Research map / Image generation / Text-to-image diffusion
Rethinking DPO-style Diffusion Aligning Frameworks
Xun Wu, Shaohan Huang, Lingjie Jiang, Furu Wei
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 penalty term on dispreferred sample probability (ablation) | RDPO* · Animal prompts | Aesthetic scorer: 7.28 -> 7.80 | Improved |
| Add penalty term on dispreferred sample probability (ablation) | RDPO* · Pick-a-pic prompts | Aesthetic scorer: 5.71 -> 5.90 | Improved |
| Add penalty term to D3PO (ablation) | D3PO · Animal prompts | Aesthetic scorer: 6.46 -> 7.33 | Improved |
| Add penalty term to SPO (ablation) | SPO · Animal prompts | ImageReward: 0.95 -> 1.21 | Improved |
| Add penalty term to TailorPO (ablation) | TailorPO · Animal prompts | Aesthetic scorer: 6.66 -> 7.51 | Improved |
| Add revised step-wise reward (Delta Q) to TailorPO-G (ablation) | TailorPO-G · Animal prompts | Aesthetic scorer: 6.96 -> 7.28 | Improved |
| Add revised step-wise reward (Delta Q) to TailorPO-G (ablation) | TailorPO-G · Pick-a-pic prompts | ImageReward: 0.25 -> 0.29 | Improved |
| Add revised step-wise reward (Delta Q) to TailorPO-G (ablation) | TailorPO-G · Animal prompts | HPSv2: 28.03 -> 28.57 | Improved |
| Use r* instead of r in best-of-N intermediate selection (N=8) (ablation) | SD v1.5 · Animal prompts | Aesthetic scorer: 6.02 -> 6.24 | Improved |
| Use r* instead of r in best-of-N intermediate selection (N=32) (ablation) | SD v1.5 · Animal prompts | Aesthetic scorer: 6.17 -> 6.40 | Improved |
| Increase penalty weight lambda to 10,000 (ablation) | RDPO · Animal prompts | aesthetic reward | Hurt |
| Train baseline DPO on animal prompts and test OOD on Pick-a-pic | D3PO · OOD Pick-a-pic | Aesthetic scorer: 5.28 -> 5.20 | Hurt |
Closest papers by meaning
- Smoothed Preference Optimization via ReNoise Inversion for Aligning Diffusion Models with Varied Human Preferences (ICML 2025 · similarity 0.93)
- InPO: Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment (CVPR 2025 · similarity 0.92)
- D-Fusion: Direct Preference Optimization for Aligning Diffusion Models with Visually Consistent Samples (ICML 2025 · similarity 0.91)
- Diffusion Model Alignment Using Direct Preference Optimization (CVPR 2024 · similarity 0.91)
- Calibrated Multi-Preference Optimization for Aligning Diffusion Models (CVPR 2025 · similarity 0.90)
- DyMO: Training-Free Diffusion Model Alignment with Dynamic Multi-Objective Scheduling (CVPR 2025 · similarity 0.90)
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.
