Research map / Reinforcement learning / RL from feedback
DPOK: Reinforcement Learning for Fine-tuning Text-to-Image Diffusion Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Fine-tune online with RL and KL regularization instead of using the original model | Stable Diffusion v1.5 · MS-CoCo (104 prompts) | ImageReward score: 0.22 -> 0.55 | Helped |
| Fine-tune online with RL and KL regularization instead of using the original model | Stable Diffusion v1.5 · MS-CoCo (104 prompts) | Aesthetic score: 5.39 -> 5.43 | Helped |
| Fine-tune online with RL and KL regularization instead of using the original model | Stable Diffusion v1.5 · Drawbench (183 prompts) | ImageReward score: 0.13 -> 0.58 | Helped |
| Fine-tune online with RL and KL regularization instead of using the original model | Stable Diffusion v1.5 · Drawbench (183 prompts) | Aesthetic score: 5.31 -> 5.35 | Helped |
| Add value learning for variance reduction (ablation) | DPOK · A dog on the moon | ImageReward: 0.86 -> 1.51 | Helped |
| Add value learning for variance reduction (ablation) | DPOK · A dog on the moon | Aesthetic score: 5.57 -> 5.60 | Helped |
| Remove KL regularization (ablation) | DPOK · A green colored rabbit | Aesthetic score | Hurt |
| Add KL-O regularization (ablation) | Supervised fine-tuning · A green colored rabbit | Aesthetic score | Helped |
| Add KL-O regularization (ablation) | Supervised fine-tuning · A green colored rabbit | ImageReward | Hurt |
| Use KL-D regularization rather than unregularized SFT (ablation) | Supervised fine-tuning · A green colored rabbit | Aesthetic score | Helped |
| Use KL-D regularization rather than unregularized SFT (ablation) | Supervised fine-tuning · A green colored rabbit | ImageReward | Hurt |
| Use KL-O rather than KL-D regularization (ablation) | Supervised fine-tuning · A green colored rabbit | Aesthetic score | Helped |
Closest papers by meaning
- Towards Better Alignment: Training Diffusion Models with Reinforcement Learning Against Sparse Rewards (CVPR 2025 · similarity 0.92)
- Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model (CVPR 2024 · similarity 0.91)
- Training Diffusion Models with Reinforcement Learning (ICLR 2024 · similarity 0.90)
- David and Goliath: Small One-step Model Beats Large Diffusion with Score Post-training (ICML 2025 · similarity 0.90)
- Self-Play Fine-tuning of Diffusion Models for Text-to-image Generation (NeurIPS 2024 · similarity 0.90)
- Directly Fine-Tuning Diffusion Models on Differentiable Rewards (ICLR 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.