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.
More from Multimodal
- Magnific launches Illusion, an AI video editor that detects scene elements — techhalla · 2026-11-19
- Matt Wolfe rounds up 12 under-the-radar AI breakthroughs, from phone-brained robot to silent-speech glasses — Matt Wolfe · 2026-11-19
- Fable 5.2 vs Astra 6.2: vork creates, rigs and animates game assets head-to-head — ravisparikh · 2026-11-19
- 'Cambridge-ish' 3D generation earns a perfect 10/10 in rating thread — SamuelAlbanie · 2026-11-19
- 3D generation showcase: King's College Chapel cutaway scores 8/10 — SamuelAlbanie · 2026-11-19
- A copy-paste prompt that fixes GPT-6.2's washed-out, color-graded images — Angaisb_ · 2026-11-19