app.py

import gradio as gr
import torch
from diffusers import I2VGenXLPipeline
from diffusers.utils import export_to_gif, load_image
import tempfile
import spaces

@spaces.GPU
def initialize_pipeline():
    # Initialize the pipeline without CUDA support
    pipeline = I2VGenXLPipeline.from_pretrained("ali-vilab/i2vgen-xl", torch_dtype=torch.float16, variant="fp16")
    return pipeline

def generate_gif(prompt, image, negative_prompt, num_inference_steps, guidance_scale, seed):
    # Initialize the pipeline within the function
    pipeline = initialize_pipeline()

    # Set the generator seed
    generator = torch.Generator().manual_seed(seed)

    # Check if an image is provided
    if image is not None:
        image = load_image(image).convert("RGB")
        frames = pipeline(
            prompt=prompt,
            image=image,
            num_inference_steps=num_inference_steps,
            negative_prompt=negative_prompt,
            guidance_scale=guidance_scale,
            generator=generator
        ).frames[0]
    else:
        frames = pipeline(
            prompt=prompt,
            num_inference_steps=num_inference_steps,
            negative_prompt=negative_prompt,
            guidance_scale=guidance_scale,
            generator=generator
        ).frames[0]

    # Export to GIF
    with tempfile.NamedTemporaryFile(delete=False, suffix=".gif") as tmp_gif:
        gif_path = tmp_gif.name
        export_to_gif(frames, gif_path)

    return gif_path

# Create the Gradio interface with tabs
with gr.Blocks() as demo:
    with gr.Tabs():
        with gr.TabItem("Generate from Text"):
            with gr.Row():
                with gr.Column():
                    text_prompt = gr.Textbox(lines=2, placeholder="Enter your prompt here...", label="Prompt")
                    text_negative_prompt = gr.Textbox(lines=2, placeholder="Enter your negative prompt here...", label="Negative Prompt")
                    text_num_inference_steps = gr.Slider(1, 100, step=1, value=50, label="Number of Inference Steps")
                    text_guidance_scale = gr.Slider(1, 20, step=0.1, value=9.0, label="Guidance Scale")
                    text_seed = gr.Number(label="Seed", value=8888)
                    text_generate_button = gr.Button("Generate GIF")

                with gr.Column():
                    text_output_video = gr.Video(label="Generated GIF")

            # Fix: Do not use `None` in the inputs, instead pass an empty input for image
            text_generate_button.click(
                fn=generate_gif,
                inputs=[text_prompt, None, text_negative_prompt, text_num_inference_steps, text_guidance_scale, text_seed],
                outputs=text_output_video
            )

        with gr.TabItem("Generate from Image"):
            with gr.Row():
                with gr.Column():
                    image_prompt = gr.Textbox(lines=2, placeholder="Enter your prompt here...", label="Prompt")
                    image_input = gr.Image(type="filepath", label="Input Image")
                    image_negative_prompt = gr.Textbox(lines=2, placeholder="Enter your negative prompt here...", label="Negative Prompt")
                    image_num_inference_steps = gr.Slider(1, 100, step=1, value=50, label="Number of Inference Steps")
                    image_guidance_scale = gr.Slider(1, 20, step=0.1, value=9.0, label="Guidance Scale")
                    image_seed = gr.Number(label="Seed", value=8888)
                    image_generate_button = gr.Button("Generate GIF")

                with gr.Column():
                    image_output_video = gr.Video(label="Generated GIF")

            image_generate_button.click(
                fn=generate_gif,
                inputs=[image_prompt, image_input, image_negative_prompt, image_num_inference_steps, image_guidance_scale, image_seed],
                outputs=image_output_video
            )

# Launch the interface
demo.launch()