Research map / Image generation / Text-to-image diffusion
Diffusion Negative Preference Optimization Made Simple
Joshua Tian Jin Tee, Hee Suk Yoon, Sunjae Yoon, Tri Ton, Chang D. Yoo
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 |
|---|---|---|---|
| use Diff-SNPO instead of Diff-DPO | SD1.5 · Pick-a-Pic v2 test | PickScore: 21.01 -> 22.24 | Improved |
| use single-model Diff-SNPO instead of dual-model Diff-NPO | SD1.5 · Pick-a-Pic v2 test | HPSv2: 26.92 -> 27.23 | Improved |
| use Diff-SNPO instead of Diff-NPO on SDXL | SDXL · Pick-a-Pic v2 test | Aesthetic Score: 5.9449 -> 5.8129 | Hurt |
| use Diff-SNPO instead of Diff-NPO on SDXL | SDXL · Pick-a-Pic v2 test | HPSv2: 28.30 -> 28.33 | No clear effect |
| add interpolation (lambda 0.9) to avoid naive single-model NPO (lambda 1.0) (ablation) | Diff-SNPO · Pick-a-Pic v2 (SD1.5) | HPSv2 and Aesthetic | Improved |
| vary lambda around 0.9 (ablation) | Diff-SNPO · Pick-a-Pic v2 (SD1.5) | HPSv2 and Aesthetic | No clear effect |
| vary beta from 1000 to 3000 (ablation) | Diff-SNPO · Pick-a-Pic v2 (SD1.5) | HPSv2: 27.30 -> 27.23 | No clear effect |
| avoid weight merging by using single model | Diff-SNPO vs Diff-NPO · Pick-a-Pic v2 test | negative implicit accuracy (%): 52.34 -> 57.45 | Improved |
| single-model design vs dual-model training | Diff-SNPO · SD1.5 training cost | relative training speed: 1.00x -> 2.00x | Improved |
| single-model design vs dual-model inference | Diff-SNPO · SD1.5 inference | throughput (img/s): 0.27 -> 0.48 | Improved |
| swap sampler from DDIM to UniPC (ablation) | Diff-SNPO · Pick-a-Pic v2 (SD1.5) | HPSv2: 27.23 -> 27.25 | No clear effect |
| use CHATS instead of Diff-SNPO on SD1.5 | CHATS · Pick-a-Pic v2 (SD1.5) | Aesthetic Score: 5.6258 -> 5.6845 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- CHATS — CHATS: Combining Human-Aligned Optimization and Test-Time Sampling for Text-to-Image Generation (ICML 2025) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Diffusion-NPO: Negative Preference Optimization for Better Preference Aligned Generation of Diffusion Models (ICLR 2025 · similarity 0.92)
- Ranking-based Preference Optimization for Diffusion Models from Implicit User Feedback (NeurIPS 2025 · similarity 0.88)
- Calibrated Multi-Preference Optimization for Aligning Diffusion Models (CVPR 2025 · similarity 0.87)
- InPO: Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment (CVPR 2025 · similarity 0.87)
- Smoothed Preference Optimization via ReNoise Inversion for Aligning Diffusion Models with Varied Human Preferences (ICML 2025 · similarity 0.86)
- TFG: Unified Training-Free Guidance for Diffusion Models (NeurIPS 2024 · similarity 0.86)
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.
