Research map / Reinforcement learning / RL from feedback
iGRPO: Fast Online RL for Flow Matching Model with Instant Reward
Sucheng Ren, Chen Chen, Zhenbang Wang, Liangchen Song, Xiangxin Zhu, Yinfei Yang, Jiasen Lu
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 full-trajectory terminal reward with single-step instant reward (iGRPO) | FlowGRPO · GenEval, CLIPScore reward | GenEval overall: 0.74 -> 0.82 | Improved |
| use iGRPO with GenEval reward | FlowGRPO 8k iter · GenEval, GenEval reward | GenEval overall: 0.95 -> 0.96 | Improved |
| iGRPO with terminal reward instead of FlowGRPO terminal (ablation) | FlowGRPO · SD3-M GenEval | training cost / GenEval: 2.1k hours -> 860 hours; 0.74 -> 0.78 | Improved |
| add light terminal reward to instant reward (hybrid) vs instant only (ablation) | iGRPO · SD3-M GenEval | GenEval overall: 0.82 -> 0.80 | Hurt |
| instant reward only (single-step collapse) (ablation) | iGRPO · SD3-M GenEval | training cost: 2.1k hours -> 205 hours | Improved |
| use base model as KL reference instead of EMA (ablation) | iGRPO · SD3-M GenEval | GenEval overall: 0.78 -> 0.82 | Improved |
| use EMA as KL reference instead of base model (ablation) | FlowGRPO · SD3-M GenEval | GenEval overall: 0.74 -> 0.76 | Improved |
| iGRPO instead of DanceGRPO on Flux with hybrid reward | DanceGRPO · HPS-v2.1 | HPS-v2.1: 0.346 -> 0.351 | Improved |
| iGRPO instead of DanceGRPO on Flux with hybrid reward | DanceGRPO · ImageReward (out-of-domain) | ImageReward: 1.314 -> 1.441 | Improved |
Closest papers by meaning
- Flow-GRPO: Training Flow Matching Models via Online RL (NeurIPS 2025 · similarity 0.92)
- Online Reward-Weighted Fine-Tuning of Flow Matching with Wasserstein Regularization (ICLR 2025 · similarity 0.90)
- PRECISE: SDE-Consistent Stochastic Sampling for RL Post-Training of Flow-Matching Models (NeurIPS 2026 · similarity 0.89)
- Flow-Based Policy for Online Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
- Flow-DPPO: Divergence Proximal Policy Optimization for Flow Matching Models (NeurIPS 2026 · similarity 0.88)
- TMPO: Trajectory Matching Policy Optimization for Diverse and Efficient Diffusion Alignment (NeurIPS 2026 · similarity 0.88)
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.
