Arena's 5M-vote post-training lifts FLUX.2-dev to #2 on live T2I leaderboard
infwinston · x · 2026-10-03
Arena shares a post-training recipe for text-to-image models that combines human preference signals with explicit rubric-based rewards.
- A reward model is trained on 5 million pairwise human preference votes from Arena, while vision-language models auto-generate rubrics checking prompt adherence, constraints, and known reward-hacking behaviors
- Preference alone misses things users care about (nice image but missing objects, added elements, style drift); rubrics fill that gap
- Post-trained FLUX.2-dev gains 69 points to reach #2 on Arena's live T2I leaderboard; post-trained Ideogram 4 hits 1224, surpassing every publicly listed open-source model
Related event: Arena Combines Preference and Rubric Rewards to Post-Train FLUX.2-dev(3 posts)→
More from Multimodal
- Gemini 4 Neon (xhigh) demo: generating a clean Ana de Armas portrait as SVG — lyraxana · 2026-11-19
- First human-vs-AI pop star interview: Scobleizer talks to virtual singer NOVA — Scobleizer · 2026-11-19
- Infinidream 0.17.2 brings 1080p Electric Sheep and screen-as-installation modes — scott_draves · 2026-11-19
- Fitting 5 character LoRAs in one image: 4 methods tested, best hits 3/5 — PoleRacers · 2026-11-19
- AI-generated robot priest praying in a candlelit church goes viral on Reddit — MrWoohoo · 2026-11-19
- AI-made retro anime short 'The Gatekeeper Episode 1' nails hand-drawn cel style — oultimobuilder · 2026-11-19