Research map / Reinforcement learning / RL from feedback
B-Spar: Bayesian Sparse-Reward Modeling for RL-based Image Editing
Shusong Xu, Peiye Liu, Yongbin Liu, Bangjie Yin, Tianyi Zheng, Zhaomang Sun, Zhenyu Chen, Peng-Tao Jiang, Jian Zhang, Yuzhao Wang, Zhen Gu, Jinwei Chen and 1 more
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 Bayesian augmentation of positive samples before RL (ablation) | B-Spar (RL only) · MIT-FiveK test | HyperIQA: 64.83 -> 68.58 | Improved |
| add Gaussian reward densification on top of augmentation (ablation) | B-Spar (RL only) · MIT-FiveK test | LPIPS x100: 1.22 -> 1.17 | Improved |
| add SFT before RL (ablation) | B-Spar · MIT-FiveK test | HyperIQA: 68.63 -> 69.02 | Improved |
| use Bayesian reward modeling instead of GRPO with a learned reward model (ablation) | B-Spar · MIT-FiveK test | HyperIQA: 64.80 -> 69.02 | Improved |
| use Bayesian reward modeling instead of direct optimization (DPO) (ablation) | B-Spar · MIT-FiveK test | HyperIQA: 66.56 -> 69.02 | Improved |
| vary minimum improvement threshold delta in 1, 2, 3 (ablation) | B-Spar · MIT-FiveK test | HyperIQA: 68.75 / 69.02 / 68.69 | No clear effect |
| use overly wide Gaussian kernels (2.0) (ablation) | B-Spar · MIT-FiveK test | HyperIQA: slight drop in average performance, increased variance | Hurt |
| use Bayesian-augmented rollout instead of vanilla offline rollout (ablation) | B-Spar sampling · 5000 images action sampling | reward proportion: 0.093 -> 0.236 | Improved |
| add dense reward signals in GRPO training (ablation) | B-Spar · GRPO training reward curve | mean reward: stagnant near zero -> approximately 0.3 | Improved |
| use B-Spar instead of other agents on fidelity to expert reference | B-Spar · MIT-FiveK test | PSNR: 16.34 vs 20.75 (PhotoArt) | Hurt |
Closest papers by meaning
- RetouchIQ: MLLM Agents for Instruction-Based Image Retouching with Generalist Reward (CVPR 2026 · similarity 0.92)
- Edit-R2: Context-Aware Reinforcement Learning for Multi-Turn Image Editing (NeurIPS 2026 · similarity 0.89)
- The Promise of RL for Autoregressive Image Editing (NeurIPS 2025 · similarity 0.88)
- Multi-Reward as Condition for Instruction-based Image Editing (ICLR 2025 · similarity 0.87)
- AutoEdit: Automatic Hyperparameter Tuning for Image Editing (NeurIPS 2025 · similarity 0.87)
- Towards Better Alignment: Training Diffusion Models with Reinforcement Learning Against Sparse Rewards (CVPR 2025 · similarity 0.86)
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.
