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.
More from Multimodal
- AI-generated Roman emperor marble bust shows striking sculptural realism — evilrabbit_ · 2026-11-17
- MiniMax-H3 V2V Masterclass Released With Eight ComfyUI Workflows — No-Property3068 · 2026-11-17
- MiniMax-H3 V2V Masterclass Released With Eight ComfyUI Workflows — No-Property3068 · 2026-11-17
- AI-generated 'Rowan Atkinson as 007' poster amuses X, Bean-as-Bond goes viral — bradneuberg · 2026-11-17
- Sprite stacking trick turns pixel art into any angle, confirmed working with NB 2.1 — jnack · 2026-11-17
- Bauhaus-style Spartan King AI art shows off a striking prompt — LudovicCreator · 2026-11-17