app.py

import gradio as gr
import torch
import torchvision
from diffusers import I2VGenXLPipeline, DiffusionPipeline
from torchvision.transforms.functional import to_tensor
from PIL import Image

if gr.NO_RELOAD:
    n_steps = 50
    high_noise_frac = 0.8
    negative_prompt = "Distorted, discontinuous, Ugly, blurry, low resolution, motionless, static, disfigured, disconnected limbs, Ugly faces, incomplete arms"
    generator = torch.manual_seed(8888)

    base = DiffusionPipeline.from_pretrained(
        "stabilityai/stable-diffusion-xl-base-1.0", 
        torch_dtype=torch.float16, 
        variant="fp16", 
        use_safetensors=True,
    )
    refiner = DiffusionPipeline.from_pretrained(
        "stabilityai/stable-diffusion-xl-refiner-1.0",
        text_encoder_2=base.text_encoder_2,
        vae=base.vae,
        torch_dtype=torch.float16,
        use_safetensors=True,
        variant="fp16",
    )
    pipeline = I2VGenXLPipeline.from_pretrained("ali-vilab/i2vgen-xl", torch_dtype=torch.float16, variant="fp16")

    base.to("cuda")
    refiner.to("cuda")
    pipeline.to("cuda")

    base.unet = torch.compile(base.unet, mode="reduce-overhead", fullgraph=True)
    refiner.unet = torch.compile(refiner.unet, mode="reduce-overhead", fullgraph=True)
    pipeline.unet = torch.compile(pipeline.unet, mode="reduce-overhead", fullgraph=True)

def generate(prompt: str):
    image = base(
        prompt=prompt,
        num_inference_steps=n_steps,
        denoising_end=high_noise_frac,
        output_type="latent",
    ).images[0]
    image = refiner(
        prompt=prompt,
        num_inference_steps=n_steps,
        denoising_start=high_noise_frac,
        image=image,
    ).images[0]
    image = to_tensor(image)
    frames: list[Image.Image] = pipeline(
        prompt=prompt,
        image=image,
        num_inference_steps=50,
        negative_prompt=negative_prompt,
        guidance_scale=9.0,
        generator=generator,
    ).frames[0]
    frames = [to_tensor(frame.convert("RGB")).mul(255).byte().permute(1, 2, 0) for frame in frames]
    frames = torch.stack(frames)
    torchvision.io.write_video("video.mp4", frames, fps=8)
    return "video.mp4"

app = gr.Interface(
    fn=generate,
    inputs=["text"],
    outputs=gr.Video()
)

if __name__ == "__main__":
    app.launch()