AnisoFlow learns direction-dependent flow targets, cutting ImageNet FID from 1.90 to 1.55

GordonWetzstein · x · 2026-11-17

AnisoFlow (Wetzstein lab et al.) lets a flow model learn a direction-dependent prediction target while keeping architecture, loss, and sampler unchanged. It ships a fully learnable anisotropic target plus theory explaining what it learns. Results: ImageNet FID improves from 1.90 to 1.55 over JiT without REPA, and pixel-space T2I removes patch artifacts even in 8-bit.

Original post →

More from Multimodal

Multimodal channel →