Research map / Image generation / Text-to-image diffusion
Stable Preference: Redefining training paradigm of human preference model for Text-to-Image Synthesis
Hanting Li, Hongjing Niu, Feng Zhao
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 anti-interference loss (ablation) | CLIP-H on ImageReward · ImageReward test | pair selection accuracy: 63.9 -> 64.5 | Improved |
| use stable preference two-step paradigm instead of cross-entropy (ablation) | CLIP-H on ImageReward · ImageReward test | pair selection accuracy: 63.9 -> 66.2 | Improved |
| combine anti-interference loss with stable preference (ablation) | CLIP-H · ImageReward test | pair selection accuracy: 66.2 -> 66.8 | Improved |
| use stable preference on CLIP-L (ablation) | CLIP-L · ImageReward test | pair selection accuracy: 63.1 -> 66.3 | Improved |
| weight step-2 loss toward larger score-gap pairs instead of uniform (ablation) | Stable Preference · HPD v2 | pair selection accuracy: 80.1 -> 80.7 | Improved |
| weight step-2 loss toward similar pairs instead of large-gap pairs (ablation) | Stable Preference · HPD v2 | pair selection accuracy: 80.7 -> 79.5 | Hurt |
| train CLIP-H with stable preference vs cross-entropy, evaluated cross-domain | CLIP-H · HPD v2 test | pair selection accuracy: 74.8 -> 80.7 | Improved |
| initialize CLIP-H from HPS v2 checkpoint | Stable Preference · ImageReward test | pair selection accuracy: 66.8 -> 68.0 | Improved |
| apply horizontal flip and center crop at test time | SP (CLIP-H) vs HPS v2 · HPD v2 | pair selection accuracy: HPS v2 83.3 -> 81.7 vs SP 80.7 -> 80.7 | 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.
- ImageReward — ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation (NeurIPS 2023) · better on 1 of 1 shared result
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
- Multimodal Large Language Models Make Text-to-Image Generative Models Align Better (NeurIPS 2024 · similarity 0.92)
- Learning Multi-Dimensional Human Preference for Text-to-Image Generation (CVPR 2024 · similarity 0.91)
- Diffusion Model Alignment Using Direct Preference Optimization (CVPR 2024 · similarity 0.89)
- Diffusion Model as a Noise-Aware Latent Reward Model for Step-Level Preference Optimization (NeurIPS 2025 · similarity 0.89)
- Aligning Visual Contrastive learning models via Preference Optimization (ICLR 2025 · similarity 0.89)
- RoVRM: A Robust Visual Reward Model Optimized via Auxiliary Textual Preference Data (AAAI 2025 · 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.
