Research map / Reinforcement learning / RL from feedback
Aligning Few-Step Diffusion Models With Dense Reward Difference Learning
Ziyi Zhang, Shen Li, Sen Zhang, Deheng Ye, Yong Luo, Miaojing Shi, Dongjing Shan, Bo Du, Dacheng Tao
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 |
|---|---|---|---|
| use SDPO instead of DDPO | SD-Turbo · Unseen prompts, 1-step | Aesthetic Score: 6.644 -> 7.062 | Improved |
| use SDPO instead of DDPO | SD-Turbo · Unseen prompts, 8-step | Aesthetic Score: 6.720 -> 8.163 | Improved |
| use SDPO instead of DDPO | SD-Turbo · Unseen prompts, 1-step | ImageReward: 1.380 -> 1.719 | Improved |
| use SDPO instead of DDPO | SD-Turbo · Unseen prompts, 8-step | PickScore: 23.29 -> 24.05 | Improved |
| use SDPO instead of DDPO | SD-Turbo · Unseen prompts, 1-step | PickScore: 22.65 -> 23.02 | Improved |
| RL finetuning with DDPO on few-step model | SD-Turbo · Unseen prompts, 1-step | PickScore: 22.69 -> 22.65 | No clear effect |
| adaptive anchor step (vs fixed anchor) for dense reward prediction (ablation) | SDPO · Reward prediction accuracy (Aesthetic Score) | cosine similarity: 0.3654 -> 0.4463 | Improved |
| adaptive anchor step (vs random anchor) (ablation) | SDPO · Reward prediction accuracy (Aesthetic Score) | cosine similarity: 0.3561 -> 0.4463 | Improved |
| add third query at anchor step (vs 2 queries & similarity) (ablation) | SDPO · Reward prediction accuracy (PickScore) | cosine similarity: 0.4117 -> 0.4853 | Improved |
| latent similarity interpolation (vs simple interpolation, 2 queries) (ablation) | SDPO · Reward prediction accuracy (Aesthetic Score) | cosine similarity: 0.3368 -> 0.3861 | Improved |
| dense reward prediction vs single final-step query copied (ablation) | SDPO · Reward prediction accuracy (Aesthetic Score) | cosine similarity: 2x10^-8 -> 0.4463 | Improved |
| sparse 3-query dense reward prediction vs full-step queries (ablation) | SDPO · Aesthetic Score, 2-step | reward | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- REBEL — REBEL: Reinforcement Learning via Regressing Relative Rewards (NeurIPS 2024) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- TMPO: Trajectory Matching Policy Optimization for Diverse and Efficient Diffusion Alignment (NeurIPS 2026 · similarity 0.91)
- Designing Reinforcement Learning for Diffusion Models: A Unified Path-Space View (NeurIPS 2026 · similarity 0.90)
- Reward Fine-Tuning Two-Step Diffusion Models via Learning Differentiable Latent-Space Surrogate Reward (CVPR 2025 · similarity 0.90)
- Diff-Instruct with Diffused Reward: Towards Principled One-step Generator RL (NeurIPS 2026 · similarity 0.90)
- Training Diffusion Models with Reinforcement Learning (ICLR 2024 · similarity 0.89)
- PRDP: Proximal Reward Difference Prediction for Large-Scale Reward Finetuning of Diffusion Models (CVPR 2024 · similarity 0.89)
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.
