Research map / Reinforcement learning / RL from feedback
Distribution Matching Distillation Meets Reinforcement Learning
Dengyang Jiang, Dongyang Liu, Zanyi Wang, Qilong Wu, Liuzhuozheng Li, Heng-Zhuang Li, Xin Jin, Zhen Li, Changsheng Lu, Mengmeng Wang, Steven Hoi, Peng Gao 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 joint RL stage 2 to stage 1 init (ablation) | DMDR · DPG-Bench (SD3-M) | overall: 82.4371 -> 85.30 | Improved |
| use ReFL instead of DPO in stage 2 (ablation) | DMDR · DPG-Bench (SD3-M) | overall: 85.00 -> 85.30 | Improved |
| use ReFL instead of GRPO in stage 2 (ablation) | DMDR · DPG-Bench (SD3-M) | overall: 85.08 -> 85.30 | Improved |
| raise lambda_rl from 0.5 to 2.0 (ablation) | DMDR · DPG-Bench (SD3-M) | overall: 85.30 -> 82.48 | Hurt |
| lower lambda_rl from 0.5 to 0.1 (ablation) | DMDR · HP Score (SD3-M) | HP Score: 32.1145 -> 31.6565 | Hurt |
| raise lambda_rl from 0.5 to 1.0 (ablation) | DMDR · HP Score (SD3-M) | HP Score: 32.1145 -> 32.8982 | Improved |
| combine CLIP and HPS rewards instead of CLIP only (ablation) | DMDR · HP Score (SD3-M) | HP Score: 28.0507 -> 32.1145 | Improved |
| combine CLIP and HPS rewards instead of HPS only (ablation) | DMDR · DPG-Bench (SD3-M) | overall: 85.04 -> 85.30 | Improved |
| increase lambda_rl 0.1 to 1 (diversity cost) (ablation) | DMDR · LPIPS diversity (SD3-M) | LPIPS: 0.5602 -> 0.5266 | Hurt |
| add RL reward to distillation (no CFG) (ablation) | one-step SiT distillation · ImageNet 256 | FID: 2.13 -> 6.95 | Hurt |
| add CFG 4 to distillation (no RL) (ablation) | one-step SiT distillation · ImageNet 256 | FID: 2.13 -> 15.70 | Hurt |
| distill and RL jointly (DMDR) instead of distill then RL (DMD2-PSO) | SDXL-Base 4-step · ShareGPT-4o-Image prompts | HP Score: 29.8732 -> 32.9832 | Improved |
Closest papers by meaning
- GDMD: Guiding Distribution Matching Distillation with Gradient-Based Reinforcement Learning (NeurIPS 2026 · similarity 0.91)
- Designing Reinforcement Learning for Diffusion Models: A Unified Path-Space View (NeurIPS 2026 · similarity 0.90)
- TMPO: Trajectory Matching Policy Optimization for Diverse and Efficient Diffusion Alignment (NeurIPS 2026 · similarity 0.89)
- Forward KL Regularized Preference Optimization for Aligning Diffusion Policies (AAAI 2025 · similarity 0.88)
- PRDP: Proximal Reward Difference Prediction for Large-Scale Reward Finetuning of Diffusion Models (CVPR 2024 · similarity 0.88)
- Towards Better Alignment: Training Diffusion Models with Reinforcement Learning Against Sparse Rewards (CVPR 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.
