Initial commit

.gitattributes

@@ -0,0 +1,36 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+*.JPG filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,43 @@
+__pycache__/
+*.py[cod]
+*$py.class
+*.so
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+*.mp4
+*.avi
+*.mov
+output_video.*
+*.log
+.DS_Store
+.vscode/
+.idea/
+*.swp
+*.swo
+*~
+gradio_cached_examples/
+flagged/
+
+
+*.JPG

README.md

@@ -0,0 +1,15 @@
+---
+title: Dream-wan2-2-faster-Pro
+emoji: 🎬
+colorFrom: blue
+colorTo: purple
+sdk: gradio
+sdk_version: 5.44.1  # يطابق requirements.txt
+app_file: app.py
+pinned: true
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+
+# Dream-wan2-2-faster-Pro
+مولد فيديو من صورة واقعي فائق السرعة، مدعوم بـ Wan2.2 مع Lightning LoRA وAoT لـ ZeroGPU.

USAGE.md

@@ -0,0 +1,166 @@
+# Usage Guide - WAN 2.2 Image-to-Video LoRA Demo
+
+## Quick Start
+
+### 1. Deploying to Hugging Face Spaces
+
+To deploy this demo to Hugging Face Spaces:
+
+```bash
+# Install git-lfs if not already installed
+git lfs install
+
+# Create a new Space on huggingface.co
+# Then clone your space repository
+git clone https://huggingface.co/spaces/YOUR_USERNAME/YOUR_SPACE_NAME
+cd YOUR_SPACE_NAME
+
+# Copy all files from this demo
+cp -r * YOUR_SPACE_NAME/
+
+# Commit and push
+git add .
+git commit -m "Initial commit: WAN 2.2 Image-to-Video LoRA Demo"
+git push
+```
+
+### 2. Running Locally
+
+```bash
+# Create a virtual environment
+python -m venv venv
+source venv/bin/activate  # On Windows: venv\Scripts\activate
+
+# Install dependencies
+pip install -r requirements.txt
+
+# Run the app
+python app.py
+```
+
+The app will be available at `http://localhost:7860`
+
+## Using the Demo
+
+### Basic Usage
+
+1. **Upload Image**: Click the image upload area and select an image file
+2. **Enter Prompt**: Type a description of the motion you want (e.g., "A person walking forward, cinematic")
+3. **Click Generate**: Wait for the video to be generated (first run will download the model)
+4. **View Result**: The generated video will appear in the output area
+
+### Advanced Settings
+
+Expand the "Advanced Settings" accordion to access:
+
+- **Inference Steps** (20-100): More steps = higher quality but slower generation
+  - 20-30: Fast, lower quality
+  - 50: Balanced (recommended)
+  - 80-100: Slow, highest quality
+
+- **Guidance Scale** (1.0-15.0): How closely to follow the prompt
+  - 1.0-3.0: More creative, less faithful to prompt
+  - 6.0: Balanced (recommended)
+  - 10.0-15.0: Very faithful to prompt, less creative
+
+- **Use LoRA**: Enable/disable LoRA fine-tuning
+
+- **LoRA Type**:
+  - **High-Noise**: Best for dynamic, action-heavy scenes
+  - **Low-Noise**: Best for subtle, smooth motions
+
+## Example Prompts
+
+### Good Prompts
+
+- "A cat walking through a garden, sunny day, high quality"
+- "Waves crashing on a beach, sunset lighting, cinematic"
+- "A car driving down a highway, fast motion, 4k"
+- "Smoke rising from a campfire, slow motion"
+
+### Tips for Better Results
+
+1. **Be Specific**: Include details about motion, lighting, and quality
+2. **Use Keywords**: "cinematic", "high quality", "4k", "smooth"
+3. **Describe Motion**: Clearly state what should move and how
+4. **Consider Style**: Add style descriptors like "photorealistic" or "animated"
+
+## Troubleshooting
+
+### Out of Memory Error
+
+If you encounter OOM errors:
+
+1. The model requires significant VRAM (16GB+ recommended)
+2. On Hugging Face Spaces, ensure you're using at least `gpu-medium` hardware
+3. For local runs, try reducing the number of frames or using CPU offloading
+
+### Slow Generation
+
+- First generation will be slower (model downloads)
+- Reduce inference steps for faster results
+- Ensure GPU is being used (check logs for "Loading model on cuda")
+
+### Model Not Loading
+
+If the model fails to load:
+
+1. Check your internet connection (model is ~20GB)
+2. Ensure sufficient disk space
+3. For Hugging Face Spaces, check your Space's logs
+
+## Customization
+
+### Using Your Own LoRA Files
+
+To use your own LoRA weights:
+
+1. Upload LoRA `.safetensors` files to Hugging Face
+2. Update the URLs in `app.py`:
+
+```python
+HIGH_NOISE_LORA_URL = "https://huggingface.co/YOUR_USERNAME/YOUR_REPO/resolve/main/your_lora.safetensors"
+```
+
+3. Uncomment and implement the LoRA loading code in the `generate_video` function
+
+### Changing the Model
+
+To use a different model:
+
+1. Update `MODEL_ID` in `app.py`
+2. Ensure the model is compatible with `CogVideoXImageToVideoPipeline`
+3. Adjust memory optimizations if needed
+
+## Performance Notes
+
+- **GPU (A10G/T4)**: ~2-3 minutes per video
+- **GPU (A100)**: ~1-2 minutes per video
+- **CPU**: Not recommended (20+ minutes)
+
+## API Access
+
+For programmatic access, you can use the Gradio Client:
+
+```python
+from gradio_client import Client
+
+client = Client("YOUR_USERNAME/YOUR_SPACE_NAME")
+result = client.predict(
+    image="path/to/image.jpg",
+    prompt="A cat walking",
+    api_name="/predict"
+)
+```
+
+## Credits
+
+- Model: CogVideoX by THUDM
+- Framework: Hugging Face Diffusers
+- Interface: Gradio
+
+## License
+
+Apache 2.0 - See LICENSE file for details
+
+

aoti.py

@@ -0,0 +1,35 @@
+"""
+"""
+
+from typing import cast
+
+import torch
+from huggingface_hub import hf_hub_download
+from spaces.zero.torch.aoti import ZeroGPUCompiledModel
+from spaces.zero.torch.aoti import ZeroGPUWeights
+from torch._functorch._aot_autograd.subclass_parametrization import unwrap_tensor_subclass_parameters
+
+
+def _shallow_clone_module(module: torch.nn.Module) -> torch.nn.Module:
+    clone = object.__new__(module.__class__)
+    clone.__dict__ = module.__dict__.copy()
+    clone._parameters = module._parameters.copy()
+    clone._buffers = module._buffers.copy()
+    clone._modules = {k: _shallow_clone_module(v) for k, v in module._modules.items() if v is not None}
+    return clone
+
+
+def aoti_blocks_load(module: torch.nn.Module, repo_id: str, variant: str | None = None):
+    repeated_blocks = cast(list[str], module._repeated_blocks)
+    aoti_files = {name: hf_hub_download(
+        repo_id=repo_id,
+        filename='package.pt2',
+        subfolder=name if variant is None else f'{name}.{variant}',
+    ) for name in repeated_blocks}
+    for block_name, aoti_file in aoti_files.items():
+        for block in module.modules():
+            if block.__class__.__name__ == block_name:
+                block_ = _shallow_clone_module(block)
+                unwrap_tensor_subclass_parameters(block_)
+                weights = ZeroGPUWeights(block_.state_dict())
+                block.forward = ZeroGPUCompiledModel(aoti_file, weights)

app.py

@@ -0,0 +1,263 @@
+import os
+import spaces
+import torch
+from diffusers.pipelines.wan.pipeline_wan_i2v import WanImageToVideoPipeline
+from diffusers.models.transformers.transformer_wan import WanTransformer3DModel
+from diffusers.utils.export_utils import export_to_video
+import gradio as gr
+import tempfile
+import numpy as np
+from PIL import Image
+import random
+import gc
+
+from torchao.quantization import quantize_
+from torchao.quantization import Float8DynamicActivationFloat8WeightConfig, Int8WeightOnlyConfig
+import aoti
+
+# =========================================================
+# MODEL CONFIGURATION
+# =========================================================
+MODEL_ID = "Wan-AI/Wan2.2-I2V-A14B-Diffusers" 
+HF_TOKEN = os.environ.get("HF_TOKEN")  
+
+MAX_DIM = 832
+MIN_DIM = 480
+SQUARE_DIM = 640
+MULTIPLE_OF = 16
+
+MAX_SEED = np.iinfo(np.int32).max
+
+FIXED_FPS = 16
+MIN_FRAMES_MODEL = 8
+MAX_FRAMES_MODEL = 7720
+
+MIN_DURATION = round(MIN_FRAMES_MODEL / FIXED_FPS, 1)
+MAX_DURATION = round(MAX_FRAMES_MODEL / FIXED_FPS, 1)
+
+# =========================================================
+# LOAD PIPELINE
+# =========================================================
+pipe = WanImageToVideoPipeline.from_pretrained(
+    MODEL_ID,
+    transformer=WanTransformer3DModel.from_pretrained(
+        MODEL_ID,
+        subfolder="transformer",
+        torch_dtype=torch.bfloat16,
+        device_map="cuda",
+        token=HF_TOKEN
+    ),
+    transformer_2=WanTransformer3DModel.from_pretrained(
+        MODEL_ID,
+        subfolder="transformer_2",
+        torch_dtype=torch.bfloat16,
+        device_map="cuda",
+        token=HF_TOKEN
+    ),
+    torch_dtype=torch.bfloat16,
+).to("cuda")
+
+# =========================================================
+# LOAD LORA ADAPTERS
+# =========================================================
+pipe.load_lora_weights(
+    "Kijai/WanVideo_comfy",
+    weight_name="Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors",
+    adapter_name="lightx2v"
+)
+pipe.load_lora_weights(
+    "obsxrver/wan2.2-i2v-scat",
+    weight_name="WAN2.2-I2V-HighNoise_scat-xxi-i2v.safetensors",
+    adapter_name="i2v_scat"
+)
+pipe.load_lora_weights(
+    "Kijai/WanVideo_comfy",
+    weight_name="Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors",
+    adapter_name="lightx2v_2",
+    load_into_transformer_2=True
+)
+pipe.load_lora_weights(
+    "obsxrver/wan2.2-i2v-scat",
+    weight_name="WAN2.2-I2V-LowNoise_scat-xxi-i2v.safetensors",
+    adapter_name="i2v_scat_2"
+)
+
+pipe.set_adapters(["lightx2v", "i2v_scat","lightx2v_2","i2v_scat_2"], adapter_weights=[1., 1., 1., 1.])
+pipe.fuse_lora(adapter_names=["lightx2v"], lora_scale=3., components=["transformer"])
+pipe.fuse_lora(adapter_names=["i2v_scat"], lora_scale=1., components=["transformer"])
+pipe.fuse_lora(adapter_names=["lightx2v_2"], lora_scale=1., components=["transformer_2"])
+pipe.fuse_lora(adapter_names=["i2v_scat_2"], lora_scale=1., components=["transformer_2"])
+pipe.unload_lora_weights()
+
+# =========================================================
+# QUANTIZATION & AOT OPTIMIZATION
+# =========================================================
+quantize_(pipe.text_encoder, Int8WeightOnlyConfig())
+quantize_(pipe.transformer, Float8DynamicActivationFloat8WeightConfig())
+quantize_(pipe.transformer_2, Float8DynamicActivationFloat8WeightConfig())
+
+aoti.aoti_blocks_load(pipe.transformer, 'zerogpu-aoti/Wan2', variant='fp8da')
+aoti.aoti_blocks_load(pipe.transformer_2, 'zerogpu-aoti/Wan2', variant='fp8da')
+
+# =========================================================
+# DEFAULT PROMPTS
+# =========================================================
+default_prompt_i2v = "the video cuts, in the next scene, she takes off her clothes and is nude and covered in feces, on her back with her with legs spread, looking at the camera, she defecates and rubs her pussy, no camera movement"
+default_negative_prompt = (
+    "色调艳丽, 过曝, 静态, 细节模糊不清, 字幕, 风格, 作品, 画作, 画面, 静止, 整体发灰, "
+    "最差质量, 低质量, JPEG压缩残留, 丑陋的, 残缺的, 多余的手指, 画得不好的手部, 画得不好的脸部, "
+    "畸形的, 毁容的, 形态畸形的肢体, 手指融合, 静止不动的画面, 杂乱的背景, 三条腿, 背景人很多, 倒着走"
+)
+
+# =========================================================
+# IMAGE RESIZING LOGIC
+# =========================================================
+def resize_image(image: Image.Image) -> Image.Image:
+    width, height = image.size
+    if width == height:
+        return image.resize((SQUARE_DIM, SQUARE_DIM), Image.LANCZOS)
+
+    aspect_ratio = width / height
+    MAX_ASPECT_RATIO = MAX_DIM / MIN_DIM
+    MIN_ASPECT_RATIO = MIN_DIM / MAX_DIM
+
+    image_to_resize = image
+
+    if aspect_ratio > MAX_ASPECT_RATIO:
+        crop_width = int(round(height * MAX_ASPECT_RATIO))
+        left = (width - crop_width) // 2
+        image_to_resize = image.crop((left, 0, left + crop_width, height))
+    elif aspect_ratio < MIN_ASPECT_RATIO:
+        crop_height = int(round(width / MIN_ASPECT_RATIO))
+        top = (height - crop_height) // 2
+        image_to_resize = image.crop((0, top, width, top + crop_height))
+
+    if width > height:
+        target_w = MAX_DIM
+        target_h = int(round(target_w / aspect_ratio))
+    else:
+        target_h = MAX_DIM
+        target_w = int(round(target_h * aspect_ratio))
+
+    final_w = round(target_w / MULTIPLE_OF) * MULTIPLE_OF
+    final_h = round(target_h / MULTIPLE_OF) * MULTIPLE_OF
+
+    final_w = max(MIN_DIM, min(MAX_DIM, final_w))
+    final_h = max(MIN_DIM, min(MAX_DIM, final_h))
+
+    return image_to_resize.resize((final_w, final_h), Image.LANCZOS)
+
+# =========================================================
+# UTILITY FUNCTIONS
+# =========================================================
+def get_num_frames(duration_seconds: float):
+    return 1 + int(np.clip(int(round(duration_seconds * FIXED_FPS)), MIN_FRAMES_MODEL, MAX_FRAMES_MODEL))
+
+def get_duration(
+    input_image, prompt, steps, negative_prompt,
+    duration_seconds, guidance_scale, guidance_scale_2,
+    seed, randomize_seed, progress,
+):
+    BASE_FRAMES_HEIGHT_WIDTH = 81 * 832 * 624
+    BASE_STEP_DURATION = 15
+    width, height = resize_image(input_image).size
+    frames = get_num_frames(duration_seconds)
+    factor = frames * width * height / BASE_FRAMES_HEIGHT_WIDTH
+    step_duration = BASE_STEP_DURATION * factor ** 1.5
+    return 10 + int(steps) * step_duration
+
+# =========================================================
+# MAIN GENERATION FUNCTION
+# =========================================================
+@spaces.GPU(duration=get_duration)
+def generate_video(
+    input_image,
+    prompt,
+    steps=4,
+    negative_prompt=default_negative_prompt,
+    duration_seconds=MAX_DURATION,
+    guidance_scale=1,
+    guidance_scale_2=1,
+    seed=42,
+    randomize_seed=False,
+    progress=gr.Progress(track_tqdm=True),
+):
+    if input_image is None:
+        raise gr.Error("Please upload an input image.")
+
+    num_frames = get_num_frames(duration_seconds)
+    current_seed = random.randint(0, MAX_SEED) if randomize_seed else int(seed)
+    resized_image = resize_image(input_image)
+
+    output_frames_list = pipe(
+        image=resized_image,
+        prompt=prompt,
+        negative_prompt=negative_prompt,
+        height=resized_image.height,
+        width=resized_image.width,
+        num_frames=num_frames,
+        guidance_scale=float(guidance_scale),
+        guidance_scale_2=float(guidance_scale_2),
+        num_inference_steps=int(steps),
+        generator=torch.Generator(device="cuda").manual_seed(current_seed),
+    ).frames[0]
+
+    with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmpfile:
+        video_path = tmpfile.name
+    export_to_video(output_frames_list, video_path, fps=FIXED_FPS)
+    return video_path, current_seed
+
+# =========================================================
+# GRADIO UI
+# =========================================================
+with gr.Blocks() as demo:
+    gr.Markdown("# Wan 2.2 I2V LoRA Demo")
+    gr.Markdown("Try it out ⚡")
+
+    with gr.Row():
+        with gr.Column():
+            input_image_component = gr.Image(type="pil", label="Input Image")
+            prompt_input = gr.Textbox(label="Prompt", value=default_prompt_i2v)
+            duration_seconds_input = gr.Slider(
+                minimum=MIN_DURATION, maximum=MAX_DURATION, step=0.1, value=3.5,
+                label="Duration (seconds)",
+                info=f"Model range: {MIN_FRAMES_MODEL}-{MAX_FRAMES_MODEL} frames at {FIXED_FPS}fps."
+            )
+
+            with gr.Accordion("Advanced Settings", open=False):
+                negative_prompt_input = gr.Textbox(label="Negative Prompt", value=default_negative_prompt, lines=3)
+                seed_input = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=42)
+                randomize_seed_checkbox = gr.Checkbox(label="Randomize seed", value=True)
+                steps_slider = gr.Slider(minimum=1, maximum=30, step=1, value=6, label="Inference Steps")
+                guidance_scale_input = gr.Slider(minimum=0.0, maximum=10.0, step=0.5, value=1, label="Guidance Scale (high noise)")
+                guidance_scale_2_input = gr.Slider(minimum=0.0, maximum=10.0, step=0.5, value=1, label="Guidance Scale 2 (low noise)")
+
+            generate_button = gr.Button("🎬 Generate Video", variant="primary")
+
+        with gr.Column():
+            video_output = gr.Video(label="Generated Video", autoplay=True)
+
+    ui_inputs = [
+        input_image_component, prompt_input, steps_slider,
+        negative_prompt_input, duration_seconds_input,
+        guidance_scale_input, guidance_scale_2_input,
+        seed_input, randomize_seed_checkbox
+    ]
+    generate_button.click(fn=generate_video, inputs=ui_inputs, outputs=[video_output, seed_input])
+
+    gr.Examples(
+        examples=[
+            [
+                "wan_i2v_input.JPG",
+                "POV selfie video, white cat with sunglasses standing on surfboard, relaxed smile, tropical beach behind (clear water, green hills, blue sky with clouds). Surfboard tips, cat falls into ocean, camera plunges underwater with bubbles and sunlight beams. Brief underwater view of cat’s face, then cat resurfaces, still filming selfie, playful summer vacation mood.",
+                4,
+            ],
+        ],
+        inputs=[input_image_component, prompt_input, steps_slider],
+        outputs=[video_output, seed_input],
+        fn=generate_video,
+        cache_examples="lazy"
+    )
+
+if __name__ == "__main__":
+    demo.queue().launch(mcp_server=True)

example.py

@@ -0,0 +1,294 @@
+Hugging Face's logo
+Hugging Face
+Models
+Datasets
+Spaces
+Community
+Docs
+Enterprise
+Pricing
+
+
+Spaces:
+dream2589632147
+/
+Dream-wan2-2-faster-Pro
+
+
+like
+232
+App
+Files
+Community
+2
+Dream-wan2-2-faster-Pro
+/
+app.py
+
+dream2589632147's picture
+dream2589632147
+Update app.py
+5b9c736
+verified
+5 days ago
+raw
+
+Copy download link
+history
+blame
+contribute
+delete
+
+10.2 kB
+import os
+import spaces
+import torch
+from diffusers.pipelines.wan.pipeline_wan_i2v import WanImageToVideoPipeline
+from diffusers.models.transformers.transformer_wan import WanTransformer3DModel
+from diffusers.utils.export_utils import export_to_video
+import gradio as gr
+import tempfile
+import numpy as np
+from PIL import Image
+import random
+import gc
+
+from torchao.quantization import quantize_
+from torchao.quantization import Float8DynamicActivationFloat8WeightConfig, Int8WeightOnlyConfig
+import aoti
+
+# =========================================================
+# MODEL CONFIGURATION
+# =========================================================
+MODEL_ID = "Wan-AI/Wan2.2-I2V-A14B-Diffusers"  # المسار الجديد للنموذج
+HF_TOKEN = os.environ.get("HF_TOKEN")  # ضع توكن Hugging Face هنا إذا كان النموذج خاصًا
+
+MAX_DIM = 832
+MIN_DIM = 480
+SQUARE_DIM = 640
+MULTIPLE_OF = 16
+
+MAX_SEED = np.iinfo(np.int32).max
+
+FIXED_FPS = 16
+MIN_FRAMES_MODEL = 8
+MAX_FRAMES_MODEL = 7720
+
+MIN_DURATION = round(MIN_FRAMES_MODEL / FIXED_FPS, 1)
+MAX_DURATION = round(MAX_FRAMES_MODEL / FIXED_FPS, 1)
+
+# =========================================================
+# LOAD PIPELINE
+# =========================================================
+pipe = WanImageToVideoPipeline.from_pretrained(
+    MODEL_ID,
+    transformer=WanTransformer3DModel.from_pretrained(
+        MODEL_ID,
+        subfolder="transformer",
+        torch_dtype=torch.bfloat16,
+        device_map="cuda",
+        token=HF_TOKEN
+    ),
+    transformer_2=WanTransformer3DModel.from_pretrained(
+        MODEL_ID,
+        subfolder="transformer_2",
+        torch_dtype=torch.bfloat16,
+        device_map="cuda",
+        token=HF_TOKEN
+    ),
+    torch_dtype=torch.bfloat16,
+).to("cuda")
+
+# =========================================================
+# LOAD LORA ADAPTERS
+# =========================================================
+pipe.load_lora_weights(
+    "Kijai/WanVideo_comfy",
+    weight_name="Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors",
+    adapter_name="lightx2v"
+)
+pipe.load_lora_weights(
+    "Kijai/WanVideo_comfy",
+    weight_name="Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors",
+    adapter_name="lightx2v_2",
+    load_into_transformer_2=True
+)
+
+pipe.set_adapters(["lightx2v", "lightx2v_2"], adapter_weights=[1., 1.])
+pipe.fuse_lora(adapter_names=["lightx2v"], lora_scale=3., components=["transformer"])
+pipe.fuse_lora(adapter_names=["lightx2v_2"], lora_scale=1., components=["transformer_2"])
+pipe.unload_lora_weights()
+
+# =========================================================
+# QUANTIZATION & AOT OPTIMIZATION
+# =========================================================
+quantize_(pipe.text_encoder, Int8WeightOnlyConfig())
+quantize_(pipe.transformer, Float8DynamicActivationFloat8WeightConfig())
+quantize_(pipe.transformer_2, Float8DynamicActivationFloat8WeightConfig())
+
+aoti.aoti_blocks_load(pipe.transformer, 'zerogpu-aoti/Wan2', variant='fp8da')
+aoti.aoti_blocks_load(pipe.transformer_2, 'zerogpu-aoti/Wan2', variant='fp8da')
+
+# =========================================================
+# DEFAULT PROMPTS
+# =========================================================
+default_prompt_i2v = "make this image come alive, cinematic motion, smooth animation"
+default_negative_prompt = (
+    "色调艳丽, 过曝, 静态, 细节模糊不清, 字幕, 风格, 作品, 画作, 画面, 静止, 整体发灰, "
+    "最差质量, 低质量, JPEG压缩残留, 丑陋的, 残缺的, 多余的手指, 画得不好的手部, 画得不好的脸部, "
+    "畸形的, 毁容的, 形态畸形的肢体, 手指融合, 静止不动的画面, 杂乱的背景, 三条腿, 背景人很多, 倒着走"
+)
+
+# =========================================================
+# IMAGE RESIZING LOGIC
+# =========================================================
+def resize_image(image: Image.Image) -> Image.Image:
+    width, height = image.size
+    if width == height:
+        return image.resize((SQUARE_DIM, SQUARE_DIM), Image.LANCZOS)
+
+    aspect_ratio = width / height
+    MAX_ASPECT_RATIO = MAX_DIM / MIN_DIM
+    MIN_ASPECT_RATIO = MIN_DIM / MAX_DIM
+
+    image_to_resize = image
+
+    if aspect_ratio > MAX_ASPECT_RATIO:
+        crop_width = int(round(height * MAX_ASPECT_RATIO))
+        left = (width - crop_width) // 2
+        image_to_resize = image.crop((left, 0, left + crop_width, height))
+    elif aspect_ratio < MIN_ASPECT_RATIO:
+        crop_height = int(round(width / MIN_ASPECT_RATIO))
+        top = (height - crop_height) // 2
+        image_to_resize = image.crop((0, top, width, top + crop_height))
+
+    if width > height:
+        target_w = MAX_DIM
+        target_h = int(round(target_w / aspect_ratio))
+    else:
+        target_h = MAX_DIM
+        target_w = int(round(target_h * aspect_ratio))
+
+    final_w = round(target_w / MULTIPLE_OF) * MULTIPLE_OF
+    final_h = round(target_h / MULTIPLE_OF) * MULTIPLE_OF
+
+    final_w = max(MIN_DIM, min(MAX_DIM, final_w))
+    final_h = max(MIN_DIM, min(MAX_DIM, final_h))
+
+    return image_to_resize.resize((final_w, final_h), Image.LANCZOS)
+
+# =========================================================
+# UTILITY FUNCTIONS
+# =========================================================
+def get_num_frames(duration_seconds: float):
+    return 1 + int(np.clip(int(round(duration_seconds * FIXED_FPS)), MIN_FRAMES_MODEL, MAX_FRAMES_MODEL))
+
+def get_duration(
+    input_image, prompt, steps, negative_prompt,
+    duration_seconds, guidance_scale, guidance_scale_2,
+    seed, randomize_seed, progress,
+):
+    BASE_FRAMES_HEIGHT_WIDTH = 81 * 832 * 624
+    BASE_STEP_DURATION = 15
+    width, height = resize_image(input_image).size
+    frames = get_num_frames(duration_seconds)
+    factor = frames * width * height / BASE_FRAMES_HEIGHT_WIDTH
+    step_duration = BASE_STEP_DURATION * factor ** 1.5
+    return 10 + int(steps) * step_duration
+
+# =========================================================
+# MAIN GENERATION FUNCTION
+# =========================================================
+@spaces.GPU(duration=get_duration)
+def generate_video(
+    input_image,
+    prompt,
+    steps=4,
+    negative_prompt=default_negative_prompt,
+    duration_seconds=MAX_DURATION,
+    guidance_scale=1,
+    guidance_scale_2=1,
+    seed=42,
+    randomize_seed=False,
+    progress=gr.Progress(track_tqdm=True),
+):
+    if input_image is None:
+        raise gr.Error("Please upload an input image.")
+
+    num_frames = get_num_frames(duration_seconds)
+    current_seed = random.randint(0, MAX_SEED) if randomize_seed else int(seed)
+    resized_image = resize_image(input_image)
+
+    output_frames_list = pipe(
+        image=resized_image,
+        prompt=prompt,
+        negative_prompt=negative_prompt,
+        height=resized_image.height,
+        width=resized_image.width,
+        num_frames=num_frames,
+        guidance_scale=float(guidance_scale),
+        guidance_scale_2=float(guidance_scale_2),
+        num_inference_steps=int(steps),
+        generator=torch.Generator(device="cuda").manual_seed(current_seed),
+    ).frames[0]
+
+    with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmpfile:
+        video_path = tmpfile.name
+    export_to_video(output_frames_list, video_path, fps=FIXED_FPS)
+    return video_path, current_seed
+
+# =========================================================
+# GRADIO UI
+# =========================================================
+with gr.Blocks() as demo:
+    gr.Markdown("# 🚀 Dream Wan 2.2 Faster Pro (14B) — Ultra Fast I2V with Lightning LoRA")
+    gr.Markdown("Optimized FP8 quantized pipeline with AoT blocks & 4-step fast inference ⚡")
+
+    with gr.Row():
+        with gr.Column():
+            input_image_component = gr.Image(type="pil", label="Input Image")
+            prompt_input = gr.Textbox(label="Prompt", value=default_prompt_i2v)
+            duration_seconds_input = gr.Slider(
+                minimum=MIN_DURATION, maximum=MAX_DURATION, step=0.1, value=3.5,
+                label="Duration (seconds)",
+                info=f"Model range: {MIN_FRAMES_MODEL}-{MAX_FRAMES_MODEL} frames at {FIXED_FPS}fps."
+            )
+
+            with gr.Accordion("Advanced Settings", open=False):
+                negative_prompt_input = gr.Textbox(label="Negative Prompt", value=default_negative_prompt, lines=3)
+                seed_input = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=42)
+                randomize_seed_checkbox = gr.Checkbox(label="Randomize seed", value=True)
+                steps_slider = gr.Slider(minimum=1, maximum=30, step=1, value=6, label="Inference Steps")
+                guidance_scale_input = gr.Slider(minimum=0.0, maximum=10.0, step=0.5, value=1, label="Guidance Scale (high noise)")
+                guidance_scale_2_input = gr.Slider(minimum=0.0, maximum=10.0, step=0.5, value=1, label="Guidance Scale 2 (low noise)")
+
+            generate_button = gr.Button("🎬 Generate Video", variant="primary")
+
+        with gr.Column():
+            video_output = gr.Video(label="Generated Video", autoplay=True)
+
+    ui_inputs = [
+        input_image_component, prompt_input, steps_slider,
+        negative_prompt_input, duration_seconds_input,
+        guidance_scale_input, guidance_scale_2_input,
+        seed_input, randomize_seed_checkbox
+    ]
+    generate_button.click(fn=generate_video, inputs=ui_inputs, outputs=[video_output, seed_input])
+
+    gr.Examples(
+        examples=[
+            [
+                "wan_i2v_input.JPG",
+                "POV selfie video, white cat with sunglasses standing on surfboard, relaxed smile, tropical beach behind (clear water, green hills, blue sky with clouds). Surfboard tips, cat falls into ocean, camera plunges underwater with bubbles and sunlight beams. Brief underwater view of cat’s face, then cat resurfaces, still filming selfie, playful summer vacation mood.",
+                4,
+            ],
+        ],
+        inputs=[input_image_component, prompt_input, steps_slider],
+        outputs=[video_output, seed_input],
+        fn=generate_video,
+        cache_examples="lazy"
+    )
+
+if __name__ == "__main__":
+    demo.queue().launch(mcp_server=True)
+

requirements.txt

@@ -0,0 +1,11 @@
+git+https://github.com/linoytsaban/diffusers.git@wan22-loras
+
+transformers
+accelerate
+safetensors
+sentencepiece
+peft
+ftfy
+imageio-ffmpeg
+opencv-python
+torchao==0.11.0