---
license: apache-2.0
pipeline_tag: image-to-image
---

# Flow to the Mode: Mode-Seeking Diffusion Autoencoders for State-of-the-Art Image Tokenization

This repository contains **FlowMo**, a transformer-based diffusion autoencoder that achieves state-of-the-art performance for image tokenization at multiple compression rates. It is introduced in the paper [Flow to the Mode: Mode-Seeking Diffusion Autoencoders for State-of-the-Art Image Tokenization](https://huggingface.co/papers/2503.11056).

FlowMo operates without using convolutions, adversarial losses, spatially-aligned two-dimensional latent codes, or distilling from other tokenizers. Its key insight is that training should be broken into a mode-matching pre-training stage and a mode-seeking post-training stage.

<p align="center">
  <img src="https://github.com/kylesargent/FlowMo/raw/main/demo.gif" alt="FlowMo demo GIF" />
</p>

## Links

*   **Project Page:** [https://kylesargent.github.io/flowmo](https://kylesargent.github.io/flowmo)
*   **Code Repository:** [https://github.com/kylesargent/FlowMo](https://github.com/kylesargent/FlowMo)

## Usage

The official GitHub repository provides comprehensive instructions for installation, data preparation, training, and evaluation. A Jupyter notebook, `example.ipynb`, is available to demonstrate how to use the FlowMo tokenizer for image reconstruction.

## Citation

If you find FlowMo useful, please cite our paper:

```bibtex
@misc{sargent2025flowmodemodeseekingdiffusion,
      title={Flow to the Mode: Mode-Seeking Diffusion Autoencoders for State-of-the-Art Image Tokenization},
      author={Kyle Sargent and Kyle Hsu and Justin Johnson and Li Fei-Fei and Jiajun Wu},
      year={2025},
      eprint={2503.11056},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.11056},
}
```