Research map / Image generation / Text-to-image diffusion
$\alpha$-DPO: Robust Preference Alignment for Diffusion Models via $\alpha$ Divergence
Yang Li, Songlin Yang, Wei Wang, Xiaoxuan Han, Jing Dong
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 |
|---|---|---|---|
| replace forward-KL (DPO) with alpha-divergence objective | SDXL · Pick-a-Pic test (trained on Pick-a-Pic V2) | ImageReward: 0.9725 -> 1.054 | Improved |
| replace forward-KL (DPO) with alpha-divergence objective | SD1.5 · Pick-a-Pic test (trained on Pick-a-Pic V2) | HPSv2: 26.11 -> 27.59 | Improved |
| use alpha-DPO instead of DPO under 20% label flipping | SDXL · Pick-a-Pic test (20% label flipping) | ImageReward: 0.9102 -> 1.001 | Improved |
| use alpha-DPO instead of DPO under 20% label flipping | SD1.5 · Pick-a-Pic test (20% label flipping) | HPSv2: 25.86 -> 26.69 | Improved |
| add dynamic alpha scheduling (vs fixed alpha) (ablation) | alpha-DPO SDXL · Pick-a-Pic test | ImageReward: 1.019 -> 1.054 | Improved |
| add dynamic alpha scheduling (vs fixed alpha) (ablation) | alpha-DPO SDXL · Pick-a-Pic test | PickScore: 22.45 -> 22.51 | Improved |
| raise mu from 0.1 to 0.9999 (ablation) | alpha-DPO SDXL · Pick-a-Pic test | ImageReward: 1.013 -> 1.054 | Improved |
| raise mu from 0.1 to 0.9999 (ablation) | alpha-DPO SDXL · Pick-a-Pic test | PickScore: 22.31 -> 22.51 | Improved |
| start dynamic alpha at step 0 rather than step 200 (ablation) | alpha-DPO SDXL · Pick-a-Pic test | ImageReward: 1.031 -> 1.054 | Improved |
| set beta to 4000 rather than 2000 (ablation) | alpha-DPO SDXL · Pick-a-Pic test | ImageReward: 1.022 -> 1.054 | Improved |
| set beta to 4000 rather than 6000 (ablation) | alpha-DPO SDXL · Pick-a-Pic test | ImageReward: 1.002 -> 1.054 | Improved |
| replace DPO component of SPO with alpha-DPO | SPO SDXL · Pick-a-Pic test | ImageReward: 1.116 -> 1.134 | Improved |
Closest papers by meaning
- Smoothed Preference Optimization via ReNoise Inversion for Aligning Diffusion Models with Varied Human Preferences (ICML 2025 · similarity 0.92)
- InPO: Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment (CVPR 2025 · similarity 0.90)
- Diffusion-NPO: Negative Preference Optimization for Better Preference Aligned Generation of Diffusion Models (ICLR 2025 · similarity 0.90)
- Diffusion Model Alignment Using Direct Preference Optimization (CVPR 2024 · similarity 0.89)
- Inference-Time Alignment of Diffusion Models with Direct Noise Optimization (ICML 2025 · similarity 0.89)
- Ranking-based Preference Optimization for Diffusion Models from Implicit User Feedback (NeurIPS 2025 · 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.
