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 huggingface_hub import hf_hub_download

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

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"
)

model_id = "THUDM/CogVideoX-5b-I2V"
transformer = CogVideoXTransformer3DModel.from_pretrained(model_id, subfolder="transformer", torch_dtype=torch.float16)
text_encoder = T5EncoderModel.from_pretrained(model_id, subfolder="text_encoder", torch_dtype=torch.float16)
vae = AutoencoderKLCogVideoX.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float16)
tokenizer = T5Tokenizer.from_pretrained(model_id, subfolder="tokenizer")


def find_and_move_object_to_cpu():
    for obj in gc.get_objects():
        try:
            # Check if the object is a PyTorch model
            if isinstance(obj, torch.nn.Module):
                # Check if any parameter of the model is on CUDA
                if any(param.is_cuda for param in obj.parameters()):
                    print(f"Found PyTorch model on CUDA: {type(obj).__name__}")
                    # Move the model to CPU
                    obj.to('cpu')
                    print(f"Moved {type(obj).__name__} to CPU.")
                    
                # Optionally check if buffers are on CUDA
                if any(buf.is_cuda for buf in obj.buffers()):
                    print(f"Found buffer on CUDA in {type(obj).__name__}")
                    obj.to('cpu')
                    print(f"Moved buffers of {type(obj).__name__} to CPU.")

        except Exception as e:
            # Handle any exceptions if obj is not a torch model
            pass


def clear_gpu():
    """Clear GPU memory by removing tensors, freeing cache, and moving data to CPU."""
    # List memory usage before clearing
    print(f"Memory allocated before clearing: {torch.cuda.memory_allocated() / (1024 ** 2)} MB")
    print(f"Memory reserved before clearing: {torch.cuda.memory_reserved() / (1024 ** 2)} MB")

    # Move any bound tensors back to CPU if needed
    if torch.cuda.is_available():
        torch.cuda.empty_cache()
        torch.cuda.synchronize()  # Ensure that all operations are completed
        print("GPU memory cleared.")
    
    print(f"Memory allocated after clearing: {torch.cuda.memory_allocated() / (1024 ** 2)} MB")
    print(f"Memory reserved after clearing: {torch.cuda.memory_reserved() / (1024 ** 2)} MB")


def infer(image_path, prompt, orbit_type, progress=gr.Progress(track_tqdm=True)):
    
    pipe = CogVideoXImageToVideoPipeline.from_pretrained(model_id, tokenizer=tokenizer, text_encoder=text_encoder, transformer=transformer, vae=vae, torch_dtype=torch.float16)
    lora_path = "checkpoints/"
    adapter_name = None
    if orbit_type == "Left":
        weight_name = "orbit_left_lora_weights.safetensors"
    elif orbit_type == "Up":
        weight_name = "orbit_up_lora_weights.safetensors"
    lora_rank = 256

    # Generate a timestamp for adapter_name
    adapter_timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")

    pipe.load_lora_weights(lora_path, weight_name=weight_name)
    pipe.fuse_lora(lora_scale=1 / lora_rank)
    pipe.to("cuda")


    prompt = f"{prompt}. High quality, ultrarealistic detail and breath-taking movie-like camera shot."
    image = load_image(image_path)
    seed = random.randint(0, 2**8 - 1)
    
    video = pipe(
        image, 
        prompt, 
        num_inference_steps=50,  # NOT Changed
        guidance_scale=7.0,  # NOT Changed
        use_dynamic_cfg=True,
        generator=torch.Generator(device="cpu").manual_seed(seed)
    )

    find_and_move_object_to_cpu()
    pipe.to("cpu")
    clear_gpu()
    
    # Generate a timestamp for the output filename
    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
    export_to_video(video.frames[0], f"output_{timestamp}.mp4", fps=8)
    
    return f"output_{timestamp}.mp4"

with gr.Blocks() 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)