AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

RebRL: Reinforcing Discrete Visual Diffusion Models with Rebalanced Timestep Credits

Mu Zhang, Tianren Ma, Yunfan Liu, Kun Hu, Qixiang Ye

CVPR 2026 · Reinforcement learning › RL from feedback · approach: RL fine-tuning of generative models

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add token-level and temporal credit rebalancing to GRPO for discrete diffusionMMaDA-8B-Base · GenEvalOverall: 0.74 -> 0.86Improved
RebRL instead of UniGRPOMMaDA-8B-Base · GenEval CountingCounting: 0.77 -> 0.81Improved
RebRL instead of diffu-GRPOMMaDA-8B-Base · GenEval PositionPosition: 0.78 -> 0.83Improved
RebRL instead of MaskGRPOMMaDA-8B-Base · GenEval ColorsColors: 0.90 -> 0.88Hurt
RebRL instead of MaskGRPOMMaDA-8B-Base · HPSv3HPSv3: 12.35 -> 13.01Improved
add temporal 1-order and 2-order rebalancing to UniGRPO baseline (ablation)UniGRPO · HPSv3 reward during trainingHPSv3 rewardImproved
add token-level rebalancing on top of temporal rebalancing (ablation)RebRL · HPSv3 reward during trainingHPSv3 rewardImproved
add rebalancing to reduce policy divergenceUniGRPO · trainingKL lossImproved
RebRL reduces training stepsUniGRPO · training efficiencytraining steps: ~40% fewer stepsImproved
fully mask rollout sequence for every update (diffu-GRPO) instead of multi-ratio maskingMMaDA-8B-Base · GenEval training curveGenEvalMixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy