Research map / Image generation / Text-to-image diffusion
Getting it Right: Improving Spatial Consistency in Text-to-Image Models
Agneet Chatterjee, Gabriela Ben Melech Stan, Estelle Guez Aflalo, Sayak Paul, Dhruba Ghosh, Tejas Gokhale, Ludwig Schmidt, Hanna Hajishirzi, Vasudev Lal, Chitta R Baral, Yezhou Yang
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 |
|---|---|---|---|
| fine-tune SD 2.1 on spatially focused SPRIGHT captions (<15k images) | Stable Diffusion 2.1 · T2I-CompBench spatial | spatial score: 0.1507 -> 0.1840 | Improved |
| fine-tune SD 2.1 on SPRIGHT captions (<15k images) | Stable Diffusion 2.1 · COCO-30K | ZS-FID: 21.646 -> 14.925 | Improved |
| train only on images with more than 18 objects (444 images) (ablation) | Stable Diffusion 2.1 + SPRIGHT · T2I-CompBench spatial | spatial score: 0.1309 -> 0.2133 | Improved |
| use 50% spatial captions instead of 100% spatial captions in fine-tuning mix (ablation) | Stable Diffusion 2.1 · T2I-CompBench spatial | spatial score: 0.140 -> 0.178 | Improved |
| use only 25% spatial captions instead of 50% (ablation) | Stable Diffusion 2.1 · T2I-CompBench spatial | spatial score: 0.178 -> 0.154 | Hurt |
| use long spatial captions instead of short single-sentence captions (ablation) | Stable Diffusion 2.1 with CLIP fine-tuning · T2I-CompBench spatial | spatial score: 0.1230 -> 0.1777 | Improved |
| fine-tune CLIP text encoder together with U-Net (ablation) | Stable Diffusion 2.1 with long captions · T2I-CompBench spatial | spatial score: 0.1605 -> 0.1777 | Improved |
| fine-tune CLIP text encoder with short spatial captions (ablation) | Stable Diffusion 2.1 · T2I-CompBench spatial | spatial score: 0.1420 -> 0.1230 | Hurt |
| replace spatial phrases with negations during training (ablation) | Stable Diffusion 2.1 + SPRIGHT · T2I-CompBench spatial (negation prompts) | spatial score: 0.066 -> 0.069 (negation only); 0.1376 -> 0.1427 (mixed) | No clear effect |
| fine-tune on SPRIGHT, effect on GenEval attribute binding | Stable Diffusion 2.1 · GenEval | attribute binding: 0.17 -> 0.15 | Hurt |
| fine-tune on SPRIGHT, effect on GenEval counting and position | Stable Diffusion 2.1 · GenEval | counting / position: 0.44 -> 0.49 (counting); 0.07 -> 0.11 (position) | Improved |
Closest papers by meaning
- Stitch and Tell: A Structured Data Augmentation Method for Spatial Understanding (NeurIPS 2025 · similarity 0.89)
- Is A Picture Worth A Thousand Words? Delving Into Spatial Reasoning for Vision Language Models (NeurIPS 2024 · similarity 0.89)
- MMSI-Bench: A Benchmark for Multi-Image Spatial Intelligence (ICLR 2026 · similarity 0.89)
- TextRegion: Text-Aligned Region Tokens from Frozen Image-Text Models (NeurIPS 2026 · similarity 0.89)
- Why Is Spatial Reasoning Hard for VLMs? An Attention Mechanism Perspective on Focus Areas (ICML 2025 · similarity 0.88)
- Spatial Transport Optimization by Repositioning Attention Map for Training-Free Text-to-Image Synthesis (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.
