Enhance-A-Video

Paper | Blog | Twitter | Reddit

This repository is the official implementation of Enhance-A-Video: Better Generated Video for Free.

🎥 Demo

demo.mp4

The video has been heavily compressed to GitHub's policy. For more demos, please visit our blog.

🔥🔥🔥News

• 2025-02-11: Release Enhance-A-Video paper: Enhance-A-Video: Better Generated Video for Free.
• 2024-12-22: Our work achieves improvements on LTX-Video and has been added to ComfyUI-LTX. Many thanks to kijai 👏!
• 2024-12-22: Our work is added to ComfyUI-Hunyuan 🚀!
• 2024-12-20: Enhance-A-Video is now available for CogVideoX and HunyuanVideo!
• 2024-12-20: We have released code and blog for Enhance-A-Video!

🎉 Method

We design an Enhance Block as a parallel branch. This branch computes the average of non-diagonal elements of temporal attention maps as cross-frame intensity (CFI). An enhanced temperature parameter multiplies the CFI to enhance the temporal attention output.

🛠️ Dependencies and Installation

Install the dependencies:

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

📜 Requirements

The following table shows the requirements for running HunyuanVideo/CogVideoX model (batch size = 1) to generate videos:

Model	Setting (height/width/frame)	Denoising step	GPU Peak Memory
HunyuanVideo	720px1280px129f	50	60GB
CogVideoX-2B	480px720px49f	50	10GB

🧱 Inference

Generate videos:

python cogvideox.py
python hunyuanvideo.py

🔗 BibTeX

@misc{luo2025enhanceavideobettergeneratedvideo,
      title={Enhance-A-Video: Better Generated Video for Free}, 
      author={Yang Luo and Xuanlei Zhao and Mengzhao Chen and Kaipeng Zhang and Wenqi Shao and Kai Wang and Zhangyang Wang and Yang You},
      year={2025},
      eprint={2502.07508},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2502.07508}, 
}