Paper 2601.20354
Everything in Its Place: Benchmarking Spatial Intelligence of Text-to-Image Models
- Published
- Jan 2026
- Research lab
- Independent
- Citations
- 8
- GitHub
- 132 stars
01 In brief
Summary
The paper introduces SpatialGenEval, a benchmark for evaluating the spatial intelligence of text-to-image (T2I) models.
It uses 1,230 long, information-dense prompts across 25 real-world scenes, each integrating 10 spatial sub-domains (object, attribute, position, orientation, layout, comparison, proximity, occlusion, motion, causal) and paired with 10 multiple-choice questions.
Evaluation of 23 state-of-the-art models reveals that spatial reasoning (comparison, occlusion) is a primary bottleneck, with scores often near random (20%).
The benchmark also includes SpatialT2I, a dataset of 15,400 text-image pairs for fine-tuning, which yields consistent performance gains on Stable Diffusion-XL (+4.2%), UniWorld-V1 (+5.7%), and OmniGen2 (+4.4%).
Key findings include that open-source models are catching up to closed-source ones, text encoder capability is a key determinant, and model scale and architecture are pathways for improvement.
The benchmark's validity is supported by correlation with other benchmarks and human alignment studies.
02 From the paper
Abstract
Text-to-image (T2I) models have achieved remarkable success in generating high-fidelity images, but they often fail in handling complex spatial relationships, e.g., spatial perception, reasoning, or interaction. These critical aspects are largely overlooked by current benchmarks due to their short or information-sparse prompt design. In this paper, we introduce SpatialGenEval, a new benchmark designed to systematically evaluate the spatial intelligence of T2I models, covering two key aspects: (1) SpatialGenEval involves 1,230 long, information-dense prompts across 25 real-world scenes. Each prompt integrates 10 spatial sub-domains and corresponding 10 multi-choice question-answer pairs, ranging from object position and layout to occlusion and causality. Our extensive evaluation of 21 state-of-the-art models reveals that higher-order spatial reasoning remains a primary bottleneck. (2) To demonstrate that the utility of our information-dense design goes beyond simple evaluation, we also construct the SpatialT2I dataset. It contains 15,400 text-image pairs with rewritten prompts to ensure image consistency while preserving information density. Fine-tuned results on current foundation models (i.e., Stable Diffusion-XL, Uniworld-V1, OmniGen2) yield consistent performance gains (+4.2%, +5.7%, +4.4%) and more realistic effects in spatial relations, highlighting a data-centric paradigm to achieve spatial intelligence in T2I models.