AI

Post-Training Frontier Text-to-Image Models by Composing Preference and Rubric Rewards

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 09/10/2026, 04:18:24 UTC.

Briefings

This study introduces a post-training recipe for text-to-image models by combining Bradley-Terry preference rewards with rubric-based evaluations. The authors report that their RL-trained Flux2dev model achieved an Elo rating 69 points above the base model, while their post-trained Ideogram-4 surpassed all open-source models on the Arena leaderboard, reaching an Elo of 1223.5. The work concludes that effective rewards must cover broad user intent and resist reward hacking.

Key points

• The authors developed a post-training recipe combining Bradley-Terry preference rewards with rubric-based evaluations to improve text-to-image models.
• The post-trained Ideogram-4 model surpassed every open-source model on the Arena leaderboard, achieving an Elo rating of 1223.5.

Context

Recent text-to-image generation models have achieved remarkable visual quality, but improving them through post-training remains challenging because no single reward signal captures the full range of human preference. The study addresses this by proposing a composition strategy for heterogeneous reward signals.

Limitations

The Arena leaderboard snapshot used for performance claims is dated September 4, 2026. The authors released Arena-T2I-Training as a 1K subset of training data to support future research, but this data does not represent full-scale training conditions.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 14/11/2026, 12:16:12 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeieojb3iwaukzprmfkntqpvhshwtjfbltnseet6eltlyodgsj2dnp4