HiT Learns Its Own Abstraction Hierarchy End-to-End, Hits 1.48 FID on ImageNet 256

niloofar_mire · x · 2026-11-19

HiT is an image generator trained end-to-end from pixels that learns its own hierarchy of abstractions without a separately trained encoder: it dynamically groups patches across levels, and semantic structure emerges without masks or region labels. It achieves 1.48 FID on ImageNet 256, uses 2.4x less training compute than JiT-H/16 at matched quality, and beats models with 6x more parameters on text-to-image benchmarks. Paper and code are public.

Related event: HiT: encoder-free end-to-end image generator sets new ImageNet 256 mark(6 posts)→

Original post →

More from Research

Research channel →