Research map / Image generation / Text-to-image diffusion
Direct Diffusion Score Preference Optimization via Stepwise Contrastive Policy-Pair Supervision
Dohyun Kim, Seungwoo Lyu, Seung Kim, Paul Hongsuck Seo
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-posterior targets with contrastive-policy-pair targets in DSPO | DSPO · T2I alignment (SD-1.4) | GenEval: .4841 -> .4978 | Improved |
| replace forward-process targets with DDSPO stepwise targets vs Diffusion DPO | Diffusion DPO · T2I alignment (SD-1.4) | GenEval: .4841 -> .5045 | Improved |
| DDSPO versus D-DPO on image distribution shift | Diffusion DPO · MS-COCO 30K (SD-1.4) | FID: 18.02 -> 16.39 | Improved |
| DSPO+CPP vs DDSPO on CompBench | DDSPO · T2I alignment (SD-1.4) | T2I-CompBench: .3823 vs .3854 | Mixed |
| apply DDSPO to SDXL | SDXL · T2I alignment | GenEval: .5229 -> .6049 | Improved |
| apply DDSPO to SD3-M | SD3-M · T2I alignment | T2I-CompBench: .5155 -> .5740 | Improved |
| DDSPO without human annotation vs CaPO | SD3-M · T2I alignment | GenEval: .7100 -> .7384 | Improved |
| train DDSPO with Pick-a-Pic (DD-CPP) vs D-DPO | SDXL · Aesthetic quality | HPSv2: 28.55 -> 29.09 | Improved |
| use SPO instead of DDSPO (SDXL) | SDXL · Aesthetic quality | PickScore: 22.71 -> 23.11 | Hurt |
| use random token removal instead of LLaMA (AQ) degraded prompts (ablation) | DDSPO · T2I alignment (SD-1.4) | GenEval: .4758 -> .5045 | Improved |
| use LLaMA (AQ) degraded prompts instead of random removal (ablation) | DDSPO · Aesthetic quality (SD-1.4) | HPSv2: 27.28 -> 27.51 | Improved |
| generate explicit negative images instead of reusing random positive image (ablation) | DDSPO · T2I alignment (SD-1.4) | GenEval: .4866 -> .5045 | Improved |
Closest papers by meaning
- Ranking-based Preference Optimization for Diffusion Models from Implicit User Feedback (NeurIPS 2025 · similarity 0.92)
- Aesthetic Post-Training Diffusion Models from Generic Preferences with Step-by-step Preference Optimization (CVPR 2025 · similarity 0.90)
- Personalized Preference Fine-tuning of Diffusion Models (CVPR 2025 · similarity 0.90)
- Diffusion Model as a Noise-Aware Latent Reward Model for Step-Level Preference Optimization (NeurIPS 2025 · similarity 0.90)
- Training Diffusion Models with Reinforcement Learning (ICLR 2024 · similarity 0.89)
- Diffusion Model Alignment Using Direct Preference Optimization (CVPR 2024 · 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.
