attn_dim / find("pixel_blocks.0.attn.q_norm.weight").ne[0];
+            config.depth = config.dual_depth = config.pixel_depth = 0;
+            for (const auto& [name, tensor] : weights) {
+                if (!starts_with(name, prefix + ".")) {
+                    continue;
+                }
+                auto parts = split_string(name.substr(prefix.size() + 1), '.');
+                if (parts.size() > 2 && parts[0] == "blocks") {
+                    int64_t index = std::stoi(parts[1]);
+                    config.depth  = std::max(config.depth, index + 1);
+                    if (parts[2] == "adaln_img") {
+                        config.dual_depth = std::max(config.dual_depth, index + 1);
+                    }
+                } else if (parts.size() > 2 && parts[0] == "pixel_blocks") {
+                    config.pixel_depth = std::max(config.pixel_depth, int64_t(std::stoi(parts[1]) + 1));
+                }
+            }
+            if (config.text_layers != 12 || config.text_dim != 2560 || config.text_len != 300 ||
+                find("modulation_cores.adaln_img.weight").ne[1] != 6 * config.hidden_size ||
+                find("pixel_blocks.0.adaln.weight").ne[1] != 4 * config.pixel_dim * config.patch_size * config.patch_size) {
+                throw std::runtime_error("Iris-3B: unsupported checkpoint configuration");
+            }
+            LOG_VERBOSE("iris: hidden_size = %" PRId64 ", depth = %" PRId64 ", dual_depth = %" PRId64 ", heads = %" PRId64 "/%" PRId64 ", pixel_depth = %" PRId64,
+                        config.hidden_size, config.depth, config.dual_depth, config.num_heads, config.kv_heads, config.pixel_depth);
+            return config;
+        }
+    };
+
+    struct SelfAttention : public GGMLBlock {
+        int64_t dim;
+        int64_t heads;
+        bool qk_norm;
+
+        SelfAttention(int64_t dim, int64_t heads, bool qk_norm = true)
+            : dim(dim), heads(heads), qk_norm(qk_norm) {
+            blocks["qkv"]  = std::make_shared<Linear>(dim, 3 * dim, false);
+            blocks["proj"] = std::make_shared<Linear>(dim, dim, true);
+            if (qk_norm) {
+                blocks["q_norm"] = std::make_shared<RMSNorm>(dim / heads, 1e-6f);
+                blocks["k_norm"] = std::make_shared<RMSNorm>(dim / heads, 1e-6f);
+            }
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* pe = nullptr, ggml_tensor* mask = nullptr) {
+            auto c   = ctx->ggml_ctx;
+            auto qkv = split_qkv(c, std::dynamic_pointer_cast<Linear>(blocks["qkv"])->forward(ctx, x));
+            auto q   = ggml_reshape_4d(c, qkv[0], dim / heads, heads, x->ne[1], x->ne[2]);
+            auto k   = ggml_reshape_4d(c, qkv[1], dim / heads, heads, x->ne[1], x->ne[2]);
+            auto v   = ggml_reshape_4d(c, qkv[2], dim / heads, heads, x->ne[1], x->ne[2]);
+            if (qk_norm) {
+                q = std::dynamic_pointer_cast<RMSNorm>(blocks["q_norm"])->forward(ctx, q);
+                k = std::dynamic_pointer_cast<RMSNorm>(blocks["k_norm"])->forward(ctx, k);
+            }
+            if (pe) {
+                x = Rope::attention(ctx, q, k, v, pe, mask);
+            } else {
+                q = ggml_reshape_3d(c, q, dim, x->ne[1], x->ne[2]);
+                k = ggml_reshape_3d(c, k, dim, x->ne[1], x->ne[2]);
+                v = ggml_reshape_3d(c, v, dim, x->ne[1], x->ne[2]);
+                x = ggml_ext_attention_ext(ctx, q, k, v, heads, mask, false, ctx->flash_attn_enabled);
+            }
+            return std::dynamic_pointer_cast<Linear>(blocks["proj"])->forward(ctx, x);
+        }
+    };
+
+    struct TextBlock : public GGMLBlock {
+        bool layerwise;
+
+        TextBlock(int64_t dim, int64_t heads, int64_t mlp_dim, bool layerwise)
+            : layerwise(layerwise) {
+            blocks["norm1"] = std::make_shared<RMSNorm>(dim, 1e-6f);
+            blocks["norm2"] = std::make_shared<RMSNorm>(dim, 1e-6f);
+            blocks["attn"]  = std::make_shared<SelfAttention>(dim, heads, !layerwise);
+            if (layerwise) {
+                blocks["mlp.0"] = std::make_shared<Linear>(dim, mlp_dim, true);
+                blocks["mlp.2"] = std::make_shared<Linear>(mlp_dim, dim, true);
+            } else {
+                blocks["mlp"] = std::make_shared<Pid::FeedForward>(dim, mlp_dim);
+            }
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* mask = nullptr) {
+            auto h = std::dynamic_pointer_cast<RMSNorm>(blocks["norm1"])->forward(ctx, x);
+            h      = std::dynamic_pointer_cast<SelfAttention>(blocks["attn"])->forward(ctx, h, nullptr, mask);
+            x      = ggml_add(ctx->ggml_ctx, x, h);
+            h      = std::dynamic_pointer_cast<RMSNorm>(blocks["norm2"])->forward(ctx, x);
+            if (layerwise) {
+                h = std::dynamic_pointer_cast<Linear>(blocks["mlp.0"])->forward(ctx, h);
+                h = ggml_silu(ctx->ggml_ctx, h);
+                h = std::dynamic_pointer_cast<Linear>(blocks["mlp.2"])->forward(ctx, h);
+            } else {
+                h = std::dynamic_pointer_cast<Pid::FeedForward>(blocks["mlp"])->forward(ctx, h);
+            }
+            return ggml_add(ctx->ggml_ctx, x, h);
+        }
+    };
+
+    struct TextEmbedder : public GGMLBlock {
+        IrisConfig config;
+
+        TextEmbedder(const IrisConfig& config)
+            : config(config) {
+            for (int i = 0; i < 2; ++i) {
+                blocks["layer_blocks." + std::to_string(i)]   = std::make_shared<TextBlock>(config.text_dim, config.layer_heads, config.layer_mlp_dim, true);
+                blocks["refiner.blocks." + std::to_string(i)] = std::make_shared<TextBlock>(config.hidden_size, config.num_heads, config.mlp_dim, false);
+            }
+            blocks["layer_pool"]   = std::make_shared<Linear>(config.text_layers, 1, true);
+            blocks["refiner.proj"] = std::make_shared<Linear>(config.text_dim, config.hidden_size, true);
+            blocks["refiner.norm"] = std::make_shared<RMSNorm>(config.hidden_size, 1e-6f);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* mask) {
+            auto c         = ctx->ggml_ctx;
+            int64_t tokens = x->ne[1];
+            int64_t batch  = x->ne[2];
+            // The encoder concatenates selected layers within each token's feature dimension.
+            x = ggml_reshape_3d(c, x, config.text_dim, config.text_layers, tokens * batch);
+            for (int i = 0; i < 2; ++i) {
+                x = std::dynamic_pointer_cast<TextBlock>(blocks["layer_blocks." + std::to_string(i)])->forward(ctx, x);
+            }
+            x = ggml_cont(c, ggml_permute(c, x, 1, 0, 2, 3));
+            x = std::dynamic_pointer_cast<Linear>(blocks["layer_pool"])->forward(ctx, x);
+            x = ggml_reshape_3d(c, x, config.text_dim, tokens, batch);
+            x = std::dynamic_pointer_cast<Linear>(blocks["refiner.proj"])->forward(ctx, x);
+            for (int i = 0; i < 2; ++i) {
+                x = std::dynamic_pointer_cast<TextBlock>(blocks["refiner.blocks." + std::to_string(i)])->forward(ctx, x, mask);
+            }
+            return std::dynamic_pointer_cast<RMSNorm>(blocks["refiner.norm"])->forward(ctx, x);
+        }
+    };
+
+    struct TrunkBlock : public GGMLBlock {
+        IrisConfig config;
+        bool dual;
+
+        TrunkBlock(const IrisConfig& config, bool dual)
+            : config(config), dual(dual) {
+            int64_t d = config.hidden_size;
+            for (const std::string stream : dual ? std::vector<std::string>{"x", "y"} : std::vector<std::string>{""}) {
+                std::string suffix                                = stream.empty() ? "" : "_" + stream;
+                std::string attn                                  = dual ? "attn." : "";
+                blocks[attn + "q_proj" + suffix]                  = std::make_shared<Linear>(d, d, false);
+                blocks[attn + "k_proj" + suffix]                  = std::make_shared<Linear>(d, config.kv_heads * config.head_dim, false);
+                blocks[attn + "v_proj" + suffix]                  = std::make_shared<Linear>(d, config.kv_heads * config.head_dim, false);
+                blocks[attn + "q_norm" + suffix]                  = std::make_shared<RMSNorm>(config.head_dim, 1e-6f);
+                blocks[attn + "k_norm" + suffix]                  = std::make_shared<RMSNorm>(config.head_dim, 1e-6f);
+                blocks[dual ? "attn.proj" + suffix : "attn_proj"] = std::make_shared<Linear>(d, d, true);
+                blocks["attn_gate" + suffix]                      = std::make_shared<Linear>(d, d, false);
+                blocks["norm" + suffix + "1"]                     = std::make_shared<RMSNorm>(d, 1e-6f);
+                blocks["norm" + suffix + "2"]                     = std::make_shared<RMSNorm>(d, 1e-6f);
+                blocks["attn_post_norm" + suffix]                 = std::make_shared<RMSNorm>(d, 1e-6f);
+                blocks["mlp_post_norm" + suffix]                  = std::make_shared<RMSNorm>(d, 1e-6f);
+                blocks["mlp" + suffix]                            = std::make_shared<Pid::FeedForward>(d, config.mlp_dim);
+            }
+        }
+
+        void init_params(ggml_context* ctx, const String2TensorStorage& = {}, const std::string = "") override {
+            if (dual) {
+                params["adaln_img.bias"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 6 * config.hidden_size);
+                params["adaln_txt.bias"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 6 * config.hidden_size);
+            } else {
+                params["adaln.bias"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 6 * config.hidden_size);
+            }
+        }
+
+        ggml_tensor* norm(GGMLRunnerContext* ctx, const std::string& name, ggml_tensor* x) {
+            return std::dynamic_pointer_cast<RMSNorm>(blocks[name])->forward(ctx, x);
+        }
+
+        ggml_tensor* linear(GGMLRunnerContext* ctx, const std::string& name, ggml_tensor* x) {
+            return std::dynamic_pointer_cast<Linear>(blocks[name])->forward(ctx, x);
+        }
+
+        std::array<ggml_tensor*, 3> project(GGMLRunnerContext* ctx, ggml_tensor* h, const std::string& suffix) {
+            auto c        = ctx->ggml_ctx;
+            std::string a = dual ? "attn." : "";
+            auto q        = linear(ctx, a + "q_proj" + suffix, h);
+            auto k        = linear(ctx, a + "k_proj" + suffix, h);
+            auto v        = linear(ctx, a + "v_proj" + suffix, h);
+            q             = ggml_reshape_4d(c, q, config.head_dim, config.num_heads, h->ne[1], h->ne[2]);
+            k             = ggml_reshape_4d(c, k, config.head_dim, config.kv_heads, h->ne[1], h->ne[2]);
+            v             = ggml_reshape_4d(c, v, config.head_dim, config.kv_heads, h->ne[1], h->ne[2]);
+            return {norm(ctx, a + "q_norm" + suffix, q), norm(ctx, a + "k_norm" + suffix, k), v};
+        }
+
+        ggml_tensor* finish(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* h, ggml_tensor* out, const std::vector<ggml_tensor*>& mods, const std::string& suffix) {
+            auto c = ctx->ggml_ctx;
+            out    = ggml_mul(c, out, ggml_sigmoid(c, linear(ctx, "attn_gate" + suffix, h)));
+            out    = linear(ctx, dual ? "attn.proj" + suffix : "attn_proj", out);
+            out    = norm(ctx, "attn_post_norm" + suffix, out);
+            x      = ggml_add(c, x, ggml_mul(c, out, mods[2]));
+            h      = Pid::apply_adaln(c, norm(ctx, "norm" + suffix + "2", x), mods[3], mods[4]);
+            h      = std::dynamic_pointer_cast<Pid::FeedForward>(blocks["mlp" + suffix])->forward(ctx, h);
+            h      = norm(ctx, "mlp_post_norm" + suffix, h);
+            return ggml_add(c, x, ggml_mul(c, h, mods[5]));
+        }
+
+        std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* y, ggml_tensor* img_mod, ggml_tensor* txt_mod, ggml_tensor* pe) {
+            auto c     = ctx->ggml_ctx;
+            int64_t nt = y->ne[1];
+            auto mx    = ggml_ext_chunk(c, ggml_add(c, img_mod, params[dual ? "adaln_img.bias" : "adaln.bias"]), 6, 0);
+            if (!dual) {
+                auto tokens = ggml_concat(c, y, x, 1);
+                auto h      = Pid::apply_adaln(c, norm(ctx, "norm1", tokens), mx[0], mx[1]);
+                auto qkv    = project(ctx, h, "");
+                auto out    = Rope::attention(ctx, qkv[0], qkv[1], qkv[2], pe, nullptr);
+                tokens      = finish(ctx, tokens, h, out, mx, "");
+                return {ggml_ext_slice(c, tokens, 1, nt, tokens->ne[1]), ggml_ext_slice(c, tokens, 1, 0, nt)};
+            }
+            auto my  = ggml_ext_chunk(c, ggml_add(c, txt_mod, params["adaln_txt.bias"]), 6, 0);
+            auto hx  = Pid::apply_adaln(c, norm(ctx, "norm_x1", x), mx[0], mx[1]);
+            auto hy  = Pid::apply_adaln(c, norm(ctx, "norm_y1", y), my[0], my[1]);
+            auto qx  = project(ctx, hx, "_x");
+            auto qy  = project(ctx, hy, "_y");
+            auto out = Rope::attention(ctx, ggml_concat(c, qy[0], qx[0], 2), ggml_concat(c, qy[1], qx[1], 2),
+                                       ggml_concat(c, qy[2], qx[2], 2), pe, nullptr);
+            x        = finish(ctx, x, hx, ggml_ext_slice(c, out, 1, nt, out->ne[1]), mx, "_x");
+            y        = finish(ctx, y, hy, ggml_ext_slice(c, out, 1, 0, nt), my, "_y");
+            return {x, y};
+        }
+    };
+
+    struct PixelBlock : public GGMLBlock {
+        IrisConfig config;
+
+        PixelBlock(const IrisConfig& config)
+            : config(config) {
+            auto d             = config.pixel_dim;
+            auto flat          = d * config.patch_size * config.patch_size;
+            blocks["norm1"]    = std::make_shared<RMSNorm>(d, 1e-6f);
+            blocks["norm2"]    = std::make_shared<RMSNorm>(d, 1e-6f);
+            blocks["compress"] = std::make_shared<Linear>(flat, config.pixel_attn_dim, true);
+            blocks["expand"]   = std::make_shared<Linear>(config.pixel_attn_dim, flat, true);
+            blocks["adaln"]    = std::make_shared<Linear>(config.hidden_size, 4 * flat, true);
+            blocks["attn"]     = std::make_shared<SelfAttention>(config.pixel_attn_dim, config.pixel_heads);
+            blocks["mlp.fc1"]  = std::make_shared<Linear>(d, d * 4, true);
+            blocks["mlp.fc2"]  = std::make_shared<Linear>(d * 4, d, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* cond, ggml_tensor* pe) {
+            auto c       = ctx->ggml_ctx;
+            auto d       = config.pixel_dim;
+            auto p2      = config.patch_size * config.patch_size;
+            auto patches = cond->ne[1];
+            auto batch   = cond->ne[2];
+            auto mods    = std::dynamic_pointer_cast<Linear>(blocks["adaln"])->forward(ctx, cond);
+            mods         = ggml_reshape_3d(c, mods, 4 * d, p2, patches * batch);
+            auto m       = ggml_ext_chunk(c, mods, 4, 0);
+            auto h       = std::dynamic_pointer_cast<RMSNorm>(blocks["norm1"])->forward(ctx, x);
+            h            = ggml_reshape_3d(c, h, d * p2, patches, batch);
+            h            = std::dynamic_pointer_cast<Linear>(blocks["compress"])->forward(ctx, h);
+            h            = std::dynamic_pointer_cast<SelfAttention>(blocks["attn"])->forward(ctx, h, pe);
+            h            = std::dynamic_pointer_cast<Linear>(blocks["expand"])->forward(ctx, h);
+            h            = ggml_reshape_3d(c, h, d, p2, patches * batch);
+            x            = ggml_add(c, x, Pid::apply_adaln(c, h, m[1], m[0]));
+            h            = std::dynamic_pointer_cast<RMSNorm>(blocks["norm2"])->forward(ctx, x);
+            h            = std::dynamic_pointer_cast<Linear>(blocks["mlp.fc1"])->forward(ctx, h);
+            h            = ggml_gelu_erf(c, h);
+            h            = std::dynamic_pointer_cast<Linear>(blocks["mlp.fc2"])->forward(ctx, h);
+            return ggml_add(c, x, Pid::apply_adaln(c, h, m[3], m[2]));
+        }
+    };
+
+    struct IrisModel : public GGMLBlock {
+        IrisConfig config;
+
+        IrisModel(const IrisConfig& config)
+            : config(config) {
+            blocks["s_embedder"]                 = std::make_shared<Pid::PatchTokenEmbedder>(3 * config.patch_size * config.patch_size, config.hidden_size);
+            blocks["t_embedder"]                 = std::make_shared<Pid::PixelDiTTimestepEmbedder>(config.hidden_size);
+            blocks["y_embedder"]                 = std::make_shared<TextEmbedder>(config);
+            blocks["modulation_cores.adaln_img"] = std::make_shared<Linear>(config.hidden_size, 6 * config.hidden_size, true);
+            blocks["modulation_cores.adaln_txt"] = std::make_shared<Linear>(config.hidden_size, 6 * config.hidden_size, true);
+            for (int64_t i = 0; i < config.depth; ++i) {
+                blocks["blocks." + std::to_string(i)] = std::make_shared<TrunkBlock>(config, i < config.dual_depth);
+            }
+            blocks["pixel_embedder"] = std::make_shared<Pid::PixelTokenEmbedder>(3, config.pixel_dim);
+            for (int64_t i = 0; i < config.pixel_depth; ++i) {
+                blocks["pixel_blocks." + std::to_string(i)] = std::make_shared<PixelBlock>(config);
+            }
+            blocks["final_layer"] = std::make_shared<Pid::FinalLayer>(config.pixel_dim, 3);
+        }
+
+        void init_params(ggml_context* ctx, const String2TensorStorage& = {}, const std::string = "") override {
+            params["y_pos_embedding"] = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, config.hidden_size, config.text_len, 1);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* image, ggml_tensor* t, ggml_tensor* text, ggml_tensor* mask, ggml_tensor* joint_pe, ggml_tensor* pixel_pe, ggml_tensor* pixel_pos) {
+            auto c    = ctx->ggml_ctx;
+            int p     = static_cast<int>(config.patch_size);
+            auto x    = DiT::patchify(c, image, p, p, true);
+            x         = std::dynamic_pointer_cast<Pid::PatchTokenEmbedder>(blocks["s_embedder"])->forward(ctx, x);
+            auto te   = std::dynamic_pointer_cast<Pid::PixelDiTTimestepEmbedder>(blocks["t_embedder"])->forward(ctx, t);
+            te        = ggml_reshape_3d(c, te, config.hidden_size, 1, image->ne[3]);
+            auto cond = ggml_silu(c, te);
+            auto im   = std::dynamic_pointer_cast<Linear>(blocks["modulation_cores.adaln_img"])->forward(ctx, cond);
+            auto tm   = std::dynamic_pointer_cast<Linear>(blocks["modulation_cores.adaln_txt"])->forward(ctx, cond);
+            auto y    = std::dynamic_pointer_cast<TextEmbedder>(blocks["y_embedder"])->forward(ctx, text, mask);
+            y         = ggml_add(c, y, params["y_pos_embedding"]);
+            for (int64_t i = 0; i < config.depth; ++i) {
+                auto block     = std::dynamic_pointer_cast<TrunkBlock>(blocks["blocks." + std::to_string(i)]);
+                std::tie(x, y) = block->forward(ctx, x, y, im, tm, joint_pe);
+                sd::ggml_graph_cut::mark_graph_cut(x, "iris.blocks." + std::to_string(i), "x");
+                sd::ggml_graph_cut::mark_graph_cut(y, "iris.blocks." + std::to_string(i), "y");
+            }
+            cond = ggml_silu(c, ggml_add(c, x, te));
+            x    = std::dynamic_pointer_cast<Pid::PixelTokenEmbedder>(blocks["pixel_embedder"])->forward(ctx, image, p, pixel_pos);
+            for (int64_t i = 0; i < config.pixel_depth; ++i) {
+                x = std::dynamic_pointer_cast<PixelBlock>(blocks["pixel_blocks." + std::to_string(i)])->forward(ctx, x, cond, pixel_pe);
+                sd::ggml_graph_cut::mark_graph_cut(x, "iris.pixel_blocks." + std::to_string(i), "x");
+            }
+            x = std::dynamic_pointer_cast<Pid::FinalLayer>(blocks["final_layer"])->forward(ctx, x);
+            x = ggml_reshape_3d(c, x, 3 * p * p, cond->ne[1], image->ne[3]);
+            return DiT::unpatchify(c, x, image->ne[1] / p, image->ne[0] / p, p, p, false);
+        }
+    };
+
+    inline std::vector<float> image_rope(int64_t height, int64_t width, int64_t dim) {
+        std::vector<float> pe;
+        pe.reserve(height * width * dim * 2);
+        float step = 16.f / static_cast<float>(std::max<int64_t>(std::max(height, width) - 1, 1));
+        for (int64_t y = 0; y < height; ++y) {
+            for (int64_t x = 0; x < width; ++x) {
+                for (int64_t i = 0; i < dim / 4; ++i) {
+                    float frequency = std::pow(10000.f, -4.f * static_cast<float>(i) / static_cast<float>(dim));
+                    for (int64_t pos : {x, y}) {
+                        float angle = static_cast<float>(pos) * step * frequency;
+                        pe.insert(pe.end(), {std::cos(angle), -std::sin(angle), std::sin(angle), std::cos(angle)});
+                    }
+                }
+            }
+        }
+        return pe;
+    }
+
+    struct IrisRunner : public DiffusionModelRunner {
+        IrisConfig config;
+        IrisModel model;
+        std::vector<float> joint_pe_data;
+        std::vector<float> pixel_pe_data;
+        std::vector<float> pixel_pos_data;
+        std::vector<float> text_mask_data;
+
+        IrisRunner(ggml_backend_t backend, const String2TensorStorage& weights, const std::string& prefix, std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+            : DiffusionModelRunner(backend, prefix, weight_manager),
+              config(IrisConfig::detect_from_weights(weights, prefix)),
+              model(config) {
+            model.init(params_ctx, weights, prefix);
+        }
+
+        std::string get_desc() override { return "Iris-3B"; }
+
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
+            model.get_param_tensors(tensors, prefix);
+        }
+
+        sd::Tensor<float> compute(int n_threads, const DiffusionParams& inputs) override {
+            if (!inputs.x || !inputs.timesteps || !inputs.context || !inputs.y || inputs.y->empty()) {
+                LOG_ERROR("Iris-3B requires text features and their padding mask");
+                return {};
+            }
+            auto get_graph = [&]() {
+                auto gf   = new_graph_custom(196608);
+                auto x    = make_input(*inputs.x);
+                auto t    = make_input(*inputs.timesteps);
+                auto text = make_input(*inputs.context);
+                int h     = static_cast<int>(x->ne[1]);
+                int w     = static_cast<int>(x->ne[0]);
+                int nt    = static_cast<int>(config.text_len);
+                GGML_ASSERT(h % config.patch_size == 0 && w % config.patch_size == 0);
+                GGML_ASSERT(text->ne[0] == config.text_dim * config.text_layers && text->ne[1] == nt);
+                GGML_ASSERT(inputs.y->numel() == nt && x->ne[3] == 1);
+                joint_pe_data = Pid::make_rope_1d(nt, static_cast<int>(config.head_dim), 10000.f);
+                auto img      = image_rope(h / config.patch_size, w / config.patch_size, config.head_dim);
+                joint_pe_data.insert(joint_pe_data.end(), img.begin(), img.end());
+                pixel_pe_data  = image_rope(h / config.patch_size, w / config.patch_size, config.pixel_attn_dim / config.pixel_heads);
+                pixel_pos_data = Pid::make_pixel_abs_pos(h, w, static_cast<int>(config.pixel_dim));
+                text_mask_data.resize(nt * nt);
+                for (int q = 0; q < nt; ++q) {
+                    for (int k = 0; k < nt; ++k) {
+                        text_mask_data[q * nt + k] = inputs.y->values()[k] != 0.f || q == k ? 0.f : -INFINITY;
+                    }
+                }
+                auto pe   = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, config.head_dim / 2, nt + img.size() / (2 * config.head_dim));
+                auto pp   = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, config.pixel_attn_dim / config.pixel_heads / 2, (h / config.patch_size) * (w / config.patch_size));
+                auto pos  = ggml_new_tensor_3d(compute_ctx, GGML_TYPE_F32, config.pixel_dim, h * w, 1);
+                auto mask = ggml_new_tensor_2d(compute_ctx, GGML_TYPE_F32, nt, nt);
+                set_backend_tensor_data(pe, joint_pe_data.data());
+                set_backend_tensor_data(pp, pixel_pe_data.data());
+                set_backend_tensor_data(pos, pixel_pos_data.data());
+                set_backend_tensor_data(mask, text_mask_data.data());
+                auto ctx = get_context();
+                auto out = model.forward(&ctx, x, t, text, mask, pe, pp, pos);
+                ggml_build_forward_expand(gf, out);
+                return gf;
+            };
+            return restore_trailing_singleton_dims(GGMLRunner::compute(get_graph, n_threads, false), inputs.x->dim());
+        }
+    };
+}  // namespace Iris
+
+#endif  // __SD_MODEL_DIFFUSION_IRIS_HPP__

src/model/diffusion/ltxv.hpp

@@ -548,21 +548,22 @@ namespace LTXV {
         return build_rope_matrix_from_frequencies(freqs, dim);
     }
 
-    __STATIC_INLINE__ ggml_tensor* apply_hidden_rope(ggml_context* ctx,
+    __STATIC_INLINE__ ggml_tensor* apply_hidden_rope(GGMLRunnerContext* runner_ctx,
                                                      ggml_tensor* x,
                                                      ggml_tensor* pe,
                                                      int64_t heads,
                                                      int64_t dim_head,
                                                      bool rope_interleaved) {
+        auto ctx = runner_ctx->ggml_ctx;
         GGML_ASSERT(x->ne[0] == heads * dim_head);
         auto x4 = ggml_reshape_4d(ctx, x, dim_head, heads, x->ne[1], x->ne[2]);
         if (pe != nullptr && pe->ne[3] == x->ne[1] * heads) {
             auto x_flat   = ggml_reshape_4d(ctx, x4, dim_head, 1, x->ne[1] * heads, x->ne[2]);
-            auto out_flat = Rope::apply_rope(ctx, x_flat, pe, rope_interleaved);
+            auto out_flat = Rope::apply_rope(runner_ctx, x_flat, pe, rope_interleaved);
             auto out4     = ggml_reshape_4d(ctx, out_flat, dim_head, heads, x->ne[1], x->ne[2]);
             return ggml_reshape_3d(ctx, out4, heads * dim_head, x->ne[1], x->ne[2]);
         }
-        return Rope::apply_rope(ctx, x4, pe, rope_interleaved);
+        return Rope::apply_rope(runner_ctx, x4, pe, rope_interleaved);
     }
 
     struct TimestepEmbedder : public GGMLBlock {
@@ -705,8 +706,8 @@ namespace LTXV {
                 if (k_pe == nullptr) {
                     k_pe = pe;
                 }
-                q = apply_hidden_rope(ctx->ggml_ctx, q, pe, heads, dim_head, rope_interleaved);
-                k = apply_hidden_rope(ctx->ggml_ctx, k, k_pe, heads, dim_head, rope_interleaved);
+                q = apply_hidden_rope(ctx, q, pe, heads, dim_head, rope_interleaved);
+                k = apply_hidden_rope(ctx, k, k_pe, heads, dim_head, rope_interleaved);
             }
 
             auto out = ggml_ext_attention_ext(ctx,

src/model/diffusion/minimax_h3.hpp

@@ -160,12 +160,13 @@ namespace MiniMaxH3 {
         return ggml_reshape_3d(ctx, x, x->ne[0], x->ne[1], x->ne[2] * x->ne[3]);
     }
 
-    static ggml_tensor* apply_partial_rope(ggml_context* ctx,
+    static ggml_tensor* apply_partial_rope(GGMLRunnerContext* runner_ctx,
                                            ggml_tensor* x,
                                            ggml_tensor* pe) {
+        auto ctx        = runner_ctx->ggml_ctx;
         int64_t rot_dim = pe->ne[2] * 2;
         GGML_ASSERT(rot_dim <= x->ne[0]);
-        auto rotated = Rope::apply_rope(ctx,
+        auto rotated = Rope::apply_rope(runner_ctx,
                                         ggml_ext_slice(ctx, x, 0, 0, rot_dim),
                                         pe,
                                         false);
@@ -209,8 +210,8 @@ namespace MiniMaxH3 {
             q                = q_norm->forward(ctx, q);
             k                = k_norm->forward(ctx, k);
             if (pe != nullptr) {
-                q = apply_partial_rope(ctx->ggml_ctx, q, pe);
-                k = apply_partial_rope(ctx->ggml_ctx, k, pe);
+                q = apply_partial_rope(ctx, q, pe);
+                k = apply_partial_rope(ctx, k, pe);
             } else {
                 q = attention_layout(ctx->ggml_ctx, q);
                 k = attention_layout(ctx->ggml_ctx, k);

src/model/diffusion/qwen_image_2_1.hpp

@@ -188,8 +188,8 @@ namespace Qwen {
             auto v = project("to_v");
             q      = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_q"])->forward(ctx, q);
             k      = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_k"])->forward(ctx, k);
-            q      = Rope::apply_rope(ctx->ggml_ctx, q, pe);
-            k      = Rope::apply_rope(ctx->ggml_ctx, k, pe);
+            q      = Rope::apply_rope(ctx, q, pe);
+            k      = Rope::apply_rope(ctx, k, pe);
             if (cache.mode == QwenImage21PrefixCache::Mode::STORE) {
                 // Preserve query-first attention evaluation while writing each layer's
                 // prefix before its full-sequence K/V can accumulate across layers.

src/model/diffusion/unet.hpp

@@ -587,7 +587,9 @@ public:
             label_emb      = ggml_silu_inplace(ctx->ggml_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 = label_emb->ne[1] > emb->ne[1]
+                      ? ggml_add(ctx->ggml_ctx, label_emb, emb)
+                      : ggml_add(ctx->ggml_ctx, emb, label_emb);  // [N, time_embed_dim]
         }
         // sd::ggml_graph_cut::mark_graph_cut(emb, "unet.prelude", "emb");
 

src/model/vae/minimax_h3_vae.hpp

@@ -228,11 +228,12 @@ namespace MiniMaxH3VAE {
         return ggml_reshape_3d(ctx, x, x->ne[0], x->ne[1], x->ne[2] * x->ne[3]);
     }
 
-    static ggml_tensor* apply_partial_rope(ggml_context* ctx,
+    static ggml_tensor* apply_partial_rope(GGMLRunnerContext* runner_ctx,
                                            ggml_tensor* x,
                                            ggml_tensor* pe) {
+        auto ctx        = runner_ctx->ggml_ctx;
         int64_t rot_dim = pe->ne[2] * 2;
-        auto rotated    = Rope::apply_rope(ctx,
+        auto rotated    = Rope::apply_rope(runner_ctx,
                                            ggml_ext_slice(ctx, x, 0, 0, rot_dim),
                                            pe,
                                            false);
@@ -289,8 +290,8 @@ namespace MiniMaxH3VAE {
                                                   batch_size);
             q                   = ggml_rms_norm(ctx->ggml_ctx, q, 1e-5f);
             k                   = ggml_rms_norm(ctx->ggml_ctx, k, 1e-5f);
-            q                   = apply_partial_rope(ctx->ggml_ctx, q, pe);
-            k                   = apply_partial_rope(ctx->ggml_ctx, k, pe);
+            q                   = apply_partial_rope(ctx, q, pe);
+            k                   = apply_partial_rope(ctx, k, pe);
             auto out            = ggml_ext_attention_ext(ctx,
                                                          q,
                                                          k,
@@ -633,8 +634,11 @@ namespace MiniMaxH3VAE {
             auto plan   = make_vae_temporal_tile_plan(input.shape()[2], {17, 0});
             auto result = process_vae_temporal_tiles(input, plan, [&](const sd::Tensor<float>& chunk, const VAETemporalTile& tile) {
                 SD_UNUSED(tile);
-                return VAE::encode(n_threads, chunk, tiling, circular_x, circular_y);
+                // keep the runner alive across chunks; ending it here would
+                // evict and reload the encoder weights every chunk
+                return VAE::encode(n_threads, chunk, tiling, circular_x, circular_y, false);
             });
+            runner_end();
             if (result.empty()) {
                 return {};
             }
@@ -695,13 +699,16 @@ namespace MiniMaxH3VAE {
                 {static_cast<int>(tokens_per_chunk + token_overlap), static_cast<int>(token_overlap)});
             GGML_ASSERT(plan.tiles.size() == static_cast<size_t>(num_chunks));
             auto result = process_vae_temporal_tiles(input, plan, [&](const sd::Tensor<float>& chunk, const VAETemporalTile& tile) {
+                // keep the runner alive across chunks; ending it here would
+                // evict and reload the decoder weights every chunk
                 auto decoded = VAE::decode(n_threads,
                                            chunk,
                                            tiling,
                                            true,
                                            circular_x,
                                            circular_y,
-                                           silent);
+                                           silent,
+                                           false);
                 if (decoded.empty()) {
                     return sd::Tensor<float>();
                 }
@@ -728,6 +735,7 @@ namespace MiniMaxH3VAE {
                 }
                 return first;
             });
+            runner_end();
             if (result.empty()) {
                 return {};
             }

src/model/vae/tae.hpp

@@ -84,12 +84,17 @@ class TinyEncoder : public UnaryBlock {
     int channels    = 64;
     int z_channels  = 4;
     int num_blocks  = 3;
+    bool f16;
 
 public:
-    TinyEncoder(int z_channels = 4, bool use_midblock_gn = false)
-        : z_channels(z_channels) {
-        int index                       = 0;
+    TinyEncoder(int z_channels = 4, bool use_midblock_gn = false, bool f16 = false)
+        : z_channels(z_channels), f16(f16) {
+        in_channels                     = f16 ? 16 : 3;
+        int index                       = f16 ? 1 : 0;
         blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, channels, {3, 3}, {1, 1}, {1, 1}));
+        if (f16) {
+            index++;  // nn.ReLU()
+        }
         blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
 
         blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {2, 2}, {1, 1}, {1, 1}, false));
@@ -97,12 +102,14 @@ public:
             blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
         }
 
-        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {2, 2}, {1, 1}, {1, 1}, false));
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, f16 ? channels * 2 : channels, {3, 3}, {2, 2}, {1, 1}, {1, 1}, false));
+        channels *= f16 ? 2 : 1;
         for (int i = 0; i < num_blocks; i++) {
             blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
         }
 
-        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {2, 2}, {1, 1}, {1, 1}, false));
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, f16 ? channels * 2 : channels, {3, 3}, {2, 2}, {1, 1}, {1, 1}, false));
+        channels *= f16 ? 2 : 1;
         for (int i = 0; i < num_blocks; i++) {
             blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels, use_midblock_gn));
         }
@@ -114,7 +121,11 @@ public:
         // x: [n, in_channels, h, w]
         // return: [n, z_channels, h/8, w/8]
 
-        for (int i = 0; i < num_blocks * 3 + 6; i++) {
+        for (int i = f16 ? 1 : 0; i < num_blocks * 3 + 6 + (f16 ? 2 : 0); i++) {
+            if (f16 && i == 2) {
+                x = ggml_relu_inplace(ctx->ggml_ctx, x);
+                continue;
+            }
             auto block = std::dynamic_pointer_cast<UnaryBlock>(blocks[std::to_string(i)]);
 
             x = block->forward(ctx, x);
@@ -131,9 +142,11 @@ class TinyDecoder : public UnaryBlock {
     int num_blocks   = 3;
 
 public:
-    TinyDecoder(int z_channels = 4, bool use_midblock_gn = false)
+    TinyDecoder(int z_channels = 4, bool use_midblock_gn = false, bool f16 = false)
         : z_channels(z_channels) {
-        int index = 0;
+        channels     = f16 ? 256 : 64;
+        out_channels = f16 ? 16 : 3;
+        int index    = 0;
 
         blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(z_channels, channels, {3, 3}, {1, 1}, {1, 1}));
         index++;  // nn.ReLU()
@@ -142,13 +155,15 @@ public:
             blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels, use_midblock_gn));
         }
         index++;  // nn.Upsample()
-        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {1, 1}, {1, 1}, {1, 1}, false));
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, f16 ? channels / 2 : channels, {3, 3}, {1, 1}, {1, 1}, {1, 1}, false));
+        channels /= f16 ? 2 : 1;
 
         for (int i = 0; i < num_blocks; i++) {
             blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
         }
         index++;  // nn.Upsample()
-        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {1, 1}, {1, 1}, {1, 1}, false));
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, f16 ? channels / 2 : channels, {3, 3}, {1, 1}, {1, 1}, {1, 1}, false));
+        channels /= f16 ? 2 : 1;
 
         for (int i = 0; i < num_blocks; i++) {
             blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
@@ -691,6 +706,7 @@ class TAESD : public GGMLBlock {
 protected:
     bool decode_only;
     bool taef2 = false;
+    bool f16   = false;
 
 public:
     int z_channels = 4;
@@ -708,10 +724,14 @@ public:
             z_channels      = 32;
             use_midblock_gn = true;
         }
-        blocks["decoder.layers"] = std::shared_ptr<GGMLBlock>(new TinyDecoder(z_channels, use_midblock_gn));
+        f16 = version == VERSION_QWEN_IMAGE_2_1;
+        if (f16) {
+            z_channels = 64;
+        }
+        blocks["decoder.layers"] = std::shared_ptr<GGMLBlock>(new TinyDecoder(z_channels, use_midblock_gn, f16));
 
         if (!decode_only) {
-            blocks["encoder.layers"] = std::shared_ptr<GGMLBlock>(new TinyEncoder(z_channels, use_midblock_gn));
+            blocks["encoder.layers"] = std::shared_ptr<GGMLBlock>(new TinyEncoder(z_channels, use_midblock_gn, f16));
         }
     }
 
@@ -720,10 +740,15 @@ public:
         if (taef2) {
             z = unpatchify(ctx->ggml_ctx, z, 2);
         }
-        return decoder->forward(ctx, z);
+        auto x = decoder->forward(ctx, z);
+        return f16 ? unpatchify(ctx->ggml_ctx, x, 2) : x;
     }
 
     ggml_tensor* encode(GGMLRunnerContext* ctx, ggml_tensor* x) {
+        if (f16) {
+            GGML_ASSERT(x->ne[0] % 2 == 0 && x->ne[1] % 2 == 0);
+            x = patchify(ctx->ggml_ctx, x, 2);
+        }
         auto encoder = std::dynamic_pointer_cast<TinyEncoder>(blocks["encoder.layers"]);
         auto z       = encoder->forward(ctx, x);
         if (taef2) {

src/model/vae/vae.hpp

@@ -176,11 +176,20 @@ public:
         int scale_factor = 8;
         if (version == VERSION_LTXAV) {
             scale_factor = 32;
-        } else if (version == VERSION_WAN2_2_TI2V || version == VERSION_QWEN_IMAGE_2_1 || sd_version_is_hunyuan_video(version) || sd_version_is_mage_flow(version) || sd_version_is_minimax_h3(version)) {
+        } else if (version == VERSION_WAN2_2_TI2V ||
+                   version == VERSION_QWEN_IMAGE_2_1 ||
+                   sd_version_is_hunyuan_video(version) ||
+                   sd_version_is_mage_flow(version) ||
+                   sd_version_is_minimax_h3(version)) {
             scale_factor = 16;
         } else if (sd_version_uses_flux2_vae(version)) {
             scale_factor = 16;
-        } else if (version == VERSION_CHROMA_RADIANCE || version == VERSION_HIDREAM_O1 || sd_version_is_minit2i(version) || sd_version_is_sensenova_u1(version) || sd_version_is_z_image_l2p(version)) {
+        } else if (version == VERSION_CHROMA_RADIANCE ||
+                   version == VERSION_HIDREAM_O1 ||
+                   sd_version_is_minit2i(version) ||
+                   sd_version_is_sensenova_u1(version) ||
+                   sd_version_is_z_image_l2p(version) ||
+                   version == VERSION_IRIS) {
             scale_factor = 1;
         }
         return scale_factor;
@@ -241,6 +250,15 @@ public:
                                      sd_tiling_params_t tiling_params,
                                      bool circular_x = false,
                                      bool circular_y = false) {
+        return encode(n_threads, x, tiling_params, circular_x, circular_y, true);
+    }
+
+    sd::Tensor<float> encode(int n_threads,
+                             const sd::Tensor<float>& x,
+                             sd_tiling_params_t tiling_params,
+                             bool circular_x,
+                             bool circular_y,
+                             bool auto_end) {
         int64_t t0              = ggml_time_ms();
         tiling_params           = resolve_tiling_params(tiling_params);
         sd::Tensor<float> input = x;
@@ -280,7 +298,9 @@ public:
                                                   tiling_params);
         }
 
-        runner_end();
+        if (auto_end) {
+            runner_end();
+        }
 
         if (output.empty()) {
             LOG_ERROR("vae encode compute failed");
@@ -298,6 +318,17 @@ public:
                                      bool circular_x   = false,
                                      bool circular_y   = false,
                                      bool silent       = false) {
+        return decode(n_threads, x, tiling_params, decode_video, circular_x, circular_y, silent, true);
+    }
+
+    sd::Tensor<float> decode(int n_threads,
+                             const sd::Tensor<float>& x,
+                             sd_tiling_params_t tiling_params,
+                             bool decode_video,
+                             bool circular_x,
+                             bool circular_y,
+                             bool silent,
+                             bool auto_end) {
         int64_t t0              = ggml_time_ms();
         tiling_params           = resolve_tiling_params(tiling_params);
         sd::Tensor<float> input = x;
@@ -338,7 +369,9 @@ public:
                                                   tiling_params);
         }
 
-        runner_end();
+        if (auto_end) {
+            runner_end();
+        }
 
         if (output.empty()) {
             LOG_ERROR("vae decode compute failed");

src/model_loader.cpp

@@ -556,6 +556,10 @@ SDVersion ModelLoader::get_sd_version() const {
             tensor_storage_map.find("model.diffusion_model.audio_patch_proj.weight") != tensor_storage_map.end()) {
             return VERSION_MINIMAX_H3;
         }
+        if (name == "model.diffusion_model.y_embedder.layer_pool.weight" &&
+            tensor_storage_map.count("model.diffusion_model.modulation_cores.adaln_img.weight")) {
+            return VERSION_IRIS;
+        }
         if (tensor_storage.name.find("model.diffusion_model.blocks.0.cross_attn.norm_k.weight") != std::string::npos) {
             is_wan = true;
         }

src/model_manager.cpp

@@ -1904,6 +1904,49 @@ bool ModelManager::ensure_compute_backend_capacity(
         }
     }
 
+    // The loader retains mmap buffers after storage blocks are released, so
+    // demoting mapped states would only force duplicate allocations on reload.
+    std::unordered_set<TensorState*> mapped_states;
+    for (const auto& block : params_storage_blocks_) {
+        if (!block->mmap_tensor_stores.empty()) {
+            mapped_states.insert(block->states.begin(), block->states.end());
+        }
+    }
+    std::unordered_set<TensorState*> required_set(required_states.begin(), required_states.end());
+    std::vector<TensorState*> demoted_states;
+    auto revert_demotions = [&]() {
+        for (TensorState* state : demoted_states) {
+            if (state->loaded_to_params_backend) {
+                state->residency_mode = ResidencyMode::ParamBackend;
+            }
+        }
+    };
+    auto can_demote_to_disk = [&](TensorState* state) {
+        return state != nullptr && state->tensor != nullptr && state->tensor->view_src == nullptr &&
+               state->residency_mode == ResidencyMode::ParamBackend &&
+               state->params_backend != nullptr && state->params_backend == state->compute_backend &&
+               state->compute_backend == compute_backend && !sd_backend_is_cpu(state->params_backend) &&
+               state->pin_count == 0 && !state->staged_to_compute_backend && state->loaded_to_params_backend &&
+               protected_states.find(state) == protected_states.end() &&
+               required_set.find(state) == required_set.end() &&
+               mapped_states.find(state) == mapped_states.end() &&
+               !state->sources.empty() && state->applied_lora_epoch == UINT64_MAX &&
+               state->split_buffer_type == nullptr;
+    };
+    for (TensorState* state : global_candidates) {
+        if (!can_demote_to_disk(state)) {
+            continue;
+        }
+        state->residency_mode = ResidencyMode::Disk;
+        demoted_states.push_back(state);
+        add_evictable_state(state);
+        if (release_eviction_states()) {
+            revert_demotions();
+            return true;
+        }
+    }
+    revert_demotions();
+
     const auto capacity                = check_capacity(request, required_states, true);
     const std::string available_device = capacity.available_device_bytes == SIZE_MAX
                                              ? "unknown"

src/pipeline/diffusion_engine.cpp

@@ -108,6 +108,7 @@ const char* model_version_to_str[] = {
     "PixArt",
     "Ming-Image",
     "Z-Image L2P",
+    "Iris-3B",
 };
 
 static_assert(VERSION_COUNT == sizeof(model_version_to_str) / sizeof(model_version_to_str[0]),
@@ -1382,7 +1383,8 @@ bool StableDiffusionGGML::build_denoiser() {
                    sd_version_is_llada_image(version) ||
                    sd_version_is_boogu_image(version) ||
                    sd_version_is_pid(version) ||
-                   sd_version_is_ideogram4(version)) {
+                   sd_version_is_ideogram4(version) ||
+                   version == VERSION_IRIS) {
             pred_type = FLOW_PRED;
             if (sd_version_is_wan(version)) {
                 default_flow_shift = 5.f;
@@ -1390,7 +1392,7 @@ bool StableDiffusionGGML::build_denoiser() {
                 default_flow_shift = 7.f;
             } else if (sd_version_is_minimax_h3(version)) {
                 default_flow_shift = 12.f;
-            } else if (sd_version_is_ernie_image(version)) {
+            } else if (sd_version_is_ernie_image(version) || version == VERSION_IRIS) {
                 default_flow_shift = 4.f;
             } else if (sd_version_is_pid(version)) {
                 default_flow_shift = 1.5f;
@@ -2225,6 +2227,27 @@ void StableDiffusionGGML::report_sample_progress(int step,
     }
 }
 
+static sd::Tensor<float> batch_two_condition_tensors(const sd::Tensor<float>& a, const sd::Tensor<float>& b) {
+    if (a.empty() || b.empty() || a.dim() != b.dim()) {
+        return {};
+    }
+    if (a.dim() == 1) {
+        if (a.shape() != b.shape()) {
+            return {};
+        }
+        auto batched = sd::ops::concat(a, b, 0);
+        batched.reshape_({a.shape()[0], 2});
+        return batched;
+    }
+    const int64_t batch_dim = a.dim() - 1;
+    for (int64_t d = 0; d < batch_dim; d++) {
+        if (a.shape()[d] != b.shape()[d]) {
+            return {};
+        }
+    }
+    return sd::ops::concat(a, b, static_cast<size_t>(batch_dim));
+}
+
 void StableDiffusionGGML::compute_sample_controls(const sd::Tensor<float>& control_image,
                                                   const sd::Tensor<float>& noised_input,
                                                   const sd::Tensor<float>& timesteps_tensor,
@@ -2373,7 +2396,8 @@ sd::Tensor<float> StableDiffusionGGML::sample(const std::shared_ptr<DiffusionMod
         noise *= eta;
     }
 
-    int64_t last_progress_us     = ggml_time_us();
+    int64_t last_progress_us = ggml_time_us();
+    sd_begin_preview_pass(static_cast<int>(steps));
     SamplePreviewContext preview = prepare_sample_preview_context();
 
     sd::Tensor<float> processed_init_latent       = denoiser->process_latent_in(init_latent);
@@ -2601,6 +2625,57 @@ sd::Tensor<float> StableDiffusionGGML::sample(const std::shared_ptr<DiffusionMod
             return output_opt;
         };
 
+        auto run_batched_condition = [&](const SDCondition& condition,
+                                         const sd::Tensor<float>* c_concat_override) -> sd::Tensor<float> {
+            const sd::Tensor<float>& condition_concat =
+                c_concat_override != nullptr ? *c_concat_override : condition.c_concat;
+
+            sd::Tensor<float> batched_context = batch_two_condition_tensors(condition.c_crossattn, uncond.c_crossattn);
+            sd::Tensor<float> batched_y       = batch_two_condition_tensors(condition.c_vector, uncond.c_vector);
+            sd::Tensor<float> batched_concat  = batch_two_condition_tensors(condition_concat, uncond.c_concat);
+            if (!condition.c_crossattn.empty() && batched_context.empty()) {
+                return {};
+            }
+            if ((!condition.c_vector.empty() || !uncond.c_vector.empty()) && batched_y.empty()) {
+                return {};
+            }
+            if ((!condition_concat.empty() || !uncond.c_concat.empty()) && batched_concat.empty()) {
+                return {};
+            }
+
+            std::vector<sd::Tensor<float>> uncond_controls;
+            compute_sample_controls(control_image, noised_input, timesteps_tensor, uncond, &uncond_controls);
+            if (controls.size() != uncond_controls.size()) {
+                return {};
+            }
+            std::vector<sd::Tensor<float>> batched_controls;
+            batched_controls.reserve(controls.size());
+            for (size_t i = 0; i < controls.size(); i++) {
+                sd::Tensor<float> batched_control = batch_two_condition_tensors(controls[i], uncond_controls[i]);
+                if (batched_control.empty()) {
+                    return {};
+                }
+                batched_controls.push_back(std::move(batched_control));
+            }
+
+            sd::Tensor<float> batched_x =
+                sd::ops::concat(noised_input, noised_input, static_cast<size_t>(noised_input.dim() - 1));
+
+            DiffusionParams batched_params = diffusion_params;
+            batched_params.x               = &batched_x;
+            batched_params.context         = batched_context.empty() ? nullptr : &batched_context;
+            batched_params.c_concat        = batched_concat.empty() ? nullptr : &batched_concat;
+            batched_params.y               = batched_y.empty() ? nullptr : &batched_y;
+            batched_params.ref_latents     = nullptr;
+            batched_params.extra           = UNetDiffusionExtra{1, &batched_controls, control_strength};
+
+            sd::Tensor<float> output = work_diffusion_model->compute(n_threads, batched_params);
+            if (output.empty()) {
+                LOG_ERROR("batched diffusion model compute failed");
+            }
+            return output;
+        };
+
         const SDCondition* positive_condition      = &cond;
         const sd::Tensor<float>* c_concat_override = nullptr;
         for (const auto& extension : generation_extensions) {
@@ -2640,12 +2715,40 @@ sd::Tensor<float> StableDiffusionGGML::sample(const std::shared_ptr<DiffusionMod
             }
         }
 
-        cond_out = run_condition(*positive_condition, c_concat_override);
-        if (cond_out.empty()) {
-            return {};
+        const bool batch_cfg_ok = config_->params.batched_cfg &&
+                                  sd_version_is_unet(version) &&
+                                  !uncond.empty() &&
+                                  img_uncond.empty() &&
+                                  !skip_uncond &&
+                                  !cache_runtime.ucache_enabled() &&
+                                  !(is_skiplayer_step && slg_uncond) &&
+                                  ip_adapter_tokens.empty() &&
+                                  ip_adapter_uncond_tokens.empty() &&
+                                  !config_->animatediff_loaded &&
+                                  (noised_input.dim() < 4 || noised_input.shape()[3] <= 1) &&
+                                  std::none_of(generation_extensions.begin(),
+                                               generation_extensions.end(),
+                                               [](const std::shared_ptr<GenerationExtension>& extension) {
+                                                   return extension->is_enabled();
+                                               });
+
+        if (batch_cfg_ok) {
+            sd::Tensor<float> batched_out = run_batched_condition(*positive_condition, c_concat_override);
+            if (!batched_out.empty() && batched_out.dim() >= 4 && batched_out.shape()[3] == 2) {
+                auto parts    = sd::ops::chunk(batched_out, 2, 3);
+                cond_out      = std::move(parts[0]);
+                uncond_out    = std::move(parts[1]);
+            }
         }
 
-        if (!uncond.empty()) {
+        if (cond_out.empty()) {
+            cond_out = run_condition(*positive_condition, c_concat_override);
+            if (cond_out.empty()) {
+                return {};
+            }
+        }
+
+        if (uncond_out.empty() && !uncond.empty()) {
             if (!skip_uncond) {
                 const std::vector<int>* uncond_skip_layers = nullptr;
                 if (is_skiplayer_step && slg_uncond) {
@@ -2764,7 +2867,7 @@ int StableDiffusionGGML::get_diffusion_model_down_factor() {
     if (sd_version_is_dit(version)) {
         if (sd_version_is_sensenova_u1(version)) {
             down_factor = 32;
-        } else if (sd_version_is_z_image_l2p(version)) {
+        } else if (sd_version_is_z_image_l2p(version) || version == VERSION_IRIS) {
             down_factor = 16;
         } else if (version == VERSION_QWEN_IMAGE_2_1 || version == VERSION_MING_IMAGE || sd_version_is_wan(version) || sd_version_is_lingbot_video(version) || sd_version_is_minimax_h3(version) || sd_version_is_pixart(version)) {
             down_factor = 2;
@@ -2798,7 +2901,7 @@ int StableDiffusionGGML::get_latent_channel() {
             latent_channel = 3;
         } else if (sd_version_is_z_image_l2p(version)) {
             latent_channel = 3;
-        } else if (sd_version_is_pid(version)) {
+        } else if (sd_version_is_pid(version) || version == VERSION_IRIS) {
             latent_channel = 3;
         } else if (sd_version_is_sefi_image(version)) {
             latent_channel = 144;
@@ -2899,7 +3002,7 @@ sd::Tensor<float> StableDiffusionGGML::encode_first_stage(const sd::Tensor<float
 }
 
 sd::Tensor<float> StableDiffusionGGML::decode_first_stage(const sd::Tensor<float>& x, bool decode_video) {
-    if (sd_version_is_pid(version) || sd_version_is_minit2i(version)) {
+    if (sd_version_is_pid(version) || sd_version_is_minit2i(version) || version == VERSION_IRIS) {
         return sd::ops::clamp((x + 1.f) * 0.5f, 0.0f, 1.0f);
     }
     auto latents                      = first_stage_model->diffusion_to_vae_latents(x);

src/pipeline/model_builders.cpp

@@ -18,6 +18,7 @@
 #include "model/diffusion/hidream_o1.hpp"
 #include "model/diffusion/hunyuan.hpp"
 #include "model/diffusion/ideogram4.hpp"
+#include "model/diffusion/iris.hpp"
 #include "model/diffusion/krea2.hpp"
 #include "model/diffusion/lens.hpp"
 #include "model/diffusion/lingbot_video.hpp"
@@ -456,6 +457,11 @@ namespace sd::model_builders {
                                                                   tensor_storage_map,
                                                                   "model.diffusion_model",
                                                                   weight_manager);
+        } else if (version == VERSION_IRIS) {
+            result.conditioner = std::make_shared<LLMEmbedder>(ctx.backends.runtime_backend(SDBackendModule::TE),
+                                                               tensor_storage_map, version, "", false, weight_manager, tokenizers);
+            result.diffusion   = std::make_shared<Iris::IrisRunner>(ctx.backends.runtime_backend(SDBackendModule::DIFFUSION),
+                                                                  tensor_storage_map, "model.diffusion_model", weight_manager);
         } else {  // SD1.x SD2.x SDXL
             std::map<std::string, std::string> embbeding_map;
             for (uint32_t i = 0; i < sd_ctx_params->embedding_count; i++) {
@@ -527,7 +533,7 @@ namespace sd::model_builders {
         }
 
         auto create_tae = [&](bool decode_only) -> std::shared_ptr<VAE> {
-            if (sd_version_uses_wan_vae(version) || sd_version_is_hunyuan_video(version) || sd_version_is_ltxav(version) || sd_version_is_minimax_h3(version)) {
+            if ((sd_version_uses_wan_vae(version) && version != VERSION_QWEN_IMAGE_2_1) || sd_version_is_hunyuan_video(version) || sd_version_is_ltxav(version) || sd_version_is_minimax_h3(version)) {
                 return std::make_shared<TinyVideoAutoEncoder>(ctx.backends.runtime_backend(SDBackendModule::VAE),
                                                               tensor_storage_map,
                                                               "decoder",
@@ -626,7 +632,12 @@ namespace sd::model_builders {
             }
         };
 
-        if (version == VERSION_CHROMA_RADIANCE || version == VERSION_HIDREAM_O1 || sd_version_is_minit2i(version) || sd_version_is_sensenova_u1(version) || sd_version_is_z_image_l2p(version)) {
+        if (version == VERSION_CHROMA_RADIANCE ||
+            version == VERSION_HIDREAM_O1 ||
+            sd_version_is_minit2i(version) ||
+            sd_version_is_sensenova_u1(version) ||
+            sd_version_is_z_image_l2p(version) ||
+            version == VERSION_IRIS) {
             LOG_INFO("using FakeVAE");
             result.vae = std::make_shared<FakeVAE>(version,
                                                    ctx.backends.runtime_backend(SDBackendModule::VAE),

src/runtime/tiling.cpp

@@ -8,60 +8,51 @@
 #include "core/util.h"
 #include "ggml.h"
 
-static int sd_tiling_calc_num_tiles(int dimension, int tile_size, float target_overlap_factor) {
-    if (dimension <= tile_size) {
-        return 1;
-    } else if (dimension < 2 * tile_size) {
-        return 2;
-    } else if (dimension == 2 * tile_size) {
-        return 3;
-    } else {
-        float target_num_tiles = 1.0f + (dimension - tile_size) / ((1.0f - target_overlap_factor) * tile_size);
-        int num_tiles_lower    = static_cast<int>(std::floor(target_num_tiles));
-        int num_tiles_upper    = static_cast<int>(std::ceil(target_num_tiles));
-        int num_tiles_min      = 1 + (dimension - 2) / (tile_size - 1);  // positive adjacent overlap
-        int num_tiles_max      = 2 * dimension / tile_size - 1;          // no triple overlap under Bresenham placement
-        num_tiles_lower        = std::clamp(num_tiles_lower, num_tiles_min, num_tiles_max);
-        num_tiles_upper        = std::clamp(num_tiles_upper, num_tiles_min, num_tiles_max);
-        auto overlap_error     = [target_num_tiles](int num_tiles) -> float {
-            return std::abs(1.0f / (num_tiles - 1.0f) - 1.0f / (target_num_tiles - 1.0f));
-        };  // (dimension - tile_size) / tile_size factors out
-        return (overlap_error(num_tiles_upper) < overlap_error(num_tiles_lower)) ? num_tiles_upper : num_tiles_lower;  // use lower if tie
+struct TileSpan {
+    int offset;
+    int overlap_before;
+    int overlap_after;
+};
+
+static std::vector<TileSpan> sd_tiling_plan_axis(int dimension,
+                                                 int tile_size,
+                                                 float target_overlap,
+                                                 bool circular) {
+    if (tile_size >= dimension) {
+        return {{0, 0, 0}};
     }
-}
 
-static float sd_tiling_calc_average_stride_factor(int dimension, int tile_size, int num_tiles) {
-    return static_cast<float>(dimension - tile_size) / static_cast<float>(tile_size * (num_tiles - 1));
-}
-
-static void sd_tiling_calc_tiles(int& num_tiles_dim,
-                                 float& tile_overlap_factor_dim,
-                                 int small_dim,
-                                 int tile_size,
-                                 const float tile_overlap_factor,
-                                 bool circular) {
-    int tile_overlap     = static_cast<int>(tile_size * tile_overlap_factor);
-    int non_tile_overlap = tile_size - tile_overlap;
-
-    if (circular) {
-        // circular means the last and first tile are overlapping (wraping around)
-        num_tiles_dim = small_dim / non_tile_overlap;
-
-        if (num_tiles_dim < 1) {
-            num_tiles_dim = 1;
-        }
-
-        tile_overlap_factor_dim = (tile_size - small_dim / num_tiles_dim) / (float)tile_size;
-
-        // if single tile and tile_overlap_factor is not 0, add one to ensure we have at least two overlapping tiles
-        if (num_tiles_dim == 1 && tile_overlap_factor_dim > 0) {
-            num_tiles_dim++;
-            tile_overlap_factor_dim = 0.5;
-        }
-    } else {
-        num_tiles_dim           = sd_tiling_calc_num_tiles(small_dim, tile_size, tile_overlap_factor);
-        tile_overlap_factor_dim = (num_tiles_dim == 1) ? 0 : (1.0f - sd_tiling_calc_average_stride_factor(small_dim, tile_size, num_tiles_dim));
+    const int extent = circular ? dimension : dimension - tile_size;
+    int intervals    = extent;
+    if (tile_size > 1) {
+        const float target = extent / ((1.0f - target_overlap) * tile_size);
+        // Adjacent tiles must overlap, but three tiles must not cover the same point.
+        const int min_intervals = 1 + (extent - 1) / (tile_size - 1);
+        const int max_intervals = std::max(1, static_cast<int>(2LL * extent / tile_size));
+        const int lower         = std::clamp(static_cast<int>(std::floor(target)), min_intervals, max_intervals);
+        const int upper         = std::clamp(static_cast<int>(std::ceil(target)), min_intervals, max_intervals);
+        auto error              = [&](int count) {
+            return std::abs(1.0f / count - 1.0f / target);
+        };
+        intervals = error(upper) < error(lower) ? upper : lower;
     }
+
+    const int num_tiles = circular ? intervals : intervals + 1;
+    auto position       = [&](int index) -> int {
+        return static_cast<int>(int64_t(index) * extent / intervals);
+    };
+    std::vector<TileSpan> tiles;
+    tiles.reserve(num_tiles);
+    for (int i = 0; i < num_tiles; ++i) {
+        const int offset = position(i);
+        // Keep wrapped neighbors in the same unwrapped coordinate space.
+        const int previous = i > 0 ? position(i - 1) : position(num_tiles - 1) - dimension;
+        const int next     = i + 1 < num_tiles ? position(i + 1) : dimension;
+        const int before   = i > 0 || circular ? previous + tile_size - offset : 0;
+        const int after    = i + 1 < num_tiles || circular ? offset + tile_size - next : 0;
+        tiles.push_back({offset, before, after});
+    }
+    return tiles;
 }
 
 static int64_t sd_tensor_plane_size(const sd::Tensor<float>& tensor) {
@@ -96,62 +87,10 @@ static void sd_tensor_merge_2d(const sd::Tensor<float>& input,
                                sd::Tensor<float>* output,
                                int x,
                                int y,
-                               int overlap_x,
-                               int overlap_y,
-                               bool circular_x,
-                               bool circular_y,
-                               int x_skip,
-                               int y_skip) {
-    GGML_ASSERT(output != nullptr);
-    int64_t width        = input.shape()[0];
-    int64_t height       = input.shape()[1];
-    int64_t img_width    = output->shape()[0];
-    int64_t img_height   = output->shape()[1];
-    int64_t input_plane  = sd_tensor_plane_size(input);
-    int64_t output_plane = sd_tensor_plane_size(*output);
-    int64_t plane_count  = input.numel() / input_plane;
-    GGML_ASSERT(output->numel() / output_plane == plane_count);
-
-    // unclamped -> expects x in the range [0-1]
-    auto smootherstep_f32 = [](const float x) -> float {
-        GGML_ASSERT(x >= 0.f && x <= 1.f);
-        return x * x * x * (x * (6.0f * x - 15.0f) + 10.0f);
-    };
-
-    for (int iy = y_skip; iy < height; iy++) {
-        for (int ix = x_skip; ix < width; ix++) {
-            int64_t src_xy = ix + width * iy;
-            int64_t ox     = (x + ix) % img_width;
-            int64_t oy     = (y + iy) % img_height;
-            int64_t dst_xy = ox + img_width * oy;
-            for (int64_t plane = 0; plane < plane_count; ++plane) {
-                float new_value = input[plane * input_plane + src_xy];
-                if (overlap_x > 0 || overlap_y > 0) {
-                    float old_value   = (*output)[plane * output_plane + dst_xy];
-                    const float x_f_0 = (circular_x || (overlap_x > 0 && x > 0)) ? (ix - x_skip) / float(overlap_x) : 1.f;
-                    const float x_f_1 = (circular_x || (overlap_x > 0 && x < (img_width - width))) ? (width - ix) / float(overlap_x) : 1.f;
-                    const float y_f_0 = (circular_y || (overlap_y > 0 && y > 0)) ? (iy - y_skip) / float(overlap_y) : 1.f;
-                    const float y_f_1 = (circular_y || (overlap_y > 0 && y < (img_height - height))) ? (height - iy) / float(overlap_y) : 1.f;
-                    const float x_f   = std::min(std::min(x_f_0, x_f_1), 1.f);
-                    const float y_f   = std::min(std::min(y_f_0, y_f_1), 1.f);
-                    (*output)[plane * output_plane + dst_xy] =
-                        old_value + new_value * smootherstep_f32(y_f) * smootherstep_f32(x_f);
-                } else {
-                    (*output)[plane * output_plane + dst_xy] = new_value;
-                }
-            }
-        }
-    }
-}
-
-static void sd_tensor_merge_2d_non_circular(const sd::Tensor<float>& input,
-                                            sd::Tensor<float>* output,
-                                            int x,
-                                            int y,
-                                            int overlap_left,
-                                            int overlap_right,
-                                            int overlap_top,
-                                            int overlap_bottom) {
+                               int overlap_left,
+                               int overlap_right,
+                               int overlap_top,
+                               int overlap_bottom) {
     GGML_ASSERT(output != nullptr);
 
     int64_t in_width    = input.shape()[0];
@@ -164,8 +103,8 @@ static void sd_tensor_merge_2d_non_circular(const sd::Tensor<float>& input,
 
     GGML_ASSERT(output->numel() == plane_count * out_size);
     GGML_ASSERT(x >= 0 && y >= 0);
-    GGML_ASSERT(x + in_width <= out_width);
-    GGML_ASSERT(y + in_height <= out_height);
+    GGML_ASSERT(x < out_width && in_width <= out_width);
+    GGML_ASSERT(y < out_height && in_height <= out_height);
     GGML_ASSERT(overlap_left >= 0 && overlap_right >= 0);
     GGML_ASSERT(overlap_top >= 0 && overlap_bottom >= 0);
 
@@ -191,7 +130,7 @@ static void sd_tensor_merge_2d_non_circular(const sd::Tensor<float>& input,
                     x_f = static_cast<float>(in_width - ix) / overlap_right;
                 }
                 float x_weight = smootherstep_f32(std::clamp(x_f, 0.0f, 1.0f));
-                (*output)[plane * out_size + out_width * (y + iy) + (x + ix)] += x_weight * y_weight * input[plane * in_size + in_width * iy + ix];
+                (*output)[plane * out_size + out_width * ((y + iy) % out_height) + ((x + ix) % out_width)] += x_weight * y_weight * input[plane * in_size + in_width * iy + ix];
             }
         }
     }
@@ -223,13 +162,8 @@ sd::Tensor<float> process_tiles_2d(const sd::Tensor<float>& input,
     int scale_in     = decode ? 1 : scale;
     int scale_out    = decode ? scale : 1;
 
-    int num_tiles_x;
-    float tile_overlap_factor_x;
-    sd_tiling_calc_tiles(num_tiles_x, tile_overlap_factor_x, small_width, p_tile_size_w, tile_overlap_factor, circular_x);
-
-    int num_tiles_y;
-    float tile_overlap_factor_y;
-    sd_tiling_calc_tiles(num_tiles_y, tile_overlap_factor_y, small_height, p_tile_size_h, tile_overlap_factor, circular_y);
+    const auto tiles_x = sd_tiling_plan_axis(small_width, p_tile_size_w, tile_overlap_factor, circular_x);
+    const auto tiles_y = sd_tiling_plan_axis(small_height, p_tile_size_h, tile_overlap_factor, circular_y);
 
     int tile_width         = std::min(p_tile_size_w, small_width);
     int tile_height        = std::min(p_tile_size_h, small_height);
@@ -238,144 +172,37 @@ sd::Tensor<float> process_tiles_2d(const sd::Tensor<float>& input,
     int output_tile_width  = tile_width * scale_out;
     int output_tile_height = tile_height * scale_out;
 
-    int num_tiles   = num_tiles_x * num_tiles_y;
-    int tile_count  = 1;
-    float last_time = 0.0f;
+    const int num_tiles = static_cast<int>(tiles_x.size() * tiles_y.size());
+    int tile_count      = 0;
     if (!silent) {
-        LOG_VERBOSE("num tiles : %d, %d ", num_tiles_x, num_tiles_y);
-        LOG_VERBOSE("optimal overlap : %f, %f (targeting %f)", tile_overlap_factor_x, tile_overlap_factor_y, tile_overlap_factor);
+        LOG_VERBOSE("num tiles : %d, %d ", static_cast<int>(tiles_x.size()), static_cast<int>(tiles_y.size()));
         LOG_VERBOSE("processing %i tiles", num_tiles);
         pretty_progress(0, num_tiles, 0.0f);
     }
-    if (circular_x || circular_y) {
-        int tile_overlap_x     = static_cast<int32_t>(p_tile_size_w * tile_overlap_factor_x);
-        int non_tile_overlap_x = p_tile_size_w - tile_overlap_x;
-        int tile_overlap_y     = static_cast<int32_t>(p_tile_size_h * tile_overlap_factor_y);
-        int non_tile_overlap_y = p_tile_size_h - tile_overlap_y;
-
-        bool last_y = false;
-        bool last_x = false;
-
-        for (int y = 0; y < small_height && !last_y; y += non_tile_overlap_y) {
-            int dy = 0;
-            if (!circular_y && y + tile_height >= small_height) {
-                int original_y = y;
-                y              = small_height - tile_height;
-                dy             = original_y - y;
-                if (decode) {
-                    dy *= scale;
-                }
-                last_y = true;
+    for (const auto& y : tiles_y) {
+        for (const auto& x : tiles_x) {
+            int64_t t1       = ggml_time_ms();
+            auto input_tile  = sd_tensor_split_2d(input, input_tile_width, input_tile_height, x.offset * scale_in, y.offset * scale_in);
+            auto output_tile = on_processing(input_tile);
+            if (output_tile.empty()) {
+                return {};
             }
-            for (int x = 0; x < small_width && !last_x; x += non_tile_overlap_x) {
-                int dx = 0;
-                if (!circular_x && x + tile_width >= small_width) {
-                    int original_x = x;
-                    x              = small_width - tile_width;
-                    dx             = original_x - x;
-                    if (decode) {
-                        dx *= scale;
-                    }
-                    last_x = true;
-                }
-
-                int x_in  = decode ? x : scale * x;
-                int y_in  = decode ? y : scale * y;
-                int x_out = decode ? x * scale : x;
-                int y_out = decode ? y * scale : y;
-
-                int overlap_x_out = decode ? tile_overlap_x * scale : tile_overlap_x;
-                int overlap_y_out = decode ? tile_overlap_y * scale : tile_overlap_y;
-
-                int64_t t1       = ggml_time_ms();
-                auto input_tile  = sd_tensor_split_2d(input, input_tile_width, input_tile_height, x_in, y_in);
-                auto output_tile = on_processing(input_tile);
-                if (output_tile.empty()) {
-                    return {};
-                }
-                GGML_ASSERT(output_tile.shape()[0] == output_tile_width && output_tile.shape()[1] == output_tile_height);
-                if (output.empty()) {
-                    std::vector<int64_t> output_shape = output_tile.shape();
-                    output_shape[0]                   = output_width;
-                    output_shape[1]                   = output_height;
-                    output                            = sd::Tensor<float>::zeros(std::move(output_shape));
-                }
-                sd_tensor_merge_2d(output_tile, &output, x_out, y_out, overlap_x_out, overlap_y_out, circular_x, circular_y, dx, dy);
-
-                if (!silent) {
-                    int64_t t2 = ggml_time_ms();
-                    last_time  = (t2 - t1) / 1000.0f;
-                    pretty_progress(tile_count, num_tiles, last_time);
-                }
-                tile_count++;
+            GGML_ASSERT(output_tile.shape()[0] == output_tile_width && output_tile.shape()[1] == output_tile_height);
+            if (output.empty()) {
+                std::vector<int64_t> output_shape = output_tile.shape();
+                output_shape[0]                   = output_width;
+                output_shape[1]                   = output_height;
+                output                            = sd::Tensor<float>::zeros(std::move(output_shape));
             }
-            last_x = false;
-        }
-    } else {
-        for (int j = 0; j < num_tiles_y; ++j) {
-            int y              = 0;
-            int overlap_top    = 0;
-            int overlap_bottom = 0;
-            if (num_tiles_y > 1) {
-                y = j * (small_height - tile_height) / (num_tiles_y - 1);
-                if (j > 0) {
-                    int y_prev  = (j - 1) * (small_height - tile_height) / (num_tiles_y - 1);
-                    overlap_top = y_prev + tile_height - y;
-                }
-                if (j < num_tiles_y - 1) {
-                    int y_next     = (j + 1) * (small_height - tile_height) / (num_tiles_y - 1);
-                    overlap_bottom = y + tile_height - y_next;
-                }
-            }
-            for (int i = 0; i < num_tiles_x; ++i) {
-                int x             = 0;
-                int overlap_left  = 0;
-                int overlap_right = 0;
-                if (num_tiles_x > 1) {
-                    x = i * (small_width - tile_width) / (num_tiles_x - 1);
-                    if (i > 0) {
-                        int x_prev   = (i - 1) * (small_width - tile_width) / (num_tiles_x - 1);
-                        overlap_left = x_prev + tile_width - x;
-                    }
-                    if (i < num_tiles_x - 1) {
-                        int x_next    = (i + 1) * (small_width - tile_width) / (num_tiles_x - 1);
-                        overlap_right = x + tile_width - x_next;
-                    }
-                }
-
-                int64_t t1       = ggml_time_ms();
-                auto input_tile  = sd_tensor_split_2d(input, input_tile_width, input_tile_height, x * scale_in, y * scale_in);
-                auto output_tile = on_processing(input_tile);
-                if (output_tile.empty()) {
-                    return {};
-                }
-                GGML_ASSERT(output_tile.shape()[0] == output_tile_width && output_tile.shape()[1] == output_tile_height);
-                if (output.empty()) {
-                    std::vector<int64_t> output_shape = output_tile.shape();
-                    output_shape[0]                   = output_width;
-                    output_shape[1]                   = output_height;
-                    output                            = sd::Tensor<float>::zeros(std::move(output_shape));
-                }
-                sd_tensor_merge_2d_non_circular(
-                    output_tile, &output,
-                    x * scale_out, y * scale_out,
-                    overlap_left * scale_out,
-                    overlap_right * scale_out,
-                    overlap_top * scale_out,
-                    overlap_bottom * scale_out);
-                if (!silent) {
-                    last_time = (ggml_time_ms() - t1) / 1000.0f;
-                    pretty_progress(tile_count, num_tiles, last_time);
-                }
-                tile_count++;
+            sd_tensor_merge_2d(output_tile, &output,
+                               x.offset * scale_out, y.offset * scale_out,
+                               x.overlap_before * scale_out, x.overlap_after * scale_out,
+                               y.overlap_before * scale_out, y.overlap_after * scale_out);
+            ++tile_count;
+            if (!silent) {
+                pretty_progress(tile_count, num_tiles, (ggml_time_ms() - t1) / 1000.0f);
             }
         }
     }
-    if (!silent && tile_count < num_tiles) {
-        pretty_progress(num_tiles, num_tiles, last_time);
-    }
-    if (output.empty()) {
-        return {};
-    }
     return output;
 }

src/stable-diffusion.cpp

@@ -335,6 +335,7 @@ void sd_ctx_params_init(sd_ctx_params_t* sd_ctx_params) {
     sd_ctx_params->max_vram                  = nullptr;
     sd_ctx_params->disable_prefetch          = false;
     sd_ctx_params->disable_segmented_compute = false;
+    sd_ctx_params->batched_cfg               = true;
     sd_ctx_params->eager_load                = false;
     sd_ctx_params->enable_mmap               = false;
     sd_ctx_params->diffusion_flash_attn      = false;

src/tokenizers/bpe_tokenizer.cpp

@@ -242,7 +242,7 @@ bool BPETokenizer::encode(const std::string& text, std::vector<int>& result, on_
         ss << "\"" << token << "\", ";
     }
     ss << "]";
-    LOG_VERBOSE("split prompt \"%s\" to %zu tokens %s", text.c_str(), bpe_tokens.size(), ss.str().c_str());
+    LOG_VERBOSE("split prompt \"%s\" to %zu tokens %s", escape_newlines(text).c_str(), bpe_tokens.size(), escape_newlines(ss.str()).c_str());
     result = std::move(bpe_tokens);
     return true;
 }

src/tokenizers/hf_tokenizer.cpp

@@ -735,7 +735,7 @@ bool HFTokenizer::encode(const std::string& text, std::vector<int>& tokens, on_n
         }
     }
     ss << "]";
-    LOG_VERBOSE("split prompt \"%s\" to %zu tokens %s", text.c_str(), result.size(), ss.str().c_str());
+    LOG_VERBOSE("split prompt \"%s\" to %zu tokens %s", escape_newlines(text).c_str(), result.size(), escape_newlines(ss.str()).c_str());
     tokens = std::move(result);
     return true;
 }

src/tokenizers/t5_unigram_tokenizer.cpp

@@ -343,7 +343,7 @@ bool T5UniGramTokenizer::encode(const std::string& input, std::vector<int>& resu
         ss << "\"" << token_str << "\", ";
     }
     ss << "]";
-    LOG_VERBOSE("split prompt \"%s\" to tokens %s", input.c_str(), ss.str().c_str());
+    LOG_VERBOSE("split prompt \"%s\" to tokens %s", escape_newlines(input).c_str(), escape_newlines(ss.str()).c_str());
 
     result = std::move(tokens);
     return true;