MeanFlow: Unofficial Implementation on CIFAR-10

This repository provides an unofficial PyTorch implementation of the paper Mean Flows for One-step Generative Modeling on the CIFAR-10 dataset.

Installation

# Clone this repository
git clone https://github.com/Boo-0102/MeanFlow.git
cd MeanFlow

# Create and activate a conda environment:
conda create -n meanflow python==3.10
conda activate meanflow

# Install dependencies
pip install -r requirements.txt

Training

Unconditional

• To train an unconditional model, run the following script:

cd MeanFlow
sh scripts/train.sh

Conditional

• To train a class-conditional model with CFG, run the following script:

cd MeanFlow
sh scripts/train_cfg.sh

Sampling

Unconditional

cd MeanFlow
sh scripts/sample.sh

Conditional

cd MeanFlow
sh scripts/sample_cfg.sh

Evaluate

• To generate samples and calculate the FID score for a trained unconditional model, run:

cd MeanFlow
sh scripts/evaluate.sh

Acknowledgements

This implementation is based on or inspired by the code from the following repositories:

• SongUnet
• REPA
• MeanFlow

Citation

If you find this implementation useful, please cite the original paper:

@article{geng2025mean,
  title={Mean Flows for One-step Generative Modeling},
  author={Geng, Zhengyang and Deng, Mingyang and Bai, Xingjian and Kolter, J Zico and He, Kaiming},
  journal={arXiv preprint arXiv:2505.13447},
  year={2025}
}

License

MIT License