This repository contains an unofficial PyTorch implementation of the paper Mean Flows for One-step Generative Modeling by Zhengyang Geng, Mingyang Deng, Xingjian Bai, J. Zico Kolter, Kaiming He.

The paper introduces a new generative model called Mean Flow, which is a flow-based model that can generate high-quality images in a single step.

This codebase is based on DiT which is the official implementation of the paper Scalable Diffusion Models with Transformers.

Please install the following repositories: pytorch torchvision numpy tqdm diffusers accelerate timm

For training, please download ImageNet.

For FID evaluations please download the ImageNet reference stats:

wget https://openaipublic.blob.core.windows.net/diffusion/jul-2021/ref_batches/imagenet/256/VIRTUAL_imagenet256_labeled.npz

We have used pytorch-fid for FID evaluation.

Run

torchrun --standalone --nproc_per_node=<num gpus> sample.py --ckpt <checkpoint path>

Run

torchrun --standalone --nproc_per_node=<gpus> train.py --data-path <path to ImageNet train data> --cfg <CFG>

Alternatively, you can pre-compute the VAE encoded dataset with:

torchrun --standalone --nproc_per_node=<gpus> encode_imagenet.py --data-path <path to ImageNet train data> --out-path <new path for encoded data>

And then run:

torchrun --standalone --nproc_per_node=<gpus> train.py --data-path <new path for encoded data> --encoded-data --cfg <CFG>

The training hyperparams correspond to optimal hyperparameters found in Sec. 5 of the paper. To run with a different set of hyperparams, use train.py -h to view available options and syntax.

@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}
}
@article{Peebles2022DiT,
  title={Scalable Diffusion Models with Transformers},
  author={William Peebles and Saining Xie},
  year={2022},
  journal={arXiv preprint arXiv:2212.09748},
}

Our implementation is based on DiT, inhereting its license (LICENSE.txt) and copyright (ACKNOWLEDGEMENTS.txt).