import gradio as gr
import torch
import torchvision
from diffusers import I2VGenXLPipeline
from diffusers.utils.loading_utils import load_image
from PIL import Image

def generate(image: Image.Image, prompt: str):
    negative_prompt = "Distorted, discontinuous, Ugly, blurry, low resolution, motionless, static, disfigured, disconnected limbs, Ugly faces, incomplete arms"
    generator = torch.manual_seed(8888)
    image = image.convert("RGB")
    pipeline = I2VGenXLPipeline.from_pretrained("ali-vilab/i2vgen-xl", torch_dtype=torch.float16, variant="fp16")
    pipeline.enable_model_cpu_offload()
    pipeline.unet.enable_forward_chunking()
    frames = pipeline(
        prompt=prompt,
        image=image,
        num_inference_steps=50,
        negative_prompt=negative_prompt,
        guidance_scale=9.0,
        generator=generator,
        decode_chunk_size=6,
    ).frames[0]
    torchvision.io.write_video("video.mp4", frames, fps=16)
    return "video.mp4"

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

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