app_t2v.py

import os
os.system('pip install --upgrade --pre --extra-index-url https://download.pytorch.org/whl/nightly/cu126 "torch<2.9" spaces')

import torch
import gradio as gr
import tempfile
import random
import numpy as np
import spaces
from diffusers import WanPipeline, AutoencoderKLWan
from diffusers.utils import export_to_video

# Constants
MODEL_ID = "Wan-AI/Wan2.2-T2V-A14B-Diffusers"
MAX_SEED = np.iinfo(np.int32).max
FIXED_FPS = 16
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"
)

# Setup
dtype = torch.float16  # using float16 for broader compatibility
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

# Load model components on correct device
vae = AutoencoderKLWan.from_pretrained(
    MODEL_ID, subfolder="vae", torch_dtype=torch.float32
).to(device)

pipe = WanPipeline.from_pretrained(
    MODEL_ID, vae=vae, torch_dtype=dtype
).to(device)

# Warm-up call to reduce cold-start latency
_ = pipe(
    prompt="warmup",
    negative_prompt=DEFAULT_NEGATIVE_PROMPT,
    height=512,
    width=768,
    num_frames=8,
    num_inference_steps=2,
    generator=torch.Generator(device=device).manual_seed(0),
).frames[0]

# Estimate duration for Hugging Face Spaces GPU usage
def get_duration(prompt, negative_prompt, height, width, num_frames, guidance_scale, guidance_scale_2, num_steps, seed, randomize_seed):
    return int(num_steps * 15)

@spaces.GPU(duration=get_duration)
def generate_video(
    prompt,
    negative_prompt,
    height,
    width,
    num_frames,
    guidance_scale,
    guidance_scale_2,
    num_steps,
    seed,
    randomize_seed
):
    current_seed = random.randint(0, MAX_SEED) if randomize_seed else int(seed)
    generator = torch.Generator(device=device).manual_seed(current_seed)

    output = pipe(
        prompt=prompt,
        negative_prompt=negative_prompt,
        height=height,
        width=width,
        num_frames=num_frames,
        guidance_scale=guidance_scale,
        guidance_scale_2=guidance_scale_2,
        num_inference_steps=num_steps,
        generator=generator,
    ).frames[0]

    with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmpfile:
        export_to_video(output, tmpfile.name, fps=FIXED_FPS)
        return tmpfile.name, current_seed

# Gradio UI
with gr.Blocks() as demo:
    gr.Markdown("## 🎬 Wan2.2 Text-to-Video Generator with Hugging Face Spaces GPU")

    with gr.Row():
        with gr.Column():
            prompt = gr.Textbox(label="Prompt", value="Two anthropomorphic cats in comfy boxing gear fight intensely.")
            negative_prompt = gr.Textbox(label="Negative Prompt", value=DEFAULT_NEGATIVE_PROMPT, lines=3)
            height = gr.Slider(360, 1024, value=720, step=16, label="Height")
            width = gr.Slider(360, 1920, value=1280, step=16, label="Width")
            num_frames = gr.Slider(8, 81, value=81, step=1, label="Number of Frames")
            num_steps = gr.Slider(10, 60, value=40, step=1, label="Inference Steps")
            guidance_scale = gr.Slider(1.0, 10.0, value=4.0, step=0.5, label="Guidance Scale")
            guidance_scale_2 = gr.Slider(1.0, 10.0, value=3.0, step=0.5, label="Guidance Scale 2")
            seed = gr.Slider(0, MAX_SEED, value=42, step=1, label="Seed")
            randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)

            generate_button = gr.Button("🎥 Generate Video")

        with gr.Column():
            video_output = gr.Video(label="Generated Video", autoplay=True, interactive=False)
            final_seed_display = gr.Number(label="Used Seed", interactive=False)

    generate_button.click(
        fn=generate_video,
        inputs=[prompt, negative_prompt, height, width, num_frames, guidance_scale, guidance_scale_2, num_steps, seed, randomize_seed],
        outputs=[video_output, final_seed_display],
    )

if __name__ == "__main__":
    demo.queue().launch()