Stanford University · 2026-05-13 · notable
AsymFlow — Stanford's Rank-Asymmetric Velocity Parameterization Hits 1.57 FID on ImageNet 256 in Pixel Space, Beats Latent FLUX.2 Klein Base on Text-to-Image
New flow-matching scheme that predicts noise in a low-rank subspace while keeping data prediction full-dimensional. Sets a new pixel-space ImageNet 256 record at 1.57 FID and outperforms latent FLUX.2 on HPSv3.
Stanford team trains a pixel-space flow model that beats latent FLUX.2 by predicting noise in a low-rank subspace.
Key specs
| License | Apache-2.0 |
|---|---|
| GitHub stars | 318 |
| Imagenet256 fid | 1.57 |
| Hpsv3 text to image | 10.66 |
What is it?
AsymFlow is a new flow-matching technique from Hansheng Chen, Jan Ackermann, Minseo Kim, Gordon Wetzstein, and Leonidas Guibas. Instead of predicting full-rank velocity, it splits the target into a low-rank noise component and a full-dimensional data component, then reconstructs the full velocity analytically. Code, a HuggingFace demo, and weights all dropped today.
How does it work?
Standard flow/diffusion models predict a velocity field with the same dimensionality as the data. AsymFlow imposes a rank-asymmetric parameterization: noise prediction is constrained to a low-rank subspace while data prediction stays full-dimensional. Because the two pieces combine into the velocity in closed form, you recover the full prediction without changing the network architecture — and concentrate compute on the part of the field that needs it.
Why does it matter?
Pixel-space generation has been overshadowed by latent diffusion for years because predicting high-dimensional noise was wasteful. AsymFlow reaches 1.57 FID on ImageNet 256 — a new pixel-space record — and an HPSv3 score of 10.66 on text-to-image, beating latent FLUX.2 klein base. It also enables converting pretrained latent models into pixel-space versions by aligning low-rank subspaces, opening a path for finetuning latent diffusion checkpoints back into pixel space.
Who is it for?
Generative-model researchers and practitioners working on flow / diffusion model training
Try it
https://huggingface.co/spaces/Lakonik/AsymFLUX.2-klein