SensIMG-Bench is a literature-grounded benchmark for evaluating whether AI-generated images are visually appropriate for sensory-sensitive viewers, with a focus on children with autism spectrum disorder (ASD).
The benchmark computes deterministic visual features for saturation, entropy, spatial information, pattern density, spatial frequency, and brightness, then aggregates them into a composite sensory-load score. The repository also includes a public project website and curated result files for reference.
The full generated-image dataset is coming soon and will be hosted externally.
Dataset available on Hugging Face (link pending).
The repository only includes lightweight prompt metadata, selected results, and website assets. Large image datasets are intentionally not stored in git.
The optional synthesis layer uses Qwen2.5-VL to produce natural-language interpretations of benchmark scores and image-level sensory suitability. Running this step requires the Qwen/transformers stack and is best done on a CUDA GPU with sufficient VRAM.
By default, the synthesis scripts use Qwen/Qwen2.5-VL-7B-Instruct. Set SENSVIS_QWEN_MODEL=/path/to/local/model to use a local checkpoint.
The results/ folder contains selected outputs for demonstration and reproducibility:
- per-model benchmark results
- grounded and ablation synthesis outputs
- synthesis-method comparison summary
These files are curated outputs, not a dump of all intermediate experiment artifacts.
@inproceedings{adeyemi2026sensimg,
title = {SensIMG-Bench: A Literature-Grounded Benchmark for Sensory-Appropriate AI-Generated Images},
author = {Adeyemi, Morayo Danielle and Liang, Paul Pu},
booktitle = {Proceedings of the 2026 ACM SIGACCESS Conference on Computers and
Accessibility (ASSETS)},
year = {2026},
note = {Under review.}
}MIT. See LICENSE.