Research map / Reinforcement learning / RL from feedback
EditScore: Unlocking Online RL for Image Editing via High-Fidelity Reward Modeling
Xin Luo, Jiahao Wang, Chenyuan Wu, Shitao Xiao, Xiyan Jiang, Defu Lian, Jiajun Zhang, Dong Liu, Zheng Liu
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 |
|---|---|---|---|
| fine-tune EditScore-7B on curated editing reward data | Qwen2.5-VL-7B · EditReward-Bench | pairwise accuracy (Overall): 0.432 -> 0.659 | Improved |
| self-ensemble 4 stochastic passes (Avg@4) (ablation) | EditScore-72B · EditReward-Bench | pairwise accuracy (Overall): 0.703 -> 0.763 | Improved |
| require reasoning before the score (ablation) | EditScore-7B · EditReward-Bench | pairwise accuracy (Overall): 0.621 -> 0.659 | Improved |
| widen score range from [0,10] to [0,25] (ablation) | EditScore-7B · EditReward-Bench | pairwise accuracy (Overall): 0.605 -> 0.659 | Improved |
| widen score range from [0,25] to [0,30] (ablation) | EditScore-7B · EditReward-Bench | pairwise accuracy (Overall): 0.659 -> 0.629 | Hurt |
| use raw Qwen2.5-VL-72B as RL reward instead of EditScore (ablation) | OmniGen2 RL · GEdit-Bench | Overall: 6.89 -> 6.42 (vs EditScore-Qwen3-VL-8B) | Hurt |
| use raw Qwen2.5-VL-7B as RL reward (ablation) | OmniGen2 RL · GEdit-Bench | Overall: 6.28 -> 6.22 | No clear effect |
| RL with EditScore-7B (Avg@4) reward | OmniGen2 · GEdit-Bench | Overall: 6.28 -> 6.68 | Improved |
| switch RL reward backbone to Qwen3-VL-8B (ablation) | EditScore RL · GEdit-Bench | Overall: 6.68 -> 6.89 | Improved |
| use 4-pass self-ensemble reward in RL (ablation) | EditScore RL · RL learning curves | convergence and final performance | Improved |
| RL with EditScore on Flux-Kontext-dev | Flux-Kontext-dev · GEdit-Bench | Overall: 6.15 -> 6.87 | Improved |
| replace Flow-GRPO with TempFlow-GRPO (ablation) | EditScore RL on OmniGen2 · GEdit-Bench | Overall: 6.68 -> 7.21 | Improved |
Closest papers by meaning
- RetouchIQ: MLLM Agents for Instruction-Based Image Retouching with Generalist Reward (CVPR 2026 · similarity 0.92)
- Multi-Reward as Condition for Instruction-based Image Editing (ICLR 2025 · similarity 0.91)
- Edit-R2: Context-Aware Reinforcement Learning for Multi-Turn Image Editing (NeurIPS 2026 · similarity 0.91)
- The Promise of RL for Autoregressive Image Editing (NeurIPS 2025 · similarity 0.91)
- SpatialReward: Bridging the Perception Gap in Online RL for Image Editing via Explicit Spatial Reasoning (ICML 2026 · similarity 0.88)
- AutoEdit: Automatic Hyperparameter Tuning for Image Editing (NeurIPS 2025 · 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.
