---
datasets:
- ILSVRC/imagenet-1k
license: apache-2.0
pipeline_tag: unconditional-image-generation
tags:
- flow-matching
- pixel-diffusion
- pixel-generation
---

# Asymmetric Flow Models

Pixel-space flow models trained on ImageNet using the AsymFlow method proposed in the paper:

**Asymmetric Flow Models**
<br>
arXiv 2026
<br>
[Hansheng Chen](https://lakonik.github.io/),
[Jan Ackermann](https://janackermann.info/),
[Minseo Kim](https://soniaminseokim.github.io/),
[Gordon Wetzstein](http://web.stanford.edu/~gordonwz/), 
[Leonidas Guibas](https://geometry.stanford.edu/?member=guibas)<br>
Stanford University
<br>
[Project Page](https://hanshengchen.com/asymflow) | [arXiv](https://arxiv.org/abs/2605.12964) | [Code](https://github.com/Lakonik/LakonLab/blob/main/docs/AsymFlow.md) | [AsymFLUX.2 klein Demo🤗](https://huggingface.co/spaces/Lakonik/AsymFLUX.2-klein)

![asymflow_teaser](https://cdn-uploads.huggingface.co/production/uploads/638067fcb334960c987fbeda/UCU9seMTK_iBccdFNErns.jpeg)

## Abstract
Flow-based generation in high-dimensional spaces is difficult because velocity prediction requires modeling high-dimensional noise. Asymmetric Flow Modeling (AsymFlow) is a rank-asymmetric velocity parameterization that restricts noise prediction to a low-rank subspace while keeping data prediction full-dimensional. On ImageNet 256$\times$256, AsymFlow achieves a leading 1.57 FID, outperforming prior DiT/JiT-like pixel diffusion models by a large margin.

## Citation
```
@article{chen2026asymmetric,
  title={Asymmetric Flow Models},
  author={Hansheng Chen and Jan Ackermann and Minseo Kim and Gordon Wetzstein and Leonidas Guibas},
  journal={arXiv preprint arXiv:2605.12964},
  url={https://arxiv.org/abs/2605.12964},
  year={2026},
}
```