Research map / Image generation / Text-to-image diffusion
Towards Better Optimization For Listwise Preference in Diffusion Models
Jiamu Bai, Xin Yu, Meilong Xu, Weitao Lu, Xin Pan, Kiwan Maeng, Daniel Kifer, Jian Wang, Yu Wang
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 pairwise loss with listwise loss | Diffusion-DPO on SD1.5 · Pick-a-Pic | PickScore win rate: 68.2% -> 80.4% | Improved |
| replace pairwise loss with listwise loss | Diffusion-DPO on SDXL · Parti-Prompts | PickScore win rate: 66.8% -> 72.8% | Improved |
| replace pairwise loss with listwise loss | Diffusion-DPO on SDXL · HPSV2 | HPS win rate: 80.0% -> 85.0% | Improved |
| replace pairwise loss with listwise loss | Diffusion-DPO on SDXL · Pick-a-Pic | HPS win rate: 81.1% -> 87.1% | Improved |
| replace pairwise loss with listwise loss | Diffusion-DPO on SD1.5 · InstructPix2Pix editing | PickScore and HPS win rate: +9.1% PickScore, +4.4% HPS | Improved |
| use listwise loss on Pick-a-Pic v1 versus Diffusion-DPO trained on v2 (twice data) | Diffusion-LPO on SD1.5 · Pick-a-Pic | PickScore win rate: 73.3% -> 80.4% | Improved |
| replace pairwise loss with listwise loss | Diffusion-DPO on SD1.5 · Pick-a-Pic | HPS win rate: 70.7% -> 74.6% but DSPO 82.6% | Hurt |
| replace SFT with preference optimization | SFT on SDXL · Pick-a-Pic | PickScore win rate: 22.5% -> 77.1% | Improved |
| replace pairwise PPD loss with listwise loss (PPD-LPO) | PPD · held-in users | win rate: 71.1% -> 72.3% | Improved |
| replace pairwise PPD loss with listwise loss (PPD-LPO) | PPD · held-out users | win rate: 70.3% -> 80.2% | Improved |
| replace group pairwise DPO with listwise rank loss (ablation) | GP-DPO on SD1.5 · Pick-a-Pic | PickScore win rate: 52.3% win rate vs GP-DPO | Improved |
| replace group pairwise DPO with listwise rank loss (ablation) | GP-DPO on SD1.5 · HPSV2 | AES win rate: 49.1% | No clear effect |
Closest papers by meaning
- Ranking-based Preference Optimization for Diffusion Models from Implicit User Feedback (NeurIPS 2025 · similarity 0.93)
- Diffusion Model Alignment Using Direct Preference Optimization (CVPR 2024 · similarity 0.93)
- Diffusion Model as a Noise-Aware Latent Reward Model for Step-Level Preference Optimization (NeurIPS 2025 · similarity 0.92)
- Smoothed Preference Optimization via ReNoise Inversion for Aligning Diffusion Models with Varied Human Preferences (ICML 2025 · similarity 0.92)
- Aligning Diffusion Models by Optimizing Human Utility (NeurIPS 2024 · similarity 0.92)
- Personalized Preference Fine-tuning of Diffusion Models (CVPR 2025 · similarity 0.91)
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.
