Research map / Image generation / Text-to-image diffusion
Leveraging Verifier-Based Reinforcement Learning in Image Editing
Hanzhong Guo, Jie Wu, Jie Liu, Yu Gao, Zilyu Ye, Linxiao Yuan, Xionghui Wang, Yizhou Yu, Weilin Huang
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 GCPO reinforcement stage after SFT (ablation) | Edit-RRM 7B · Internal reward benchmark | accuracy: 75.4% -> 82.2% | Improved |
| add GCPO reinforcement stage after SFT (ablation) | Edit-RRM 3B · Internal reward benchmark | accuracy: 69.3% -> 72.0% | Improved |
| add GCPO reinforcement stage after SFT (ablation) | Edit-RRM 7B · EditRewardBench | accuracy: 73.3% -> 78.2% | Improved |
| add external Verify filtering to Think-only SFT data (ablation) | Edit-RRM 7B · Internal reward benchmark | accuracy: 68.9% -> 75.4% | Improved |
| use principle-decomposition SFT data instead of VIESCORE prompts (ablation) | Edit-RRM 7B · Internal reward benchmark | accuracy: 68.3% -> 75.4% | Improved |
| scale reward model from 3B to 7B (ablation) | Edit-RRM · Internal reward benchmark | accuracy (Think+Verify+GCPO): 72.0% -> 82.2% | Improved |
| use verifier-based RRM instead of holistic scorer EditScore-7B | Edit-RRM SFT · EditRewardBench | accuracy: 65.9% -> 73.3% | Improved |
| use RL-RRM (7B) as GRPO reward for FLUX.Kontext | FLUX.Kontext · GEdit-Bench-EN | overall O: 5.77 -> 6.24 | Improved |
| use RL-RRM (7B) as GRPO reward for FLUX.Kontext | FLUX.Kontext · GEdit-Bench-EN | SC: 6.27 -> 6.86 | Improved |
| use RL-RRM (7B) as GRPO reward for FLUX.Kontext | FLUX.Kontext · GEdit-Bench-EN | PQ: 7.25 -> 7.20 | No clear effect |
| use GCPO-refined RRM instead of SFT-only RRM as reward (7B) (ablation) | FLUX.Kontext · GEdit-Bench-EN | overall O: 6.20 -> 6.24 | Improved |
| use GCPO-refined RRM instead of SFT-only RRM as reward (3B) (ablation) | FLUX.Kontext · GEdit-Bench-EN | overall O: 5.63 -> 6.10 | Improved |
Closest papers by meaning
- RetouchIQ: MLLM Agents for Instruction-Based Image Retouching with Generalist Reward (CVPR 2026 · similarity 0.91)
- Edit-R2: Context-Aware Reinforcement Learning for Multi-Turn Image Editing (NeurIPS 2026 · similarity 0.90)
- The Promise of RL for Autoregressive Image Editing (NeurIPS 2025 · similarity 0.90)
- Multi-Reward as Condition for Instruction-based Image Editing (ICLR 2025 · similarity 0.89)
- ReasonEdit: Towards Reasoning-Enhanced Image Editing Models (CVPR 2026 · similarity 0.88)
- SpatialReward: Bridging the Perception Gap in Online RL for Image Editing via Explicit Spatial Reasoning (ICML 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.
