Research map / Image generation / Text-to-image diffusion
BranchGRPO: Stable and Efficient GRPO with Structured Branching in Diffusion Models
Yuming Li, Yikai Wang, Yuying zhu, Zhongyu Zhao, Ming Lu, Qi She, Shanghang Zhang
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 sequential rollouts with branching tree + depth-wise advantages | DanceGRPO (tf=1.0) · HPDv2.1 | HPS-v2.1: 0.360 -> 0.363 | Improved |
| replace sequential rollouts with branching | DanceGRPO · HPDv2.1 | iteration time (s): 698 -> 493 | Improved |
| add depth pruning (sliding window) to BranchGRPO (ablation) | BranchGRPO · HPDv2.1 | HPS-v2.1: 0.363 -> 0.369 | Improved |
| add width pruning (Parent-Top1) to BranchGRPO (ablation) | BranchGRPO · HPDv2.1 | HPS-v2.1: 0.363 -> 0.364 | No clear effect |
| add width pruning (ablation) | BranchGRPO · HPDv2.1 | iteration time (s): 493 -> 314 | Improved |
| use hybrid ODE-SDE (Mix) (ablation) | BranchGRPO · HPDv2.1 | iteration time (s): 493 -> 148 | Improved |
| path-weighted reward fusion instead of uniform averaging (ablation) | BranchGRPO · HPDv2.1 | HPS-v2.1: 0.361 -> 0.363 | Improved |
| sparser branching steps (0,5,10,15) instead of dense (0,3,6,9) (ablation) | BranchGRPO · HPDv2.1 | HPS-v2.1: 0.363 -> 0.354 | Hurt |
| moderate branch correlation s=4 vs s=1,2 or 8 (ablation) | BranchGRPO · HPDv2.1 | training reward | Mixed |
| later branching steps (9,12,15,18) instead of early (ablation) | BranchGRPO · HPDv2.1 | training reward | Hurt |
| use fewer denoising timesteps in training (tf=0.6) (ablation) | DanceGRPO · HPDv2.1 | HPS-v2.1: 0.360 -> 0.353 | Hurt |
| increase branch factor or branching steps (larger group size) (ablation) | BranchGRPO · HPDv2.1 | training reward | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MixGRPO (20,5) — MixGRPO: Unlocking Flow-based GRPO Efficiency with Mixed ODE-SDE (ECCV 2026) · better on 3 of 3 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
- MixGRPO: Unlocking Flow-based GRPO Efficiency with Mixed ODE-SDE (ECCV 2026 · similarity 0.91)
- Flash-GRPO: Efficient Alignment for Video Diffusion via One-Step Policy Optimization (ICML 2026 · similarity 0.90)
- TreeGRPO: Tree-Advantage GRPO for Online RL Post-Training of Diffusion Models (ICLR 2026 · similarity 0.90)
- TEMPFLOW-GRPO: WHEN TIMING MATTERS FOR GRPO IN FLOW MODELS (ICLR 2026 · similarity 0.90)
- Flow-GRPO: Training Flow Matching Models via Online RL (NeurIPS 2025 · similarity 0.87)
- Reinforcement Learning Meets Masked Generative Models: Mask-GRPO for Text-to-Image Generation (NeurIPS 2025 · similarity 0.87)
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.
