HiT image generator learns its own abstraction hierarchy, hits 1.48 FID on ImageNet 256

LucaAmb · x · 2026-11-19

HiT trains end-to-end from pixels, dynamically grouping patches across levels so semantic structure emerges without masks or region labels. It reaches 1.48 FID on ImageNet 256, uses 2.4× less training compute than JiT-H/16, and beats models with 6× more parameters on text-to-image benchmarks.

Original post →

More from Multimodal

Multimodal channel →