This repository is the official implementation of the following paper:
FreqPrior: Improving Video Diffusion Models with Frequency Filtering Gaussian Noise Yunlong Yuan, Yuanfan Guo, Chunwei Wang, Wei Zhang, Hang Xu, Li Zhang
We introduce FreqPrior, an effective approach to enhancing the quality of videos generated by diffusion models. Our method features a novel frequency filtering technique specifically designed for Gaussian noise, which is fundamentally a random variable rather than a signal. FreqPrior requires no additional training, introduces no learnable parameters, and can be seamlessly integrated into any video diffusion model.
git clone https://github.com/fudan-zvg/FreqPrior.git
cd FreqPrior
cd examples/AnimateDiff
conda env create -f environment.yaml
conda activate animatediff
Please refer to the official repo of AnimateDiff. The setup guide is listed here.
After downloading the base model, motion module and personalize T2I checkpoints, run the following command to generate animations with FreqPrior. The generation results is then saved to outputs folder.
python -m scripts.animate_with_freqprior \
--config "configs/prompts/freqprior_examples/RealisticVision_v2.yaml" \
--num_iters 3 \
--ratio 0.8 \
--mid_step 321 \
--save_intermediate \
--use_fp16
We perform theoretical analysis on several methods and conduct numerical experiments to validate our findings. The numerical results can be obtained by running the script covariance_error.py. Use the following command:
python covariance_error.py --filter [FILTER TYPE] --shape [(T, H, W)]
@inproceedings{yuan2025freqprior,
title={FreqPrior: Improving Video Diffusion Models with Frequency Filtering Gaussian Noise},
author={Yunlong Yuan and Yuanfan Guo and Chunwei Wang and Wei Zhang and Hang Xu and Li Zhang},
booktitle={International Conference on Learning Representations (ICLR)},
year={2025}
}The example code is built upon FreeInit and AnimateDiff. Thanks to the team for their impressive work!