Hansheng Chen introduces AsymFlow, a flow-based image generation method that retains velocity information in a low-rank subspace instead of using JiT x0-prediction. The model operates directly in pixel space without a VAE and records an FID of 1.57 on ImageNet, the highest among published pixel-space flow models. When used to finetune FLUX.2 klein, the resulting checkpoint outperforms the base model on HPSv3, DPG, and GenEval, ranking first overall on HPSv3 while delivering sharper textures and roughly 40 percent faster inference. The work was shared by researchers including Kosta Derpanis at York University.
Alexia Jolicoeur-Martineau#617
FLUX.2 Klein in pure pixel space! No VAE. AsymFlow - hyper-realistic images by working directly in pixel space rather than using compressed latent representations.
- sharper textures, superior visual fidelity - 40% faster - low-rank noise parameterization to solve high-dimensional bottlenecks
ComfyUI support incoming
https://hanshengchen.com/asymflow/
Users praise AsymFlow's record 1.57 FID on ImageNet pixel generation because it shows academia can still advance foundation model frontiers.
No Digg Deeper questions have been answered for this story yet.
New paper: AsymFlow🔥
JiT x0-prediction is not enough for pixel generation. Better keep velocity in a low-rank subspace:
- 1.57 FID on ImageNet (best pixel flow model) - Finetunes FLUX.2 klein into pixel space, beats the original on HPSv3/DPG/GenEval (#1 overall on HPSv3)
1/7
New SOTA diffusion/flow matching model for pixel generation: 1.57 FID on ImageNet with pretraining.
Key insight: new parameterization of velocity objective.
It can also finetune a pretrained latent diffusion model into a pixel diffusion model and achieves best scores for large pixel generation.
New paper: AsymFlow🔥
JiT x0-prediction is not enough for pixel generation. Better keep velocity in a low-rank subspace:
- 1.57 FID on ImageNet (best pixel flow model) - Finetunes FLUX.2 klein into pixel space, beats the original on HPSv3/DPG/GenEval (#1 overall on HPSv3)
1/7
Cool work by my former lab at Stanford on pixel-space image diffusion!
New paper: AsymFlow🔥
JiT x0-prediction is not enough for pixel generation. Better keep velocity in a low-rank subspace:
- 1.57 FID on ImageNet (best pixel flow model) - Finetunes FLUX.2 klein into pixel space, beats the original on HPSv3/DPG/GenEval (#1 overall on HPSv3)
1/7
Core method:
Velocity prediction has a data term and a noise term. AsymFlow makes them rank-asymmetric:
- Data term is full-dimensional - Noise term is in a low-rank subspace
The full-dimensional velocity is recovered analytically for flow matching training and sampling.
2/7
Alternative view:
AsymFlow predicts velocity in the low-rank subspace, x0 in the orthogonal complement
Thus, rank controls the balance. Plain x0-pred and velocity-pred are two special cases: - rank=0➡️x0 - rank=full➡️velocity
The sweet spot is a small but non-zero rank.
3/7
We finetune FLUX.2 klein into pixel AsymFLUX.2 klein.
It beats the latent base on HPSv3, DPG, and GenEval. HPSv3=10.66 ranks #1 overall.
Not just metrics: Images are visibly more realistic, with richer styles and no plastic AI textures.
Demo: http://lakonik-asymflux-2-klein.hf.space
6/7
On ImageNet, per-patch PCA subspace (rank=8) achieves the best FID, outperforming JiT (rank=0) by a clear margin.
Final FID using JiT-H/16 net: - 1.76 with plain flow matching - 1.57 with REPA Best pixel flow model so far. Achieved with plain transformers, no extra heads.
5/7
Latent-to-pixel finetuning is easy.
Align the low-rank subspace to the pretrained latent space ➡️ get a pixel AsymFlow initialized directly from the latent base.
High-level structure/semantics are preserved. Finetuning only needs to correct low-level pixel mismatches.
4/7
arXiv: https://arxiv.org/abs/2605.12964 Code: http://github.com/Lakonik/LakonLab HF model: http://huggingface.co/Lakonik/AsymFLUX.2-klein-9B Project page: http://hanshengchen.com/asymflow/
7/7
@HanshengCh It turn out, one can push the frontier of foundation models in academia after all.
@HanshengCh Excellent work!
@HanshengCh Excellent work!
