HiT learns image abstractions from raw pixels, hits 1.48 FID with 2.4x less compute
niloofar_mire · x · 2026-11-19
- Most image generators today borrow representations from separately pretrained SSL/CLIP encoders (via alignment or direct reuse), which forces a two-stage training pipeline.
- HiT shows pixel-space generation can learn its own hierarchy of abstractions end-to-end through adaptive grouping and compression — semantic structure emerges with no masks or region labels: backgrounds collapse into few groups, subjects get many, and words light up the regions they refer to.
- Results: 1.48 FID on ImageNet 256, 2.4x less training compute than JiT-H/16 at matched quality, and it beats models with 6x more parameters on average across text-to-image benchmarks.
- Website, paper, and code are released, along with a thread explaining the mechanism and emergent visualizations.
Related event: HiT: Encoder-Free End-to-End Generation Hits FID 1.48 on ImageNet 256(7 posts)→
More from Multimodal
- Gemini 4 variants take three of top four spots on BrushArena image arena — teortaxesTex · 2026-11-19
- AI-Generated 'Tour of Armenia' Video With Muscular Guide Amuses Reddit — EternalSnow05 · 2026-11-19
- One image, 5x price gap: Gemini 4 Neon costs $1.01 vs GPT-6.2 Sol at $0.19 — Angaisb_ · 2026-11-19
- Creator Shares Opening Scene of '23:17', an 8-Episode AI Mystery Anime Series — SimMonzi · 2026-11-19
- Gemini 4 Neon (xhigh) demo: generating a clean Ana de Armas portrait as SVG — lyraxana · 2026-11-19
- Infinidream 0.17.2 brings 1080p Electric Sheep and screen-as-installation modes — scott_draves · 2026-11-19