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.
More from Multimodal
- Osmo rebrands as applied AI lab, launches Posty video agent that cuts footage in under 30 min — yawnxyz · 2026-11-19
- MOSS-Image-Base nails small real-world sign text: lost cat flyers and passive-aggressive notes — Impressive-Kick293 · 2026-11-19
- Dev ports MiniMax H3 to Swift, cuts 480p video gen to ~4 min on a 16GB M1 MacBook — Lokoto123 · 2026-11-19
- Dreamina Black Friday: unlimited Seedance 2.5 video for 30 days, credits up to 83% off — FellMentKE · 2026-11-19
- How I built a music demo with Gemini 4 and Antigravity — alexanderchen · 2026-11-19
- fal partners with Speri Labs to post-train Idol model, inviting community builds — OdinLovis · 2026-11-19