FancyVideo

This repository is the official implementation of FancyVideo.

FancyVideo: Towards Dynamic and Consistent Video Generation via Cross-frame Textual Guidance
Jiasong Feng*, Ao Ma*, Jing Wang*, Bo Cheng, Xiaodan Liang, Dawei Leng†, Yuhui Yin(*Equal Contribution, ✝Corresponding Author)

Our code builds upon AnimateDiff, and we also incorporate insights from CV-VAE, Res-Adapter, and Long-CLIP to enhance our project. We appreciate the open-source contributions of these works.

🔥 News

[2024/08/19] We initialized this github repository and released the inference code and 61-frame model.
[2024/08/15] We released the paper of FancyVideo.

Quick Demos

Video demos can be found in the webpage. Some of them are contributed by the community. You can customize your own videos using the following reasoning code.

Quick Start

0. Experimental environment

We tested our inference code on a machine with a 24GB 3090 GPU and CUDA environment version 12.1.

1. Setup repository and environment

git clone https://github.com/360CVGroup/FancyVideo.git
cd FancyVideo

conda create -n fancyvideo python=3.10
conda activate fancyvideo
pip install -r requirements.txt

2. Prepare the models

# fancyvideo-ckpts & cv-vae & res-adapter & longclip & sdv1.5-base-models
git lfs install
git clone https://huggingface.co/qihoo360/FancyVideo
mv FancyVideo/resouces/models resouces 


# stable-diffusion-v1-5
git clone https://huggingface.co/runwayml/stable-diffusion-v1-5 resources/models

After download models, your resources folder is like:

📦 resouces/
├── 📂 models/
│   └── 📂 fancyvideo_ckpts/
│   └── 📂 CV-VAE/
│   └── 📂 res-adapter/
│   └── 📂 LongCLIP-L/
│   └── 📂 sd_v1-5_base_models/
│   └── 📂 stable-diffusion-v1-5/
├── 📂 demos/
│   └── 📂 reference_images/
│   └── 📂 test_prompts/

3. Customize your own videos

3.1 Image to Video

Due to the limited image generation capabilities of the SD1.5 model, we recommend generating the initial frame using a more advanced T2I model, such as SDXL, and then using our model's I2V capabilities to create the video.

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=./ python scripts/demo.py --config configs/inference/i2v.yaml

3.2 Text to Video with different base models

Our model features universal T2V capabilities and can be customized with the SD1.5 community base model.

# use the base model of pixars
CUDA_VISIBLE_DEVICES=0 PYTHONPATH=./ python scripts/demo.py --config configs/inference/t2v_pixars.yaml

# use the base model of realcartoon3d
CUDA_VISIBLE_DEVICES=0 PYTHONPATH=./ python scripts/demo.py --config configs/inference/t2v_realcartoon3d.yaml

# use the base model of toonyou
CUDA_VISIBLE_DEVICES=0 PYTHONPATH=./ python scripts/demo.py --config configs/inference/t2v_toonyou.yaml

Reference

Animatediff: https://github.com/guoyww/AnimateDiff
CV-VAE: https://github.com/AILab-CVC/CV-VAE
Res-Adapter: https://github.com/bytedance/res-adapter
Long-CLIP: https://github.com/beichenzbc/Long-CLIP

We Are Hiring

We are seeking academic interns in the AIGC field. If interested, please send your resume to maao@360.cn.

BibTeX

@misc{feng2024fancyvideodynamicconsistentvideo,
        title={FancyVideo: Towards Dynamic and Consistent Video Generation via Cross-frame Textual Guidance}, 
        author={Jiasong Feng and Ao Ma and Jing Wang and Bo Cheng and Xiaodan Liang and Dawei Leng and Yuhui Yin},
        year={2024},
        eprint={2408.08189},
        archivePrefix={arXiv},
        primaryClass={cs.CV},
        url={https://arxiv.org/abs/2408.08189}, 
}

License

This project is licensed under the Apache License (Version 2.0).

mojowebs/FancyVideo