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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add token-level and temporal credit rebalancing to GRPO for discrete diffusion | MMaDA-8B-Base · GenEval | Overall: 0.74 -> 0.86 | Improved |
| RebRL instead of UniGRPO | MMaDA-8B-Base · GenEval Counting | Counting: 0.77 -> 0.81 | Improved |
| RebRL instead of diffu-GRPO | MMaDA-8B-Base · GenEval Position | Position: 0.78 -> 0.83 | Improved |
| RebRL instead of MaskGRPO | MMaDA-8B-Base · GenEval Colors | Colors: 0.90 -> 0.88 | Hurt |
| RebRL instead of MaskGRPO | MMaDA-8B-Base · HPSv3 | HPSv3: 12.35 -> 13.01 | Improved |
| add temporal 1-order and 2-order rebalancing to UniGRPO baseline (ablation) | UniGRPO · HPSv3 reward during training | HPSv3 reward | Improved |
| add token-level rebalancing on top of temporal rebalancing (ablation) | RebRL · HPSv3 reward during training | HPSv3 reward | Improved |
| add rebalancing to reduce policy divergence | UniGRPO · training | KL loss | Improved |
| RebRL reduces training steps | UniGRPO · training efficiency | training steps: ~40% fewer steps | Improved |
| fully mask rollout sequence for every update (diffu-GRPO) instead of multi-ratio masking | MMaDA-8B-Base · GenEval training curve | GenEval | Mixed |
Closest papers by meaning
- Towards Better Alignment: Training Diffusion Models with Reinforcement Learning Against Sparse Rewards (CVPR 2025 · similarity 0.87)
- ART for Diffusion Sampling: A Reinforcement Learning Approach to Timestep Scheduling (NeurIPS 2026 · similarity 0.86)
- Beyond the Dirac Delta: Mitigating Diversity Collapse in Reinforcement Fine-Tuning for Image Generation (NeurIPS 2026 · similarity 0.86)
- Training Diffusion Models with Reinforcement Learning (ICLR 2024 · similarity 0.86)
- Designing Reinforcement Learning for Diffusion Models: A Unified Path-Space View (NeurIPS 2026 · similarity 0.86)
- Rendering-Aware Reinforcement Learning for Vector Graphics Generation (NeurIPS 2025 · similarity 0.86)
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.
