app.py

import torch
import gradio as gr
import spaces
from diffusers.utils import export_to_video
from diffusers import AutoencoderKLWan, WanPipeline
from diffusers.schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler
from diffusers.schedulers.scheduling_flow_match_euler_discrete import FlowMatchEulerDiscreteScheduler

# Define model options
MODEL_OPTIONS = {
    "Wan2.1-T2V-1.3B": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
    "Wan2.1-T2V-14B": "Wan-AI/Wan2.1-T2V-14B-Diffusers"
}

# Define scheduler options
SCHEDULER_OPTIONS = {
    "UniPCMultistepScheduler": UniPCMultistepScheduler,
    "FlowMatchEulerDiscreteScheduler": FlowMatchEulerDiscreteScheduler
}

@spaces.GPU(duration=300)  # Set a 5-minute duration for the GPU access
def generate_video(
    model_choice,
    prompt,
    negative_prompt,
    lora_id,
    lora_scale,
    scheduler_type,
    flow_shift,
    height,
    width,
    num_frames,
    guidance_scale,
    num_inference_steps,
    output_fps
):
    """Generate a video using the Wan model and provided parameters"""
    try:
        # Get model ID from selection
        model_id = MODEL_OPTIONS[model_choice]
        
        # Load the model components
        vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
        pipe = WanPipeline.from_pretrained(model_id, vae=vae, torch_dtype=torch.bfloat16)
        
        # Set the scheduler
        scheduler_class = SCHEDULER_OPTIONS[scheduler_type]
        
        if scheduler_type == "UniPCMultistepScheduler":
            pipe.scheduler = scheduler_class.from_config(
                pipe.scheduler.config,
                prediction_type="flow_prediction",
                use_flow_sigmas=True,
                flow_shift=flow_shift
            )
        else:
            pipe.scheduler = scheduler_class(shift=flow_shift)
        
        # Move to GPU
        pipe.to("cuda")
        
        # Enable CPU offload for low VRAM
        pipe.enable_model_cpu_offload()
        
        # Load and fuse LoRA if provided
        if lora_id and lora_id.strip():
            try:
                # Load the LoRA weights
                pipe.load_lora_weights(lora_id)
                
                # Fuse LoRA with specified scale if available
                if hasattr(pipe, "fuse_lora"):
                    pipe.fuse_lora(lora_scale=lora_scale)
            except Exception as e:
                return f"Error loading/fusing LoRA: {str(e)}"
        
        # Generate the video
        output = pipe(
            prompt=prompt,
            negative_prompt=negative_prompt,
            height=height,
            width=width,
            num_frames=num_frames,
            guidance_scale=guidance_scale,
            num_inference_steps=num_inference_steps
        ).frames[0]
        
        # Export to video
        temp_file = "output.mp4"
        export_to_video(output, temp_file, fps=output_fps)
        
        return temp_file
    except Exception as e:
        return f"Error generating video: {str(e)}"

# Create the Gradio interface
with gr.Blocks() as demo:
    gr.Markdown("# Wan Video Generation with ZeroGPU")
    gr.Markdown("Generate high-quality videos using the Wan model with optional LoRA adaptations.")
    
    with gr.Row():
        with gr.Column(scale=1):
            model_choice = gr.Dropdown(
                choices=list(MODEL_OPTIONS.keys()),
                value="Wan2.1-T2V-1.3B",
                label="Model"
            )
            
            prompt = gr.Textbox(
                label="Prompt",
                value="steamboat willie style, golden era animation, an anthropomorphic cat character wearing a hat removes it and performs a courteous bow",
                lines=3
            )
            
            negative_prompt = gr.Textbox(
                label="Negative Prompt",
                value="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",
                lines=3
            )
            
            with gr.Row():
                lora_id = gr.Textbox(
                    label="LoRA ID (e.g., benjamin-paine/steamboat-willie-1.3b)",
                    value="benjamin-paine/steamboat-willie-1.3b"
                )
                lora_scale = gr.Slider(
                    label="LoRA Scale",
                    minimum=0.0,
                    maximum=1.0,
                    value=0.75,
                    step=0.05
                )
            
            with gr.Row():
                scheduler_type = gr.Dropdown(
                    choices=list(SCHEDULER_OPTIONS.keys()),
                    value="UniPCMultistepScheduler",
                    label="Scheduler"
                )
                flow_shift = gr.Slider(
                    label="Flow Shift",
                    minimum=1.0,
                    maximum=12.0,
                    value=3.0,
                    step=0.5,
                    info="2.0-5.0 for smaller videos, 7.0-12.0 for larger videos"
                )
            
            with gr.Row():
                height = gr.Slider(
                    label="Height",
                    minimum=256,
                    maximum=1024,
                    value=480,
                    step=32
                )
                width = gr.Slider(
                    label="Width",
                    minimum=256,
                    maximum=1792,
                    value=832,
                    step=32
                )
            
            with gr.Row():
                num_frames = gr.Slider(
                    label="Number of Frames (4k+1 is recommended, e.g. 81)",
                    minimum=17,
                    maximum=129,
                    value=81,
                    step=4
                )
                output_fps = gr.Slider(
                    label="Output FPS",
                    minimum=8,
                    maximum=30,
                    value=16,
                    step=1
                )
            
            with gr.Row():
                guidance_scale = gr.Slider(
                    label="Guidance Scale (CFG)",
                    minimum=1.0,
                    maximum=15.0,
                    value=5.0,
                    step=0.5
                )
                num_inference_steps = gr.Slider(
                    label="Inference Steps",
                    minimum=10,
                    maximum=100,
                    value=32,
                    step=1
                )
            
            generate_btn = gr.Button("Generate Video")
        
        with gr.Column(scale=1):
            output_video = gr.Video(label="Generated Video")
            
    generate_btn.click(
        fn=generate_video,
        inputs=[
            model_choice,
            prompt,
            negative_prompt,
            lora_id,
            lora_scale,
            scheduler_type,
            flow_shift,
            height,
            width,
            num_frames,
            guidance_scale,
            num_inference_steps,
            output_fps
        ],
        outputs=output_video
    )
    
    gr.Markdown("""
    ## Tips for best results:
    - For smaller resolution videos, try lower values of flow shift (2.0-5.0)
    - For larger resolution videos, try higher values of flow shift (7.0-12.0)
    - Number of frames should be of the form 4k+1 (e.g., 49, 81, 65)
    - The model is memory intensive, so adjust resolution according to available VRAM
    - LoRA ID should be a Hugging Face repository containing safetensors files
    """)

demo.launch()