Researchers combined human preference rewards with prompt-specific rubric rewards for text-to-image post-training. They trained a reward model on 5 million pairwise human votes from Text-to-Image Arena. Vision-language models auto-generated rubrics to check prompt adherence, constraints, and reward-hacking. FLUX.2-dev gained 69 Arena points after post-training. Ideogram 4 reached a 1224 Arena score, surpassing all listed open-source models by September 4, 2026. The full reward recipe achieved a 66.0% win rate against base models in offline tests. Anti-reward-hack gating suppressed photorealism shortcuts in non-photographic prompts. Weight-space ensembling of policies improved performance without tuning.
What changed
Text-to-image models were post-trained using human preference rewards alone.
What it unlocks
Generating images that better follow prompts, respect constraints, and avoid reward-hacking behaviors.
- 5M pairwise human preference votes
- 69 Arena points gained by FLUX.2-dev
- 1224 Arena score for Ideogram 4
- 66.0% win rate with full reward recipe
What you need to act on it
- a text-to-image model
- access to human preference data
- vision-language models for rubric evaluation
Sources