Anna Kravchenko¹, Andrey Bagrov¹, Veronica Dudarev²; ¹Radboud University, ²University of British Columbia

Image complexity affects a variety of aspects of processing, including recognition and search efficiency, attention distribution and the likelihood that scenes and objects will be remembered (Sun & Firestone, 2021; Hensellek, 2025; Kyle-Davidson et al., 2025; Endemann & Kamp, 2024). Yet, quantifying image complexity remains a persistent problem. Multiple measures of complexity have been proposed in the literature, from entropy estimate to more complex measures of the number of regions or visual congestion, however, no single metric is generally accepted as the standard, their performance varying widely between images of different types, e.g. artwork vs. photographs (Saraee et al., 2018; Saraee et al., 2020). Here we employ Multi-Scale Structural Complexity (MSSC), a measure that is designed to quantify the presence of structure at distinct spatial scales within images (Bagrov et al., 2020). It has shown reliable correlations with human ranking of image complexity and robust performance across categories (Kravchenko et al., 2025). To test behavioral relevance, we used MSSC to predict scanning patterns (number and duration of fixations) in 101 participants who were viewing naturalistic scenes (data published in Dudarev et al., 2025). Each scene was displayed for 5 seconds alongside image of a simple object. The scenes varied in emotional valence (negative, neutral) and presence of people. We demonstrate that MSSC has higher correlation with the number of fixations than entropy and visual congestion. MSSC also showed an indirect relationship with duration of fixations, in that the two correlated in images without people. When people were present in the picture, images were processed with shorter fixations overall (this effect is reported and discussed in Dudarev et al., 2025), despite no difference in MSSC. We suggest that this discrepancy reflects a distinction between perceptual and conceptual complexity and argue that MSSC provides a framework for disentangling the two.