Research map / Image generation / Fast diffusion sampling
SIPO: Stabilized and Improved Preference Optimization for Aligning Diffusion Models
Xiaomeng Yang, Mengping Yang, Junyan Wang, Zhijian Zhou, Zhiyu Tan, Hao Li
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 timestep-wise importance reweighting on top of clipping and masking (ablation) | DPO-C&M · VBench CogVideoX-2B @ 500 steps | Total: 81.37 -> 81.53 | Improved |
| add timestep-wise importance reweighting on top of clipping and masking (ablation) | DPO-C&M · VBench CogVideoX-2B @ 1000 steps | Total: 81.46 -> 81.68 | Improved |
| add timestep-wise importance reweighting on top of clipping and masking (ablation) | DPO-C&M · VBench WanX-1.3B | Total: 84.54 -> 84.78 | Improved |
| add clipping and masking of early timesteps to Diffusion-DPO (ablation) | Diffusion-DPO · VBench CogVideoX-2B @ 1000 steps | Total: 67.28 -> 81.46 | Improved |
| add clipping and masking of early timesteps to Diffusion-DPO (ablation) | Diffusion-DPO · VBench CogVideoX-5B | Total: 82.02 -> 82.17 | Improved |
| add SIPO to pretrained model for longer training (500 to 1000 steps) | SIPO · VBench CogVideoX-2B | Total: 81.53 -> 81.68 | Improved |
| replace Diffusion-DPO with SIPO | Diffusion-DPO · Pick-a-Pic | PickScore: 20.97 -> 21.68 | Improved |
| replace Diffusion-DPO with SIPO | Diffusion-DPO · Pick-a-Pic | HPSv2: 0.2656 -> 0.2783 | Improved |
| replace Diffusion-DPO with SIPO | Diffusion-DPO · Pick-a-Pic | ImageReward: 0.2989 -> 0.3051 | Improved |
| add timestep-wise importance reweighting on top of clipping and masking (ablation) | DPO-C&M · Pick-a-Pic | PickScore: 21.57 -> 21.68 | Improved |
| replace SIPO with Diffusion-KTO | SIPO · Pick-a-Pic | ImageReward: 0.3051 -> 0.6156 | Hurt |
| replace Diffusion-DPO with SIPO on FLUX.1-dev | Diffusion-DPO · GenEval | Overall: 0.82 -> 0.85 | Improved |
Closest papers by meaning
- Diffusion-NPO: Negative Preference Optimization for Better Preference Aligned Generation of Diffusion Models (ICLR 2025 · similarity 0.93)
- Calibrated Multi-Preference Optimization for Aligning Diffusion Models (CVPR 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 0.91)
- InPO: Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment (CVPR 2025 · similarity 0.91)
- Diffusion Model Alignment Using Direct Preference Optimization (CVPR 2024 · 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.
