Ikuto Hanada¹ (), Takehiro Nagai¹; ¹Institute of Science Tokyo

Understanding how color distributions influence aesthetic preference for paintings is an important issue in vision research. Previous studies have suggested that specific color statistics and the similarity of color distributions to natural-scene statistics may contribute to painting preference, yet their findings remain inconclusive. In this study, we aimed to examine the relationship between color preference for abstract paintings and the rarity of color distributions in real-world scenes quantified using a variational autoencoder (VAE) trained on naturalistic images. We first trained a VAE on a large naturalistic image dataset, such that color distributions with lower rarity were represented closer to the origin of the latent space. Accordingly, rarity was defined as the Euclidean distance from the origin in the latent space. Using this metric, we selected natural images spanning a wide range of rarity levels and created experimental stimuli by reflecting their color distributions onto several abstract base patterns. Preference was measured using Thurstone’s pairwise comparison method. As a result, although there were slight differences across base patterns, a strong positive correlation was consistently observed between the rarity of color distributions and preference: abstract paintings with rarer color distributions tended to be preferred. Analyses of color statistics showed that higher rarity was associated with greater variance in luminance, chroma, and hue, and that abstract paintings with larger variance in these dimensions were more likely to be preferred. These findings suggest that the rarity of color distributions may serve as an integrative concept that encompasses various color-related factors known to influence painting preferences. Overall, this study suggests the possibility that the rarity of color distributions underlies color preference for abstract paintings and helps unify diverse findings in the literature on painting aesthetics.

Acknowledgements: This study was supported by JST SPRING, Japan Grant Number JPMJSP2180.