Research map / Reinforcement learning / RL from feedback
Optimizing Visual Generative Models via Distribution-wise Rewards
Ruihang Li, Mengde Xu, Shuyang Gu, Leigang Qu, Fuli Feng, Han Hu, Wenjie Wang
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 distribution-wise reward RL (rejection sampling) to SiT-XL/2 | SiT-XL/2 · ImageNet 256 | FID-50K: 8.30 -> 6.98 | Improved |
| add distribution-wise reward RL to SiT-XL/2 | SiT-XL/2 · ImageNet 256 | FID-50K: 8.30 -> 5.77 | Improved |
| add distribution-wise reward RL to SiT-XL/2 | SiT-XL/2 · ImageNet 256 | FD_DINOv2: 230.39 -> 164.88 | Improved |
| use distribution-wise reward instead of sample-wise ImageReward | SiT-XL/2 RL · ImageNet 256 | FID: 34.26 -> 5.77 | Improved |
| add RL-EMA merging-coefficient optimization to EDM2-XS | EDM2-XS · ImageNet 512 | FID: 3.74 -> 3.52 | Improved |
| add RL-EMA merging-coefficient optimization to EDM2-S | EDM2-S · ImageNet 512 | FID: 2.57 -> 2.52 | Improved |
| continue RL after rejection-sampling fine-tuning (RS-then-RL) (ablation) | SiT-XL/2 · ImageNet 256 | FID-50K | No clear effect |
| use batch-level instead of group-level advantage normalization (ablation) | distribution-wise RL · ImageNet 256 | convergence speed | Improved |
| use 5,000-image reference set instead of 7,500 (ablation) | distribution-wise RL · ImageNet 256 | FID-50K@250 | Improved |
| replace 50 images per rollout instead of 100 or 200 (ablation) | distribution-wise RL · ImageNet 256 | generation quality | Improved |
| train on global top 25% samples instead of all samples (ablation) | distribution-wise RL · ImageNet 256 | performance | Improved |
| use ODE solver at evaluation for SDE-trained model (ablation) | SiT SDE-trained RL · ImageNet 256 | FID | Hurt |
Closest papers by meaning
- Reward-Instruct: A Reward-Centric Approach to Fast Photo-Realistic Image Generation (NeurIPS 2025 · similarity 0.90)
- MaskSense: Confronting the Visual Exploration Trap in Masked Image Generation (NeurIPS 2026 · similarity 0.89)
- Training Diffusion Models Towards Diverse Image Generation with Reinforcement Learning (CVPR 2024 · similarity 0.88)
- Diff-Instruct with Diffused Reward: Towards Principled One-step Generator RL (NeurIPS 2026 · similarity 0.88)
- Beyond the Dirac Delta: Mitigating Diversity Collapse in Reinforcement Fine-Tuning for Image Generation (NeurIPS 2026 · similarity 0.88)
- Towards Better Alignment: Training Diffusion Models with Reinforcement Learning Against Sparse Rewards (CVPR 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.
