Research map / Image generation / Text-to-image diffusion
Self-NPO: Data-Free Diffusion Model Enhancement via Truncated Diffusion Fine-Tuning
Fu-Yun Wang, Keqiang Sun, Yao Teng, Xihui Liu, Jiale Yuan, Jiaming Song, Hongsheng 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 Self-NPO negative branch for CFG | SDXL · Pick-a-Pic test_unique | LAION-Aesthetic: 6.114 -> 6.226 | Improved |
| add Self-NPO negative branch for CFG | SDXL · Pick-a-Pic test_unique | PickScore: 22.06 -> 22.26 | Improved |
| add Self-NPO | SDXL Diffusion-DPO · Pick-a-Pic test_unique | HPSv2.1: 29.76 -> 29.83 | Improved |
| add Self-NPO | Juggernaut · Pick-a-Pic test_unique | ImageReward: 0.9778 -> 0.9921 | Improved |
| add Self-NPO | SD-1.5 · Pick-a-Pic test_unique | ImageReward: 0.1064 -> 0.2816 | Improved |
| add Self-NPO | SePPO · Pick-a-Pic test_unique | HPSv2.1: 28.45 -> 30.28 | Improved |
| use Self-NPO instead of Diffusion-NPO (data-free) | SD-1.5 · Pick-a-Pic test_unique | HPSv2.1: 27.36 -> 27.04 | Hurt |
| lower negative-branch weight alpha from 1.0 to 0.8 (ablation) | DreamShaper + Self-NPO · Pick-a-Pic test_unique | LAION-Aesthetic: 6.196 -> 6.562 (ImageReward 0.8038 -> 0.6548) | Mixed |
| truncate simulation to 5 steps instead of full 25-step simulation (ablation) | Self-NPO training · SD1.5 training cost | A800 GPU hours: 10.4 -> 2 | Improved |
Closest papers by meaning
- Diffusion-NPO: Negative Preference Optimization for Better Preference Aligned Generation of Diffusion Models (ICLR 2025 · similarity 0.96)
- Ranking-based Preference Optimization for Diffusion Models from Implicit User Feedback (NeurIPS 2025 · similarity 0.92)
- Diffusion Model as a Noise-Aware Latent Reward Model for Step-Level Preference Optimization (NeurIPS 2025 · similarity 0.91)
- Personalized Preference Fine-tuning of Diffusion Models (CVPR 2025 · similarity 0.91)
- Calibrated Multi-Preference Optimization for Aligning Diffusion Models (CVPR 2025 · similarity 0.90)
- Self-Play Fine-tuning of Diffusion Models for Text-to-image Generation (NeurIPS 2024 · similarity 0.90)
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.
