Asymmetric Flow Models: SOTA 1.57 FID on ImageNet via Rank-Asymmetric Velocity Parameterization
Stanford University
AsymFlow introduces a rank-asymmetric velocity parameterization for flow-based generative models: noise prediction is constrained to a low-rank subspace while data prediction remains full-dimensional. This asymmetry addresses a fundamental tension in high-dimensional flow modeling. The method achieves 1.57 FID on ImageNet in pixel space, and when fine-tuned from pretrained latent flow models (e.g., FLUX.2 klein 9B), establishes state-of-the-art results for pixel-space text-to-image generation.
Why it matters
290 upvotes on HuggingFace Daily Papers for May 14 — top paper of the day. The rank-asymmetric parameterization advances pixel-space generation to near-latent-space quality, potentially simplifying future generative architectures.
Importance: 3/5
Top HF Daily Paper May 14 (290 upvotes); SOTA 1.57 FID on ImageNet pixel-space generation