HiT trains end-to-end from pixels, hits 1.48 FID with 2.4x less compute

CatAstro_Piyush · x · 2026-11-19

HiT is an image generator trained end-to-end from pixels without a separately trained encoder. It learns its own hierarchy of abstractions, dynamically grouping patches across multiple levels—semantic structure emerges without masks or region labels. In the demo video, backgrounds collapse into a few groups, the main subject gets many, and words light up the regions they refer to.

Performance: 1.48 FID on ImageNet 256, 2.4x less training compute than JiT-H/16 at matched quality, and it outperforms models with 6x more parameters on average across text-to-image benchmarks. Website, paper, and code are released.

Original post →

More from Multimodal

Multimodal channel →