Research map / Image generation / Text-to-image diffusion
SAIL: Self-Amplified Iterative Learning for Diffusion Model Alignment with Minimal Human Feedback
Xiaoxuan He, Siming Fu, Wanli Li, Zhiyuan Li, Dacheng Yin, Kang Rong, Fengyun Rao, Bo Zhang
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 |
|---|---|---|---|
| mix generated preference data with human seed data (mixup) (ablation) | SAIL · Pick-a-Pic v2 (SD1.5, Iter2) | ImageReward: 0.1564 -> 0.1729 | Improved |
| mix generated preference data with human seed data (mixup) (ablation) | SAIL · Pick-a-Pic v2 (SD1.5, Iter2) | HPSv2: 26.55 -> 26.65 | Improved |
| select best-worst pair instead of random pair (ablation) | SAIL · Pick-a-Pic v2 (SD1.5, Iter1) | ImageReward: 0.1055 -> 0.1137 | Improved |
| self-reward ranking instead of Aesthetics reward model ranking (Online DPO) (ablation) | SAIL · Pick-a-Pic v2 (SD1.5, Iter1) | ImageReward: 0.0936 -> 0.1137 | Improved |
| self-reward ranking instead of Aesthetics reward model ranking (Online DPO) (ablation) | SAIL · Pick-a-Pic v2 (SD1.5, Iter1) | Aesthetics: 5.53 -> 5.46 | Hurt |
| add more self-improvement iterations (ablation) | SAIL · Pick-a-Pic v2 (SD1.5) | ImageReward: 0.0715 -> 0.2329 | Improved |
| add a further iteration (Iter4 vs Iter3) (ablation) | SAIL · PartiPrompts (SD1.5) | ImageReward: 0.3198 -> 0.3072 | No clear effect |
| use 6% of human preference data with iterative self-rewarding instead of full DiffusionDPO data | SDXL · Pick-a-Pic v2 validation | HPSv2: 27.27 -> 27.32 | Improved |
| use 6% of human preference data instead of full DiffusionSPO data | SD1.5 · Pick-a-Pic v2 validation | PickScore: 21.20 -> 21.00 | No clear effect |
| initialize SAIL with the full Pick-a-Pic v2 data (SAIL*) | SD1.5 · Pick-a-Pic v2 validation | ImageReward: 0.1577 -> 0.4303 vs DiffusionSPO | Improved |
Closest papers by meaning
- Diffusion Model Alignment Using Direct Preference Optimization (CVPR 2024 · similarity 0.90)
- Calibrated Multi-Preference Optimization for Aligning Diffusion Models (CVPR 2025 · similarity 0.89)
- Aligning Diffusion Models by Optimizing Human Utility (NeurIPS 2024 · similarity 0.89)
- InPO: Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment (CVPR 2025 · similarity 0.89)
- AlignDiff: Aligning Diverse Human Preferences via Behavior-Customisable Diffusion Model (ICLR 2024 · similarity 0.89)
- Smoothed Preference Optimization via ReNoise Inversion for Aligning Diffusion Models with Varied Human Preferences (ICML 2025 · similarity 0.88)
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.
