Research map / Image generation / Text-to-image diffusion
Benchmarking and Improving Fine-Grained Text-to-Image Alignment via Paired Reinforcement Learning
Kaihang Pan, Wendong Bu, Yuruo Wu, Kai Shen, Yang Wu, Yun Zhu, Zehan Wang, Yunfei Li, Hang Zhao, Juncheng Li, Siliang Tang
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 |
|---|---|---|---|
| train with FocusDiff (paired data + Pair-GRPO) on X-Omni | X-Omni · DeltaBench average | sa: 61.8 -> 76.1 | Improved |
| train with FocusDiff on X-Omni | X-Omni · DeltaBench average | sa-sg stability gap: 8.6 -> 4.3 points | Improved |
| expand group to include images from both paired prompts (ablation) | Pair-GRPO · DeltaBench average | sa / sg: 73.5 -> 76.1 sa; 67.9 -> 71.8 sg | Improved |
| provide ground-truth images with annealed probability during RL (ablation) | Pair-GRPO · DeltaBench average | sa / sg: 75.1 -> 76.1 sa; 70.0 -> 71.8 sg | Improved |
| use Pair-GRPO instead of vanilla GRPO (ablation) | GRPO · DeltaBench average | sa / sg: 71.9 -> 76.1 sa; 66.0 -> 71.8 sg | Improved |
| train on FocusDiff-Data instead of commonly-used prompts (ablation) | vanilla GRPO · DeltaBench average | sa / sg: 63.0 -> 71.9 sa; 55.4 -> 66.0 sg | Improved |
| add FocusDiff training to X-Omni | X-Omni · GenEval | overall: 0.83 -> 0.90 | Improved |
| add FocusDiff training to X-Omni | X-Omni · DPG-Bench | average: 87.65 -> 88.35 | Improved |
| use FocusDiff-trained model under extended sampling (2 to 8 images per prompt) | X-Omni-FocusDiff · DeltaBench | geometric mean: 71.8 -> 69.8 | No clear effect |
| use Pair-GRPO instead of vanilla GRPO for training stability (ablation) | GRPO · RL training reward | reward trend | Improved |
| FocusDiff-trained model versus Qwen-Image on GenEval | Qwen-Image · GenEval | overall: 0.91 vs 0.90 | Hurt |
Closest papers by meaning
- Aligning Text to Image in Diffusion Models is Easier Than You Think (NeurIPS 2025 · similarity 0.90)
- Improving Long-Text Alignment for Text-to-Image Diffusion Models (ICLR 2025 · similarity 0.89)
- Information Theoretic Text-to-Image Alignment (ICLR 2025 · similarity 0.89)
- Text Embedding is Not All You Need: Attention Control for Text-to-Image Semantic Alignment with Text Self-Attention Maps (CVPR 2025 · similarity 0.88)
- Discriminative Probing and Tuning for Text-to-Image Generation (CVPR 2024 · similarity 0.88)
- Multimodal Large Language Models Make Text-to-Image Generative Models Align Better (NeurIPS 2024 · 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.
