Research map / Image generation / Text-to-image diffusion
OSPO: Object-Centric Self-Improving Preference Optimization for Text-to-Image Generation
Yoonjin Oh, Yongjin Kim, Hyomin Kim, Donghwan Chi, Sungwoong Kim
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 prompt densification before image generation (ablation) | OSPO on Janus-Pro-7B · T2I-CompBench++ Attribute | score: 0.641 -> 0.756 | Improved |
| add pair filtering and selection (with densification) (ablation) | OSPO on Janus-Pro-7B · T2I-CompBench++ Attribute | score: 0.716 -> 0.756 | Improved |
| add pair filtering alone (without densification) (ablation) | OSPO on Janus-Pro-7B · T2I-CompBench++ Attribute | score: 0.618 -> 0.549 | Hurt |
| add pair filtering and selection (without densification) (ablation) | OSPO on Janus-Pro-7B · T2I-CompBench++ Attribute | score: 0.618 -> 0.641 | Improved |
| add filtering and selection (with densification) (ablation) | OSPO on Janus-Pro-7B · GenEval Overall | score: 0.813 -> 0.831 | Improved |
| add SFT loss to preference loss (ablation) | OSPO on Janus-Pro-7B · GenEval Overall | score: 0.794 -> 0.831 | Improved |
| add SFT loss to preference loss (ablation) | OSPO on Janus-Pro-7B · T2I-CompBench++ Attribute | score: 0.776 -> 0.756 | Hurt |
| use object-weighted SimPO instead of standard SimPO (with SFT) (ablation) | OSPO on Janus-Pro-7B · GenEval Position | score: 0.795 -> 0.828 | Improved |
| use object-weighted SimPO instead of standard SimPO (with SFT) (ablation) | OSPO on Janus-Pro-7B · T2I-CompBench++ Layout | score: 0.437 -> 0.447 | Improved |
| apply OSPO to Janus-Pro-7B | Janus-Pro-7B · T2I-CompBench++ Color | score: 0.5215 -> 0.8567 | Improved |
| apply OSPO vs SUDER on Counting | Janus-Pro-7B · GenEval Counting | score: 0.70 -> 0.63 | Hurt |
| increase training sample size and candidate pair count (ablation) | OSPO · T2I-CompBench++ attribute and GenEval | score | Improved |
Closest papers by meaning
- SILMM: Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation (CVPR 2025 · similarity 0.89)
- Multimodal Large Language Models Make Text-to-Image Generative Models Align Better (NeurIPS 2024 · similarity 0.89)
- MIA-DPO: Multi-Image Augmented Direct Preference Optimization For Large Vision-Language Models (ICLR 2025 · similarity 0.88)
- Improving Long-Text Alignment for Text-to-Image Diffusion Models (ICLR 2025 · similarity 0.88)
- Generalizing Alignment Paradigm of Text-to-Image Generation with Preferences Through f-Divergence Minimization (AAAI 2025 · similarity 0.88)
- VideoDPO: Omni-Preference Alignment for Video Diffusion Generation (CVPR 2025 · similarity 0.88)
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.
