Update docs

This reverts commit 594ccf5235.

README.md

@@ -28,6 +28,8 @@ A Gradio web UI for Large Language Models.
 - 100% offline and private, with zero telemetry, external resources, or remote update requests.
 - **File attachments**: Upload text files, PDF documents, and .docx documents to talk about their contents.
 - **Vision (multimodal models)**: Attach images to messages for visual understanding ([tutorial](https://github.com/oobabooga/text-generation-webui/wiki/Multimodal-Tutorial)).
+Image generation: A dedicated tab for diffusers models like Z-Image-Turbo and Qwen-Image. Features 4-bit/8-bit quantization and a persistent gallery with metadata (tutorial).
+- **Image generation**: A dedicated tab for `diffusers` models like **Z-Image-Turbo** and **Qwen-Image**. Features 4-bit/8-bit quantization and a persistent gallery with metadata ([tutorial](https://github.com/oobabooga/text-generation-webui/wiki/Image-Generation-Tutorial)).
 - **Web search**: Optionally search the internet with LLM-generated queries to add context to the conversation.
 - Aesthetic UI with dark and light themes.
 - Syntax highlighting for code blocks and LaTeX rendering for mathematical expressions.

css/main.css

@@ -244,37 +244,46 @@ button {
     font-size: 100% !important;
 }
 
-.pretty_scrollbar::-webkit-scrollbar {
+.pretty_scrollbar::-webkit-scrollbar,
+#image-history-gallery > :nth-child(2)::-webkit-scrollbar {
     width: 8px;
     height: 8px;
 }
 
-.pretty_scrollbar::-webkit-scrollbar-track {
+.pretty_scrollbar::-webkit-scrollbar-track,
+#image-history-gallery > :nth-child(2)::-webkit-scrollbar-track {
     background: transparent;
 }
 
 .pretty_scrollbar::-webkit-scrollbar-thumb,
-.pretty_scrollbar::-webkit-scrollbar-thumb:hover {
+.pretty_scrollbar::-webkit-scrollbar-thumb:hover,
+#image-history-gallery > :nth-child(2)::-webkit-scrollbar-thumb,
+#image-history-gallery > :nth-child(2)::-webkit-scrollbar-thumb:hover {
     background: var(--neutral-300);
     border-radius: 30px;
 }
 
 .dark .pretty_scrollbar::-webkit-scrollbar-thumb,
-.dark .pretty_scrollbar::-webkit-scrollbar-thumb:hover {
+.dark .pretty_scrollbar::-webkit-scrollbar-thumb:hover,
+.dark #image-history-gallery > :nth-child(2)::-webkit-scrollbar-thumb,
+.dark #image-history-gallery > :nth-child(2)::-webkit-scrollbar-thumb:hover {
     background: rgb(255 255 255 / 6.25%);
     border-radius: 10px;
 }
 
-.pretty_scrollbar::-webkit-resizer {
+.pretty_scrollbar::-webkit-resizer,
+#image-history-gallery > :nth-child(2)::-webkit-resizer {
     background: #c5c5d2;
 }
 
-.dark .pretty_scrollbar::-webkit-resizer {
+.dark .pretty_scrollbar::-webkit-resizer,
+.dark #image-history-gallery > :nth-child(2)::-webkit-resizer {
     background: #ccc;
     border-radius: 10px;
 }
 
-.pretty_scrollbar::-webkit-scrollbar-corner {
+.pretty_scrollbar::-webkit-scrollbar-corner,
+#image-history-gallery > :nth-child(2)::-webkit-scrollbar-corner {
     background: transparent;
 }
 
@@ -1737,3 +1746,9 @@ button#swap-height-width {
     padding: 8px;
     background: var(--background-fill-secondary);
 }
+
+/* Fix a gr.Markdown UI glitch when clicking Next in the
+ * Image AI > Gallery tab */
+.min.svelte-1yrv54 {
+    min-height: 0;
+}

docs/Image Generation Tutorial.md

@@ -0,0 +1,20 @@
+# Image Generation Tutorial
+
+This feature allows you to generate images using high-speed models like Z-Image-Turbo directly within the web UI.
+
+## How to use
+
+1. Click on the **Image AI** tab at the top of the interface.
+2. Select the **Model** sub-tab.
+3. Copy and paste the following link into the **Download model** box:
+
+```
+https://huggingface.co/Tongyi-MAI/Z-Image-Turbo
+```
+
+4. Click the **Download** button and wait for the confirmation message.
+5. In the **Model** dropdown menu, select the model you just downloaded (if you don't see it, click the 🔄 refresh button).
+6. Click **Load**.
+7. Go to the **Generate** sub-tab, type a prompt, and click **GENERATE**.
+
+> **Note for Z-Image-Turbo:** For the best results with this specific model, keep the **CFG Scale** slider at **0**.

modules/image_models.py

@@ -83,9 +83,9 @@ def get_pipeline_type(pipe):
         str: 'zimage', 'qwenimage', or 'unknown'
     """
     class_name = pipe.__class__.__name__
-    if 'ZImage' in class_name:
+    if class_name == 'ZImagePipeline':
         return 'zimage'
-    elif 'QwenImage' in class_name:
+    elif class_name == 'QwenImagePipeline':
         return 'qwenimage'
     else:
         return 'unknown'

modules/shared.py

@@ -23,6 +23,7 @@ lora_names = []
 # Image model variables
 image_model = None
 image_model_name = 'None'
+image_pipeline_type = None
 
 # Generation variables
 stop_everything = False
@@ -364,7 +365,6 @@ def apply_image_model_cli_overrides():
         settings['image_quant'] = args.image_quant
 
 
-
 def fix_loader_name(name):
     if not name:
         return name

modules/ui_image_generation.py

@@ -11,8 +11,14 @@ from PIL import Image
 from PIL.PngImagePlugin import PngInfo
 
 from modules import shared, ui, utils
-from modules.image_models import load_image_model, unload_image_model
+from modules.image_models import (
+    get_pipeline_type,
+    load_image_model,
+    unload_image_model
+)
 from modules.logging_colors import logger
+from modules.text_generation import stop_everything_event
+from modules.torch_utils import get_device
 from modules.utils import gradio
 
 ASPECT_RATIOS = {
@@ -241,6 +247,12 @@ def get_paginated_images(page=0, force_refresh=False):
     return page_images, page, total_pages, total_images
 
 
+def get_initial_page_info():
+    """Get page info string for initial load."""
+    _, page, total_pages, total_images = get_paginated_images(0)
+    return f"Page {page + 1} of {total_pages} ({total_images} total images)"
+
+
 def refresh_gallery(current_page=0):
     """Refresh gallery with current page."""
     images, page, total_pages, total_images = get_paginated_images(current_page, force_refresh=True)
@@ -324,6 +336,15 @@ def send_to_generate(selected_image_path):
     return updates + [status]
 
 
+def read_dropped_image_metadata(image_path):
+    """Read metadata from a dropped/uploaded image."""
+    if not image_path:
+        return "Drop an image to view its generation settings."
+
+    metadata = read_image_metadata(image_path)
+    return format_metadata_for_display(metadata)
+
+
 def create_ui():
     if shared.settings['image_model_menu'] != 'None':
         shared.image_model_name = shared.settings['image_model_menu']
@@ -349,7 +370,7 @@ def create_ui():
                         )
 
                         shared.gradio['image_generate_btn'] = gr.Button("Generate", variant="primary", size="lg")
-                        shared.gradio['image_generating_btn'] = gr.Button("Generating...", size="lg", visible=False, interactive=False)
+                        shared.gradio['image_stop_btn'] = gr.Button("Stop", size="lg", visible=False)
                         gr.HTML("<hr style='border-top: 1px solid #444; margin: 20px 0;'>")
 
                         gr.Markdown("### Dimensions")
@@ -374,7 +395,7 @@ def create_ui():
                                 shared.gradio['image_steps'] = gr.Slider(1, 100, value=shared.settings['image_steps'], step=1, label="Steps")
                                 shared.gradio['image_cfg_scale'] = gr.Slider(
                                     0.0, 10.0,
-                                    value=0.0,
+                                    value=shared.settings['image_cfg_scale'],
                                     step=0.1,
                                     label="CFG Scale",
                                     info="Z-Image Turbo: 0.0 | Qwen: 4.0"
@@ -396,7 +417,7 @@ def create_ui():
                         with gr.Row():
                             shared.gradio['image_refresh_history'] = gr.Button("🔄 Refresh", elem_classes="refresh-button")
                             shared.gradio['image_prev_page'] = gr.Button("◀ Prev", elem_classes="refresh-button")
-                            shared.gradio['image_page_info'] = gr.Markdown("Loading...", elem_id="image-page-info")
+                            shared.gradio['image_page_info'] = gr.Markdown(value=get_initial_page_info, elem_id="image-page-info")
                             shared.gradio['image_next_page'] = gr.Button("Next ▶", elem_classes="refresh-button")
                             shared.gradio['image_page_input'] = gr.Number(value=1, label="Page", precision=0, minimum=1, scale=0, min_width=80)
                             shared.gradio['image_go_to_page'] = gr.Button("Go", elem_classes="refresh-button", scale=0, min_width=50)
@@ -423,6 +444,13 @@ def create_ui():
                         shared.gradio['image_send_to_generate'] = gr.Button("Send to Generate", variant="primary")
                         shared.gradio['image_gallery_status'] = gr.Markdown("")
 
+                        gr.Markdown("### Import Image")
+                        shared.gradio['image_drop_upload'] = gr.Image(
+                            label="Drop image here to view settings",
+                            type="filepath",
+                            height=150
+                        )
+
             # TAB 3: MODEL
             with gr.TabItem("Model"):
                 with gr.Row():
@@ -517,22 +545,26 @@ def create_event_handlers():
     # Generation
     shared.gradio['image_generate_btn'].click(
         ui.gather_interface_values, gradio(shared.input_elements), gradio('interface_state')).then(
-        lambda: [gr.update(visible=True), gr.update(visible=False)], None, gradio('image_generating_btn', 'image_generate_btn')).then(
+        lambda: [gr.update(visible=True), gr.update(visible=False)], None, gradio('image_stop_btn', 'image_generate_btn')).then(
         generate, gradio('interface_state'), gradio('image_output_gallery'), show_progress=False).then(
-        lambda: [gr.update(visible=False), gr.update(visible=True)], None, gradio('image_generating_btn', 'image_generate_btn'))
+        lambda: [gr.update(visible=False), gr.update(visible=True)], None, gradio('image_stop_btn', 'image_generate_btn'))
 
     shared.gradio['image_prompt'].submit(
         ui.gather_interface_values, gradio(shared.input_elements), gradio('interface_state')).then(
-        lambda: [gr.update(visible=True), gr.update(visible=False)], None, gradio('image_generating_btn', 'image_generate_btn')).then(
+        lambda: [gr.update(visible=True), gr.update(visible=False)], None, gradio('image_stop_btn', 'image_generate_btn')).then(
         generate, gradio('interface_state'), gradio('image_output_gallery'), show_progress=False).then(
-        lambda: [gr.update(visible=False), gr.update(visible=True)], None, gradio('image_generating_btn', 'image_generate_btn'))
+        lambda: [gr.update(visible=False), gr.update(visible=True)], None, gradio('image_stop_btn', 'image_generate_btn'))
 
     shared.gradio['image_neg_prompt'].submit(
         ui.gather_interface_values, gradio(shared.input_elements), gradio('interface_state')).then(
-        lambda: [gr.update(visible=True), gr.update(visible=False)], None, gradio('image_generating_btn', 'image_generate_btn')).then(
+        lambda: [gr.update(visible=True), gr.update(visible=False)], None, gradio('image_stop_btn', 'image_generate_btn')).then(
         generate, gradio('interface_state'), gradio('image_output_gallery'), show_progress=False).then(
-        lambda: [gr.update(visible=False), gr.update(visible=True)], None, gradio('image_generating_btn', 'image_generate_btn'))
+        lambda: [gr.update(visible=False), gr.update(visible=True)], None, gradio('image_stop_btn', 'image_generate_btn'))
 
+    # Stop button
+    shared.gradio['image_stop_btn'].click(
+        stop_everything_event, None, None, show_progress=False
+    )
 
     # Model management
     shared.gradio['image_refresh_models'].click(
@@ -620,6 +652,13 @@ def create_event_handlers():
         show_progress=False
     )
 
+    shared.gradio['image_drop_upload'].change(
+        read_dropped_image_metadata,
+        gradio('image_drop_upload'),
+        gradio('image_settings_display'),
+        show_progress=False
+    )
+
 
 def generate(state):
     """
@@ -627,83 +666,113 @@ def generate(state):
     Automatically adjusts parameters based on pipeline type.
     """
     import torch
-    import numpy as np
 
-    model_name = state['image_model_menu']
+    from modules.torch_utils import clear_torch_cache
 
-    if not model_name or model_name == 'None':
-        logger.error("No image model selected. Go to the Model tab and select a model.")
-        return []
+    clear_torch_cache()
 
-    if shared.image_model is None:
-        result = load_image_model(
-            model_name,
-            dtype=state['image_dtype'],
-            attn_backend=state['image_attn_backend'],
-            cpu_offload=state['image_cpu_offload'],
-            compile_model=state['image_compile'],
-            quant_method=state['image_quant']
-        )
-        if result is None:
-            logger.error(f"Failed to load model `{model_name}`.")
+    try:
+        model_name = state['image_model_menu']
+
+        if not model_name or model_name == 'None':
+            logger.error("No image model selected. Go to the Model tab and select a model.")
             return []
 
-        shared.image_model_name = model_name
+        if shared.image_model is None:
+            result = load_image_model(
+                model_name,
+                dtype=state['image_dtype'],
+                attn_backend=state['image_attn_backend'],
+                cpu_offload=state['image_cpu_offload'],
+                compile_model=state['image_compile'],
+                quant_method=state['image_quant']
+            )
+            if result is None:
+                logger.error(f"Failed to load model `{model_name}`.")
+                return []
 
-    seed = state['image_seed']
-    if seed == -1:
-        seed = np.random.randint(0, 2**32 - 1)
+            shared.image_model_name = model_name
 
-    generator = torch.Generator("cuda").manual_seed(int(seed))
-    all_images = []
+        seed = state['image_seed']
+        if seed == -1:
+            seed = np.random.randint(0, 2**32 - 1)
 
-    # Get pipeline type for parameter adjustment
-    pipeline_type = getattr(shared, 'image_pipeline_type', None)
-    if pipeline_type is None:
-        pipeline_type = get_pipeline_type(shared.image_model)
+        device = get_device()
+        if device is None:
+            device = "cpu"
+        generator = torch.Generator(device).manual_seed(int(seed))
 
-    # Process Prompt
-    prompt = state['image_prompt']
+        all_images = []
 
-    # Apply "Positive Magic" for Qwen models only
-    if pipeline_type == 'qwenimage':
-        magic_suffix = ", Ultra HD, 4K, cinematic composition"
-        # Avoid duplication if user already added it
-        if magic_suffix.strip(", ") not in prompt:
-            prompt += magic_suffix
+        # Get pipeline type for parameter adjustment
+        pipeline_type = getattr(shared, 'image_pipeline_type', None)
+        if pipeline_type is None:
+            pipeline_type = get_pipeline_type(shared.image_model)
 
-    # Build generation kwargs
-    gen_kwargs = {
-        "prompt": prompt,
-        "negative_prompt": state['image_neg_prompt'],
-        "height": int(state['image_height']),
-        "width": int(state['image_width']),
-        "num_inference_steps": int(state['image_steps']),
-        "num_images_per_prompt": int(state['image_batch_size']),
-        "generator": generator,
-    }
+        # Process Prompt
+        prompt = state['image_prompt']
 
-    # Add pipeline-specific parameters for CFG
-    cfg_val = state.get('image_cfg_scale', 0.0)
+        # Apply "Positive Magic" for Qwen models only
+        if pipeline_type == 'qwenimage':
+            magic_suffix = ", Ultra HD, 4K, cinematic composition"
+            # Avoid duplication if user already added it
+            if magic_suffix.strip(", ") not in prompt:
+                prompt += magic_suffix
 
-    if pipeline_type == 'qwenimage':
-        # Qwen-Image uses true_cfg_scale (typically 4.0)
-        gen_kwargs["true_cfg_scale"] = cfg_val
-    else:
-        # Z-Image and others use guidance_scale (typically 0.0 for Turbo)
-        gen_kwargs["guidance_scale"] = cfg_val
+        # Reset stop flag at start
+        shared.stop_everything = False
 
-    t0 = time.time()
-    for i in range(int(state['image_batch_count'])):
-        generator.manual_seed(int(seed + i))
-        batch_results = shared.image_model(**gen_kwargs).images
-        all_images.extend(batch_results)
+        # Callback to check for interruption during diffusion steps
+        def interrupt_callback(pipe, step_index, timestep, callback_kwargs):
+            if shared.stop_everything:
+                pipe._interrupt = True
 
-    t1 = time.time()
-    save_generated_images(all_images, state, seed)
+            return callback_kwargs
 
-    logger.info(f'Images generated in {(t1-t0):.2f} seconds ({state["image_steps"]/(t1-t0):.2f} steps/s, seed {seed})')
-    return all_images
+        # Build generation kwargs
+        gen_kwargs = {
+            "prompt": prompt,
+            "negative_prompt": state['image_neg_prompt'],
+            "height": int(state['image_height']),
+            "width": int(state['image_width']),
+            "num_inference_steps": int(state['image_steps']),
+            "num_images_per_prompt": int(state['image_batch_size']),
+            "generator": generator,
+            "callback_on_step_end": interrupt_callback,
+        }
+
+        # Add pipeline-specific parameters for CFG
+        cfg_val = state.get('image_cfg_scale', 0.0)
+
+        if pipeline_type == 'qwenimage':
+            # Qwen-Image uses true_cfg_scale (typically 4.0)
+            gen_kwargs["true_cfg_scale"] = cfg_val
+        else:
+            # Z-Image and others use guidance_scale (typically 0.0 for Turbo)
+            gen_kwargs["guidance_scale"] = cfg_val
+
+        t0 = time.time()
+        for i in range(int(state['image_batch_count'])):
+            if shared.stop_everything:
+                break
+
+            generator.manual_seed(int(seed + i))
+            batch_results = shared.image_model(**gen_kwargs).images
+            all_images.extend(batch_results)
+
+        t1 = time.time()
+        save_generated_images(all_images, state, seed)
+
+        total_images = int(state['image_batch_count']) * int(state['image_batch_size'])
+        total_steps = state["image_steps"] * int(state['image_batch_count'])
+        logger.info(f'Generated {total_images} {"image" if total_images == 1 else "images"} in {(t1 - t0):.2f} seconds ({total_steps / (t1 - t0):.2f} steps/s, seed {seed})')
+
+        return all_images
+
+    except Exception as e:
+        logger.error(f"Image generation failed: {e}")
+        traceback.print_exc()
+        return []
 
 
 def load_image_model_wrapper(model_name, dtype, attn_backend, cpu_offload, compile_model, quant_method):
@@ -734,9 +803,10 @@ def load_image_model_wrapper(model_name, dtype, attn_backend, cpu_offload, compi
 
 
 def unload_image_model_wrapper():
+    previous_name = shared.image_model_name
     unload_image_model()
-    if shared.image_model_name != 'None':
-        return f"Model: **{shared.image_model_name}** (not loaded)"
+    if previous_name != 'None':
+        return f"Model: **{previous_name}** (unloaded)"
     return "No model loaded"
 
 

modules/utils.py

@@ -157,14 +157,7 @@ def get_available_models():
 
 def get_available_image_models():
     model_dir = Path(shared.args.image_model_dir)
-
-    # Find directories with safetensors files
-    dirs_with_safetensors = set()
-    for item in os.listdir(model_dir):
-        item_path = model_dir / item
-        if item_path.is_dir():
-            if any(file.lower().endswith(('.safetensors', '.pt')) for file in os.listdir(item_path) if (item_path / file).is_file()):
-                dirs_with_safetensors.add(item)
+    model_dir.mkdir(parents=True, exist_ok=True)
 
     # Find valid model directories
     model_dirs = []

requirements/full/requirements.txt

@@ -11,6 +11,7 @@ huggingface-hub==0.36.0
 jinja2==3.1.6
 markdown
 numpy==2.2.*
+optimum-quanto==0.2.7
 pandas
 peft==0.18.*
 Pillow>=9.5.0
@@ -34,6 +35,9 @@ wandb
 gradio==4.37.*
 https://github.com/oobabooga/gradio/releases/download/custom-build/gradio_client-1.0.2+custom.1-py3-none-any.whl
 
+# Diffusers
+diffusers @ git+https://github.com/huggingface/diffusers.git@edf36f5128abf3e6ecf92b5145115514363c58e6
+
 # API
 flask_cloudflared==0.0.14
 sse-starlette==1.6.5

requirements/full/requirements_amd.txt

@@ -9,6 +9,7 @@ huggingface-hub==0.36.0
 jinja2==3.1.6
 markdown
 numpy==2.2.*
+optimum-quanto==0.2.7
 pandas
 peft==0.18.*
 Pillow>=9.5.0
@@ -32,6 +33,9 @@ wandb
 gradio==4.37.*
 https://github.com/oobabooga/gradio/releases/download/custom-build/gradio_client-1.0.2+custom.1-py3-none-any.whl
 
+# Diffusers
+diffusers @ git+https://github.com/huggingface/diffusers.git@edf36f5128abf3e6ecf92b5145115514363c58e6
+
 # API
 flask_cloudflared==0.0.14
 sse-starlette==1.6.5

requirements/full/requirements_amd_noavx2.txt

@@ -9,6 +9,7 @@ huggingface-hub==0.36.0
 jinja2==3.1.6
 markdown
 numpy==2.2.*
+optimum-quanto==0.2.7
 pandas
 peft==0.18.*
 Pillow>=9.5.0
@@ -32,6 +33,9 @@ wandb
 gradio==4.37.*
 https://github.com/oobabooga/gradio/releases/download/custom-build/gradio_client-1.0.2+custom.1-py3-none-any.whl
 
+# Diffusers
+diffusers @ git+https://github.com/huggingface/diffusers.git@edf36f5128abf3e6ecf92b5145115514363c58e6
+
 # API
 flask_cloudflared==0.0.14
 sse-starlette==1.6.5

requirements/full/requirements_apple_intel.txt

@@ -9,6 +9,7 @@ huggingface-hub==0.36.0
 jinja2==3.1.6
 markdown
 numpy==2.2.*
+optimum-quanto==0.2.7
 pandas
 peft==0.18.*
 Pillow>=9.5.0
@@ -32,6 +33,9 @@ wandb
 gradio==4.37.*
 https://github.com/oobabooga/gradio/releases/download/custom-build/gradio_client-1.0.2+custom.1-py3-none-any.whl
 
+# Diffusers
+diffusers @ git+https://github.com/huggingface/diffusers.git@edf36f5128abf3e6ecf92b5145115514363c58e6
+
 # API
 flask_cloudflared==0.0.14
 sse-starlette==1.6.5

requirements/full/requirements_apple_silicon.txt

@@ -9,6 +9,7 @@ huggingface-hub==0.36.0
 jinja2==3.1.6
 markdown
 numpy==2.2.*
+optimum-quanto==0.2.7
 pandas
 peft==0.18.*
 Pillow>=9.5.0
@@ -32,6 +33,9 @@ wandb
 gradio==4.37.*
 https://github.com/oobabooga/gradio/releases/download/custom-build/gradio_client-1.0.2+custom.1-py3-none-any.whl
 
+# Diffusers
+diffusers @ git+https://github.com/huggingface/diffusers.git@edf36f5128abf3e6ecf92b5145115514363c58e6
+
 # API
 flask_cloudflared==0.0.14
 sse-starlette==1.6.5

requirements/full/requirements_cpu_only.txt

@@ -9,6 +9,7 @@ huggingface-hub==0.36.0
 jinja2==3.1.6
 markdown
 numpy==2.2.*
+optimum-quanto==0.2.7
 pandas
 peft==0.18.*
 Pillow>=9.5.0
@@ -32,6 +33,9 @@ wandb
 gradio==4.37.*
 https://github.com/oobabooga/gradio/releases/download/custom-build/gradio_client-1.0.2+custom.1-py3-none-any.whl
 
+# Diffusers
+diffusers @ git+https://github.com/huggingface/diffusers.git@edf36f5128abf3e6ecf92b5145115514363c58e6
+
 # API
 flask_cloudflared==0.0.14
 sse-starlette==1.6.5

requirements/full/requirements_cpu_only_noavx2.txt

@@ -9,6 +9,7 @@ huggingface-hub==0.36.0
 jinja2==3.1.6
 markdown
 numpy==2.2.*
+optimum-quanto==0.2.7
 pandas
 peft==0.18.*
 Pillow>=9.5.0
@@ -32,6 +33,9 @@ wandb
 gradio==4.37.*
 https://github.com/oobabooga/gradio/releases/download/custom-build/gradio_client-1.0.2+custom.1-py3-none-any.whl
 
+# Diffusers
+diffusers @ git+https://github.com/huggingface/diffusers.git@edf36f5128abf3e6ecf92b5145115514363c58e6
+
 # API
 flask_cloudflared==0.0.14
 sse-starlette==1.6.5

requirements/full/requirements_noavx2.txt

@@ -11,6 +11,7 @@ huggingface-hub==0.36.0
 jinja2==3.1.6
 markdown
 numpy==2.2.*
+optimum-quanto==0.2.7
 pandas
 peft==0.18.*
 Pillow>=9.5.0
@@ -34,6 +35,9 @@ wandb
 gradio==4.37.*
 https://github.com/oobabooga/gradio/releases/download/custom-build/gradio_client-1.0.2+custom.1-py3-none-any.whl
 
+# Diffusers
+diffusers @ git+https://github.com/huggingface/diffusers.git@edf36f5128abf3e6ecf92b5145115514363c58e6
+
 # API
 flask_cloudflared==0.0.14
 sse-starlette==1.6.5

requirements/full/requirements_nowheels.txt

@@ -9,6 +9,7 @@ huggingface-hub==0.36.0
 jinja2==3.1.6
 markdown
 numpy==2.2.*
+optimum-quanto==0.2.7
 pandas
 peft==0.18.*
 Pillow>=9.5.0
@@ -32,6 +33,9 @@ wandb
 gradio==4.37.*
 https://github.com/oobabooga/gradio/releases/download/custom-build/gradio_client-1.0.2+custom.1-py3-none-any.whl
 
+# Diffusers
+diffusers @ git+https://github.com/huggingface/diffusers.git@edf36f5128abf3e6ecf92b5145115514363c58e6
+
 # API
 flask_cloudflared==0.0.14
 sse-starlette==1.6.5

server.py

@@ -5,6 +5,7 @@ from pathlib import Path
 
 from modules import shared
 from modules.block_requests import OpenMonkeyPatch, RequestBlocker
+from modules.image_models import load_image_model
 from modules.logging_colors import logger
 from modules.prompts import load_prompt
 
@@ -320,6 +321,22 @@ if __name__ == "__main__":
         if shared.args.lora:
             add_lora_to_model(shared.args.lora)
 
+    # Load image model if specified via CLI
+    if shared.args.image_model:
+        logger.info(f"Loading image model: {shared.args.image_model}")
+        result = load_image_model(
+            shared.args.image_model,
+            dtype=shared.settings.get('image_dtype', 'bfloat16'),
+            attn_backend=shared.settings.get('image_attn_backend', 'sdpa'),
+            cpu_offload=shared.settings.get('image_cpu_offload', False),
+            compile_model=shared.settings.get('image_compile', False),
+            quant_method=shared.settings.get('image_quant', 'none')
+        )
+        if result is not None:
+            shared.image_model_name = shared.args.image_model
+        else:
+            logger.error(f"Failed to load image model: {shared.args.image_model}")
+
     shared.generation_lock = Lock()
 
     if shared.args.idle_timeout > 0: