app.py

import gradio as gr
import torch
from PIL import Image
from pyramid_dit import PyramidDiTForVideoGeneration
from diffusers.utils import load_image, export_to_video
from huggingface_hub import snapshot_download
import os

# Download and load the model
model_path = 'pyramid_flow_model'
if not os.path.exists(model_path):
    snapshot_download("rain1011/pyramid-flow-sd3", local_dir=model_path, local_dir_use_symlinks=False, repo_type='model')

torch.cuda.set_device(0)
model_dtype, torch_dtype = 'bf16', torch.bfloat16

model = PyramidDiTForVideoGeneration(
    model_path,
    model_dtype,
    model_variant='diffusion_transformer_768p',
)

model.vae.to("cuda")
model.dit.to("cuda")
model.text_encoder.to("cuda")
model.vae.enable_tiling()

def generate_video(prompt, height, width, duration, guidance_scale, video_guidance_scale):
    temp = 16 if duration == "5s" else 31
    
    with torch.no_grad(), torch.cuda.amp.autocast(enabled=True, dtype=torch_dtype):
        frames = model.generate(
            prompt=prompt,
            num_inference_steps=[20, 20, 20],
            video_num_inference_steps=[10, 10, 10],
            height=height,
            width=width,
            temp=temp,
            guidance_scale=guidance_scale,
            video_guidance_scale=video_guidance_scale,
            output_type="pil",
        )
    
    output_path = "generated_video.mp4"
    export_to_video(frames, output_path, fps=24)
    return output_path

def generate_video_from_image(image, prompt, video_guidance_scale):
    with torch.no_grad(), torch.cuda.amp.autocast(enabled=True, dtype=torch_dtype):
        frames = model.generate_i2v(
            prompt=prompt,
            input_image=image,
            num_inference_steps=[10, 10, 10],
            temp=16,
            video_guidance_scale=video_guidance_scale,
            output_type="pil",
        )
    
    output_path = "generated_video_from_image.mp4"
    export_to_video(frames, output_path, fps=24)
    return output_path

# Gradio interface
with gr.Blocks() as demo:
    gr.Markdown("# Pyramid Flow Video Generation Demo")
    
    with gr.Tab("Text-to-Video"):
        with gr.Row():
            with gr.Column():
                txt_prompt = gr.Textbox(label="Prompt")
                txt_height = gr.Slider(384, 768, value=768, step=384, label="Height")
                txt_width = gr.Slider(640, 1280, value=1280, step=640, label="Width")
                txt_duration = gr.Radio(["5s", "10s"], value="5s", label="Duration")
                txt_guidance_scale = gr.Slider(1, 15, value=9, step=0.1, label="Guidance Scale")
                txt_video_guidance_scale = gr.Slider(1, 15, value=5, step=0.1, label="Video Guidance Scale")
                txt_generate = gr.Button("Generate Video")
            with gr.Column():
                txt_output = gr.Video(label="Generated Video")
    
    with gr.Tab("Image-to-Video"):
        with gr.Row():
            with gr.Column():
                img_input = gr.Image(type="pil", label="Input Image")
                img_prompt = gr.Textbox(label="Prompt (optional)")
                img_video_guidance_scale = gr.Slider(1, 15, value=4, step=0.1, label="Video Guidance Scale")
                img_generate = gr.Button("Generate Video")
            with gr.Column():
                img_output = gr.Video(label="Generated Video")
    
    txt_generate.click(generate_video, 
                       inputs=[txt_prompt, txt_height, txt_width, txt_duration, txt_guidance_scale, txt_video_guidance_scale], 
                       outputs=txt_output)
    
    img_generate.click(generate_video_from_image,
                       inputs=[img_input, img_prompt, img_video_guidance_scale],
                       outputs=img_output)

demo.launch()