HiT image generator hits 1.48 FID learning hierarchy from pixels, fast reproduced

niloofar_mire · x · 2026-11-19

Albert Gu highlighted rapid independent reproductions of 2D H-Net after their blog post. The cited HiT image generator trains end-to-end from pixels, dynamically grouping patches across levels—semantic structure emerges without masks or region labels. It achieves 1.48 FID on ImageNet 256, uses 2.4× less training compute than JiT-H/16 at matched quality, and beats models with 6× more parameters on text-to-image benchmarks.

Original post →

More from Multimodal

Multimodal channel →