Research map / Reinforcement learning / RL from feedback
TreeGRPO: Tree-Advantage GRPO for Online RL Post-Training of Diffusion Models
Zheng Ding, Weirui 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 |
|---|---|---|---|
| replace trajectory GRPO (DanceGRPO) with tree-structured sampling and advantages | SD3.5-M RL post-training · HPDv2 (trained on HPS-v2.1) | iteration time (s): 173.5 -> 72.0 | Improved |
| replace DanceGRPO with TreeGRPO | SD3.5-M · HPDv2 (trained on HPS-v2.1) | HPS-v2.1: 0.3556 -> 0.3735 | Improved |
| replace DanceGRPO with TreeGRPO | SD3.5-M · HPDv2 (trained on HPS-v2.1) | ImageReward: 1.3668 -> 1.3294 | Hurt |
| replace DanceGRPO with TreeGRPO | SD3.5-M · HPDv2 (trained on HPS-v2.1) | ClipScore: 0.3769 -> 0.3703 | Hurt |
| replace MixGRPO with TreeGRPO | SD3.5-M · HPDv2 (trained on HPS-v2.1) | HPS-v2.1 / time: 0.3649 -> 0.3735; 145.4s -> 72.0s | Improved |
| replace DanceGRPO with TreeGRPO in multi-reward training | SD3.5-M · HPDv2 (HPS-v2.1 + ClipScore) | HPS-v2.1 / time: 0.3485 -> 0.364; 184.0s -> 79.2s | Improved |
| increase branching from k=3 to k=4 (d=3) (ablation) | TreeGRPO · HPSv2 | HPSv2 / time: 0.3735 -> 0.3822; 70.0s -> 126.3s | Mixed |
| reduce branching from k=3 to k=2 (d=3) (ablation) | TreeGRPO · HPSv2 | HPSv2: 0.3735 -> 0.3271 | Hurt |
| use 2 trees per prompt instead of 1 (ablation) | TreeGRPO (k=3, d=3) · HPSv2 | HPSv2: 0.3735 -> 0.3771 | No clear effect |
| random window parameter r=0.5 instead of 0.3 or 0.7 (ablation) | TreeGRPO · HPSv2 | HPSv2: 0.3632 / 0.3576 -> 0.3735 | Improved |
| shifting window instead of random window (ablation) | TreeGRPO · ClipScore | ClipScore: 0.3703 -> 0.3738 | Mixed |
| use 0.8:0.2 reward weighting instead of equal weighting (ablation) | TreeGRPO multi-reward · HPDv2 | multi-metric balance | Improved |
Later papers that compared with this method
3 later accepted papers report TreeGRPO as a baseline in their results tables; 3 of them report a better number on most of the results they share with it.
- TMPO: Trajectory Matching Policy Optimization for Diverse and Efficient Diffusion Alignment (NeurIPS 2026 · better on 5 of 5 shared results)
- On Group Relative Policy Optimization Collapse in Agent Search: The Lazy Likelihood-Displacement (ICML 2026 · better on 3 of 3 shared results)
- PAIR: Prefix-Aware Internal Reward Model for Multi-Turn Agent Optimization (NeurIPS 2026 · better on 2 of 2 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Reinforcement Learning Meets Masked Generative Models: Mask-GRPO for Text-to-Image Generation (NeurIPS 2025 · similarity 0.90)
- GenPO: Generative Diffusion Models Meet On-Policy Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
- Training Diffusion Models with Reinforcement Learning (ICLR 2024 · similarity 0.89)
- Flow-GRPO: Training Flow Matching Models via Online RL (NeurIPS 2025 · similarity 0.89)
- Towards Better Alignment: Training Diffusion Models with Reinforcement Learning Against Sparse Rewards (CVPR 2025 · similarity 0.89)
- Designing Reinforcement Learning for Diffusion Models: A Unified Path-Space View (NeurIPS 2026 · similarity 0.89)
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.
