Research map / Image generation / Text-to-image diffusion
Scalable Ranked Preference Optimization for Text-to-Image Generation
Shyamgopal Karthik, Huseyin Coskun, Zeynep Akata, Sergey Tulyakov, Jian Ren, Anil Kag
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 |
|---|---|---|---|
| label preferences with reward models instead of random labels (ablation) | DPO on SDXL · DPG-Bench | DSG score: 75.66 -> 78.84 | Improved |
| aggregate 5 reward models instead of single HPSv2.1 (ablation) | DPO on SDXL · DPG-Bench | DSG score: 78.04 -> 78.84 | Improved |
| use ranked preferences with DCG-weighted RankDPO loss instead of pairwise DPO (ablation) | SDXL · DPG-Bench | DSG score: 78.84 -> 79.26 | Improved |
| add gain weights to DPO (ablation) | DPO on SDXL · DPG-Bench | DSG score: 78.84 -> 79.15 | Improved |
| generate Syn-Pic images from multiple T2I models instead of SDXL only (ablation) | RankDPO · DPG-Bench | Q-Align score: 0.74 -> 0.81 | Improved |
| synthetic Syn-Pic preferences instead of Pick-a-Picv2 human preferences for DPO (ablation) | DPO on SDXL · DPG-Bench | DSG score: 76.74 -> 78.84 | Improved |
| supervised fine-tuning on winning images instead of preference optimization (ablation) | SDXL · DPG-Bench | DSG score: 79.26 -> 76.56 vs RankDPO | Hurt |
| apply RankDPO to SDXL | SDXL · GenEval mean | mean score: 0.55 -> 0.61 | Improved |
| apply RankDPO to SD3-Medium (already DPO-tuned on 3M human preferences) | SD3-Medium · GenEval mean | mean score: 0.70 -> 0.74 | Improved |
| apply RankDPO to SDXL on position category | SDXL · GenEval position | score: 0.15 -> 0.14 | No clear effect |
| apply RankDPO to SD3-Medium visual quality | SD3-Medium · DPG-Bench | Q-Align score: 0.67 -> 0.68 | No clear effect |
| apply RankDPO to SDXL non-spatial relations | SDXL · T2I-CompBench Non-Spatial | score: 31.19 -> 31.33 | No clear effect |
Closest papers by meaning
- Smoothed Preference Optimization via ReNoise Inversion for Aligning Diffusion Models with Varied Human Preferences (ICML 2025 · similarity 0.91)
- Multimodal Large Language Models Make Text-to-Image Generative Models Align Better (NeurIPS 2024 · similarity 0.90)
- MIA-DPO: Multi-Image Augmented Direct Preference Optimization For Large Vision-Language Models (ICLR 2025 · similarity 0.90)
- Curriculum Direct Preference Optimization for Diffusion and Consistency Models (CVPR 2025 · similarity 0.90)
- DrPO: Drifting Preference Optimization for One-Step Generative Models (NeurIPS 2026 · similarity 0.90)
- Diffusion Model Alignment Using Direct Preference Optimization (CVPR 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.
