Research map / Image generation / Text-to-image diffusion
T2I-CompBench: A Comprehensive Benchmark for Open-world Compositional Text-to-image Generation
Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, Xihui Liu
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 |
|---|---|---|---|
| finetune Stable Diffusion v2 with selected high-reward samples (GORS) | Stable Diffusion v2 · Attribute-color | B-VQA: 0.5065 -> 0.6603 | Improved |
| GORS finetuning | Stable Diffusion v2 · Spatial relationship | UniDet: 0.1342 -> 0.1815 | Improved |
| GORS finetuning | Stable Diffusion v2 · Complex compositions | 3-in-1: 0.3386 -> 0.3328 | Mixed |
| GORS finetuning | Stable Diffusion v2 · Non-spatial relationship | CLIP: 0.3127 -> 0.3193 | Improved |
| finetune only U-Net instead of both CLIP and U-Net (ablation) | GORS · Attribute-color | B-VQA: 0.6570 -> 0.6216 | Hurt |
| finetune only CLIP text encoder instead of both (ablation) | GORS · Attribute-color | B-VQA: 0.6570 -> 0.5507 | Hurt |
| use half the selection threshold (ablation) | GORS · Attribute-color | B-VQA: 0.6570 -> 0.6157 | Hurt |
| use no selection threshold (all generated images) (ablation) | GORS · Attribute-color | B-VQA: 0.6570 -> 0.6130 | Hurt |
| use alternative reward models not used in evaluation (GORS-unbiased) (ablation) | GORS · Attribute-color | B-VQA: 0.6603 -> 0.6414 | No clear effect |
| add Chain-of-Thought to MiniGPT4 evaluation metric (ablation) | MiniGPT4 metric · Attribute-color human correlation | Kendall tau: 0.1197 -> 0.3156 | Improved |
| use Disentangled BLIP-VQA instead of CLIP as metric | evaluation metric · Attribute-color human correlation | Kendall tau: 0.1938 -> 0.6297 | Improved |
| Structured Diffusion on SD v2 | Stable Diffusion v2 · Attribute-color | B-VQA: 0.5065 -> 0.4990 | No clear effect |
Closest papers by meaning
- OneIG-Bench: Omni-dimensional Nuanced Evaluation for Image Generation (NeurIPS 2025 · similarity 0.91)
- ConceptMix: A Compositional Image Generation Benchmark with Controllable Difficulty (NeurIPS 2024 · similarity 0.90)
- T2V-CompBench: A Comprehensive Benchmark for Compositional Text-to-video Generation (CVPR 2025 · similarity 0.90)
- LLM Blueprint: Enabling Text-to-Image Generation with Complex and Detailed Prompts (ICLR 2024 · similarity 0.90)
- MMIG-Bench: Towards Comprehensive and Explainable Evaluation of Multi-Modal Image Generation Models (NeurIPS 2025 · similarity 0.89)
- Mastering Text-to-Image Diffusion: Recaptioning, Planning, and Generating with Multimodal LLMs (ICML 2024 · 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.
