app.py

import gradio as gr
import os
import torch
import gc
from diffusers import AutoencoderKLCogVideoX, CogVideoXImageToVideoPipeline, CogVideoXTransformer3DModel
from diffusers.utils import export_to_video, load_image
from transformers import T5EncoderModel, T5Tokenizer

from datetime import datetime
import random
from moviepy.editor import VideoFileClip
import ffmpeg

from huggingface_hub import hf_hub_download

# Ensure 'checkpoint' directory exists
os.makedirs("checkpoints", exist_ok=True)

# Download LoRA weights
hf_hub_download(
    repo_id="wenqsun/DimensionX",
    filename="orbit_left_lora_weights.safetensors",
    local_dir="checkpoints"
)

hf_hub_download(
    repo_id="wenqsun/DimensionX",
    filename="orbit_up_lora_weights.safetensors",
    local_dir="checkpoints"
)

# Load models in the global scope
model_id = "THUDM/CogVideoX-5b-I2V"
transformer = CogVideoXTransformer3DModel.from_pretrained(model_id, subfolder="transformer", torch_dtype=torch.float16).to("cpu")
text_encoder = T5EncoderModel.from_pretrained(model_id, subfolder="text_encoder", torch_dtype=torch.float16).to("cpu")
vae = AutoencoderKLCogVideoX.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float16).to("cpu")
tokenizer = T5Tokenizer.from_pretrained(model_id, subfolder="tokenizer")
pipe = CogVideoXImageToVideoPipeline.from_pretrained(model_id, tokenizer=tokenizer, text_encoder=text_encoder, transformer=transformer, vae=vae, torch_dtype=torch.float16)

# Add this near the top after imports
os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'expandable_segments:True'

def calculate_resize_dimensions(width, height, max_width=1024):
    """Calculate new dimensions maintaining aspect ratio"""
    if width <= max_width:
        return width, height
    
    aspect_ratio = height / width
    new_width = max_width
    new_height = int(max_width * aspect_ratio)
    # Make height even number for video encoding
    new_height = new_height - (new_height % 2)
    return new_width, new_height

def infer(image_path, prompt, orbit_type, progress=gr.Progress(track_tqdm=True)):
    # Move everything to CPU initially
    pipe.to("cpu")
    torch.cuda.empty_cache()

    # Load and get original image dimensions
    image = load_image(image_path)
    original_width, original_height = image.size
    print(f"IMAGE INPUT SIZE: {original_width} x {original_height}")
    
    # Calculate target dimensions maintaining aspect ratio
    target_width, target_height = calculate_resize_dimensions(original_width, original_height)
    print(f"TARGET SIZE: {target_width} x {target_height}")
    
    lora_path = "checkpoints/"
    weight_name = "orbit_left_lora_weights.safetensors" if orbit_type == "Left" else "orbit_up_lora_weights.safetensors"
    lora_rank = 256
    adapter_timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")

    # Load LoRA weights on CPU
    pipe.load_lora_weights(lora_path, weight_name=weight_name, adapter_name=f"adapter_{adapter_timestamp}")
    pipe.fuse_lora(lora_scale=1 / lora_rank)
    
    try:
        # Move to GPU just before inference
        pipe.to("cuda")
        torch.cuda.empty_cache()
        
        prompt = f"{prompt}. High quality, ultrarealistic detail and breath-taking movie-like camera shot."
        seed = random.randint(0, 2**8 - 1)
        
        with torch.inference_mode():
            video = pipe(
                image, 
                prompt, 
                num_inference_steps=50,
                guidance_scale=7.0,
                use_dynamic_cfg=True,
                generator=torch.Generator(device="cpu").manual_seed(seed)
            )
    finally:
        # Ensure cleanup happens even if inference fails
        pipe.to("cpu")
        pipe.unfuse_lora()
        pipe.unload_lora_weights()
        torch.cuda.empty_cache()
        gc.collect()
    
   # Generate initial output video
    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
    temp_path = f"output_{timestamp}_temp.mp4"
    final_path = f"output_{timestamp}.mp4"
    
    # First export the original video
    export_to_video(video.frames[0], temp_path, fps=8)
    
    try:
        # Use ffmpeg-python
        stream = ffmpeg.input(temp_path)
        stream = ffmpeg.filter(stream, 'scale', target_width, target_height)
        stream = ffmpeg.output(stream, final_path, 
                             vcodec='libx264', 
                             preset='medium', 
                             crf=23, 
                             acodec='copy')
        ffmpeg.run(stream, overwrite_output=True, capture_stdout=True, capture_stderr=True)
    except ffmpeg.Error as e:
        print('stdout:', e.stdout.decode('utf8'))
        print('stderr:', e.stderr.decode('utf8'))
        raise e
    finally:
        if os.path.exists(temp_path):
            os.remove(temp_path)
    
    return final_path

# Set up Gradio UI
with gr.Blocks(analytics_enabled=False) as demo:
    with gr.Column(elem_id="col-container"):
        gr.Markdown("# DimensionX")
        gr.Markdown("### Create Any 3D and 4D Scenes from a Single Image with Controllable Video Diffusion")
        gr.HTML("""
        <div style="display:flex;column-gap:4px;">
            <a href="https://github.com/wenqsun/DimensionX">
                <img src='https://img.shields.io/badge/GitHub-Repo-blue'>
            </a> 
            <a href="https://chenshuo20.github.io/DimensionX/">
                <img src='https://img.shields.io/badge/Project-Page-green'>
            </a>
			<a href="https://arxiv.org/abs/2411.04928">
                <img src='https://img.shields.io/badge/ArXiv-Paper-red'>
            </a>
            <a href="https://huggingface.co/spaces/fffiloni/DimensionX?duplicate=true">
				<img src="https://huggingface.co/datasets/huggingface/badges/resolve/main/duplicate-this-space-sm.svg" alt="Duplicate this Space">
			</a>
			<a href="https://huggingface.co/fffiloni">
				<img src="https://huggingface.co/datasets/huggingface/badges/resolve/main/follow-me-on-HF-sm-dark.svg" alt="Follow me on HF">
			</a>
        </div>
        """)
        with gr.Row():
            with gr.Column():
                image_in = gr.Image(label="Image Input", type="filepath")
                prompt = gr.Textbox(label="Prompt")
                orbit_type = gr.Radio(label="Orbit type", choices=["Left", "Up"], value="Left", interactive=True)
                submit_btn = gr.Button("Submit")
            with gr.Column():
                video_out = gr.Video(label="Video output")
                examples = gr.Examples(
                    examples = [
                        [
                            "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg",
                            "An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space realised in the background.",
                            "Left",
                            "./examples/output_astronaut_left.mp4"
                        ],
                        [
                            "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg",
                            "An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space realised in the background.",
                            "Up",
                            "./examples/output_astronaut_up.mp4"
                        ]
                    ],
                    inputs=[image_in, prompt, orbit_type, video_out]
                )

    submit_btn.click(
        fn=infer,
        inputs=[image_in, prompt, orbit_type],
        outputs=[video_out]
    )

demo.queue().launch(show_error=True, show_api=False, ssr_mode=False)