Research map / Reinforcement learning / RL from feedback
Entropy-Adaptive Diffusion Policy Optimization with Dynamic Step Alignment
RenYe Yan, Jikang Cheng, Yaozhong Gan, Shikun Sun, You Wu, Yunfan Yang, Liang Ling, Jinlong Lin, Yeshuang Zhu, Jie Zhou, Jinchao Zhang, Junliang Xing and 2 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 dynamic step alignment (early stop of denoising) | DDPO · Training to aesthetic score 7 | training time (hours): 11.98h -> 5.52h | Improved |
| add dynamic step alignment (early stop of denoising) | DPOK · Training to aesthetic score 7 | training time (hours): 14.76h -> 7.16h | Improved |
| add dynamic step alignment (early stop of denoising) | DDPO · Training to aesthetic score 7 | total denoising steps: 1.843e6 -> 0.242e6 | Improved |
| add dynamic step alignment (early stop of denoising) | DPOK · Training to aesthetic score 7 | total denoising steps: 2.015e6 -> 0.294e6 | Improved |
| use AdaEnt instead of DDPO | SDv1.4 aesthetic fine-tuning · Unseen metrics, Inception Score mean | Inception Score: 22.95 -> 24.34 | Improved |
| use AdaEnt instead of TDPO | SDv1.4 aesthetic fine-tuning · Unseen metrics, ImageReward mean | ImageReward: 0.8453 -> 1.1212 | Improved |
| scale entropy up to same order as KL (mu_e = 1.0) (ablation) | AdaEnt · Reward learning on Simple-animals | aesthetic reward | Hurt |
| add entropy bonus at moderate scaling factors (ablation) | AdaEnt · Reward learning on Simple-animals | aesthetic reward | Improved |
| use adaptive gamma predictor instead of fixed entropy/KL weights (ablation) | AdaEnt · Denoising trajectory | entropy-KL balance | Improved |
| use AdaEnt on SD-XL instead of DDPO, DPOK, D3PO | SD-XL aesthetic fine-tuning · SD-XL | aesthetic reward | Improved |
| use AdaEnt instead of TDPO on SDv1.5 | SDv1.5 aesthetic fine-tuning · Unseen metrics, Inception Score | Inception Score: 2.40 -> 18.99 | 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.
- DPOK — DPOK: Reinforcement Learning for Fine-tuning Text-to-Image Diffusion Models (NeurIPS 2023) · better on 8 of 8 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
- Reinforcement Learning for Diffusion LLMs with Entropy-Guided Step Selection and Stepwise Advantages (NeurIPS 2026 · similarity 0.91)
- Maximum Entropy Inverse Reinforcement Learning of Diffusion Models with Energy-Based Models (NeurIPS 2024 · similarity 0.90)
- Stochastic Control for Fine-tuning Diffusion Models: Optimality, Regularity, and Convergence (ICML 2025 · similarity 0.90)
- EVODiff: Entropy-aware Variance Optimized Diffusion Inference (NeurIPS 2025 · similarity 0.90)
- DIME: Diffusion-Based Maximum Entropy Reinforcement Learning (ICML 2025 · similarity 0.90)
- Diffusion Actor-Critic with Entropy Regulator (NeurIPS 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.
