Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model

Feng Liu^1*,  Shiwei Zhang²,  Xiaofeng Wang^1,3,  Yujie Wei⁴,  Haonan Qiu⁵
Yuzhong Zhao¹,  Yingya Zhang²,  Qixiang Ye¹,  Fang Wan^1†

¹University of Chinese Academy of Sciences,  ²Alibaba Group
³Institute of Automation, Chinese Academy of Sciences
⁴Fudan University,  ⁵Nanyang Technological University

(* Work was done during internship at Alibaba Group. † Corresponding author.)

Paper | Project Page

Introduction

We introduce Timestep Embedding Aware Cache (TeaCache), a training-free caching approach that estimates and leverages the fluctuating differences among model outputs across timesteps. For more details and visual results, please visit our project page.

Installation

Prerequisites:

• Python >= 3.10
• PyTorch >= 1.13 (We recommend to use a >2.0 version)
• CUDA >= 11.6

We strongly recommend using Anaconda to create a new environment (Python >= 3.10) to run our examples:

conda create -n teacache python=3.10 -y
conda activate teacache

Install VideoSys:

git clone https://github.com/LiewFeng/TeaCache
cd TeaCache
pip install -e .

Evaluation of TeaCache

We first generate videos according to VBench's prompts.

And then calculate Vbench, PSNR, LPIPS and SSIM based on the video generated.

1. Generate video

cd eval/teacache
python experiments/latte.py
python experiments/opensora.py
python experiments/open_sora_plan.py

2. Calculate Vbench score

# vbench is calculated independently
# get scores for all metrics
python vbench/run_vbench.py --video_path aaa --save_path bbb
# calculate final score
python vbench/cal_vbench.py --score_dir bbb

3. Calculate other metrics

# these metrics are calculated compared with original model
# gt video is the video of original model
# generated video is our methods's results
python common_metrics/eval.py --gt_video_dir aa --generated_video_dir bb

Citation

@misc{liu2024timestep,
      title={Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model},
      author={Feng Liu and Shiwei Zhang and Xiaofeng Wang and Yujie Wei and Haonan Qiu and Yuzhong Zhao and Yingya Zhang and Qixiang Ye and Fang Wan},
      year={2024},
      eprint={2411.19108},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2411.19108}
}

Acknowledgement

This repository is built based on VideoSys. Thanks for their contributions!