app.py

import os
import torch
import gradio as gr
from PIL import Image, ImageOps

from huggingface_hub import snapshot_download
from pyramid_dit import PyramidDiTForVideoGeneration
from diffusers.utils import export_to_video

import spaces 

import subprocess
subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)

# Constants
MODEL_PATH = "pyramid-flow-model"
MODEL_REPO = "rain1011/pyramid-flow-sd3"
MODEL_VARIANT = "diffusion_transformer_768p"
MODEL_DTYPE = "bf16"

def center_crop(image, target_width, target_height):
    width, height = image.size
    aspect_ratio_target = target_width / target_height
    aspect_ratio_image = width / height

    if aspect_ratio_image > aspect_ratio_target:
        # Crop the width (left and right)
        new_width = int(height * aspect_ratio_target)
        left = (width - new_width) // 2
        right = left + new_width
        top, bottom = 0, height
    else:
        # Crop the height (top and bottom)
        new_height = int(width / aspect_ratio_target)
        top = (height - new_height) // 2
        bottom = top + new_height
        left, right = 0, width

    image = image.crop((left, top, right, bottom))
    return image

# Download and load the model
def load_model():
    if not os.path.exists(MODEL_PATH):
        snapshot_download(MODEL_REPO, local_dir=MODEL_PATH, local_dir_use_symlinks=False, repo_type='model')
    
    model = PyramidDiTForVideoGeneration(
        MODEL_PATH,
        MODEL_DTYPE,
        model_variant=MODEL_VARIANT,
    )
    
    model.vae.to("cuda")
    model.dit.to("cuda")
    model.text_encoder.to("cuda")
    model.vae.enable_tiling()
    
    return model

# Global model variable
model = load_model()

# Text-to-video generation function
@spaces.GPU(duration=240)
def generate_video(prompt, duration, guidance_scale, video_guidance_scale):
    temp = int(duration * 2.4)  # Convert seconds to temp value (assuming 24 FPS)
    torch_dtype = torch.bfloat16 if MODEL_DTYPE == "bf16" else torch.float32
    
    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=768,
            width=1280,
            temp=temp,
            guidance_scale=guidance_scale,
            video_guidance_scale=video_guidance_scale,
            output_type="pil",
            save_memory=True,
        )
    
    output_path = "output_video.mp4"
    export_to_video(frames, output_path, fps=24)
    return output_path

# Image-to-video generation function
@spaces.GPU(duration=240)
def generate_video_from_image(image, prompt, duration, video_guidance_scale):
    temp = int(duration * 2.4)  # Convert seconds to temp value (assuming 24 FPS)
    torch_dtype = torch.bfloat16 if MODEL_DTYPE == "bf16" else torch.float32
    
    target_size = (1280, 720)
    cropped_image = center_crop(image, 1280, 720)
    resized_image = cropped_image.resize((1280, 720))
    
    with torch.no_grad(), torch.cuda.amp.autocast(enabled=True, dtype=torch_dtype):
        frames = model.generate_i2v(
            prompt=prompt,
            input_image=resized_image,
            num_inference_steps=[10, 10, 10],
            temp=temp,
            guidance_scale=7.0,
            video_guidance_scale=video_guidance_scale,
            output_type="pil",
            save_memory=True,
        )
    
    output_path = "output_video_i2v.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():
                t2v_prompt = gr.Textbox(label="Prompt")
                t2v_duration = gr.Slider(minimum=1, maximum=10, value=5, step=1, label="Duration (seconds)")
                t2v_guidance_scale = gr.Slider(minimum=1, maximum=15, value=9, step=0.1, label="Guidance Scale")
                t2v_video_guidance_scale = gr.Slider(minimum=1, maximum=15, value=5, step=0.1, label="Video Guidance Scale")
                t2v_generate_btn = gr.Button("Generate Video")
            with gr.Column():
                t2v_output = gr.Video(label="Generated Video")
        
        t2v_generate_btn.click(
            generate_video,
            inputs=[t2v_prompt, t2v_duration, t2v_guidance_scale, t2v_video_guidance_scale],
            outputs=t2v_output
        )
    
    with gr.Tab("Image-to-Video"):
        with gr.Row():
            with gr.Column():
                i2v_image = gr.Image(type="pil", label="Input Image")
                i2v_prompt = gr.Textbox(label="Prompt")
                i2v_duration = gr.Slider(minimum=1, maximum=10, value=5, step=1, label="Duration (seconds)")
                i2v_video_guidance_scale = gr.Slider(minimum=1, maximum=15, value=4, step=0.1, label="Video Guidance Scale")
                i2v_generate_btn = gr.Button("Generate Video")
            with gr.Column():
                i2v_output = gr.Video(label="Generated Video")
        
        i2v_generate_btn.click(
            generate_video_from_image,
            inputs=[i2v_image, i2v_prompt, i2v_duration, i2v_video_guidance_scale],
            outputs=i2v_output
        )

demo.launch()