Mathias Salvas-Hébert¹, Nicolas Dupuis-Roy², Catherine Landry¹, Ian Charest¹, Frédéric Gosselin¹; ¹Cerebrum, Université de Montréal, ²Elephant Scientific Consulting

The SHINE toolbox (Willenbockel et al., 2010; >1,350 citations) has played a central role in vision science by enabling researchers to control accidental low-level image properties that might otherwise confound behavioural or neural responses. SHINE provides tools to match or specify Fourier amplitude spectra, enforce rotational averages in Fourier space, specify luminance histograms, and specify a mean and a variance for luminance distributions. These controls remain essential not only for psychophysics but also for modern computer vision mostly developed in Python, where the training data of deep neural networks often contain unintended statistical regularities. We introduce SHINIER, a Python-based re-implementation of SHINE that preserves the full feature set of the original toolbox while extending its capabilities. SHINIER adds support for colour images, improved memory efficiency, multiple histogram-specification algorithms, dithering operations, high-precision numerical routines, and enhanced visualisation tools. It is available on GitHub and PyPI (pip install shinier) and provides legacy-compatible functions to streamline migration from MATLAB. To illustrate its utility for computer vision, we trained two convolutional neural networks (CNNs) on a cat–dog classification task while manipulating luminance histogram statistics. Images followed one of three luminance profiles: negatively skewed (A), positively skewed (B), or their average (C). One CNN was trained on images normalized to C, eliminating luminance-histogram cues; the other was trained with cats mapped to A and dogs to B, introducing a spurious correlation. At test, luminance histogram was decorrelated from category by applying a 50/50 mixture of A and B. The histogram-normalised model generalized well (77% accuracy), whereas the histogram-contaminated model collapsed to chance (50% accuracy), demonstrating how uncontrolled low-level statistics can drive model behaviour. SHINIER provides a robust, accessible solution for eliminating such confounds in both human and machine vision research.