Disentangling beauty and attractiveness in art portraits: Differential links to quantitative image properties

Abstract

Art portraits convey two related yet distinct hedonic values: beauty of the image and attractiveness of the person depicted. This study explored how these values relate to objective properties. Three hundred participants rated 192 art portraits (Baroque, Impressionism, Expressionism) on beauty and attractiveness. Quantitative image properties (QIPs) were analyzed using the QIP toolbox. Beauty and attractiveness ratings were highly correlated, yet regression analyses showed that beauty ratings were linked to style and distinct QIPs (mainly self-similarity and mirror symmetry), whereas attractiveness ratings were characterized by style and a diffuse pattern of weak QIP contributions. These findings suggest that while both ratings are perceptually distinct in art portraits, only beauty judgments are systematically related to self-similarity and symmetry.

How to cite this article

Hayn-Leichsenring, G. U. (2026). Disentangling beauty and attractiveness in art portraits: Differential links to quantitative image properties. i-Perception, 17(4), 1–6. https://doi.org/10.1177/20416695261458506

Human faces have been depicted artistically for thousands of years. In most cases, the facial features of the depicted individuals are well preserved but depicted in different styles (Graham et al., 2014). Therefore, art portraits are special aesthetic objects maintaining two hedonic values: the beauty of the image itself and the attractiveness of the person depicted. Both hedonic values are highly correlated but seem to be based on different perceptual mechanisms. Previously, in a study on gist and perceptual contrast it was shown that beauty ratings are more strongly influenced by cognitive effects than attractiveness ratings (Schulz & Hayn-Leichsenring, 2017).

To distinguish between beauty and attractiveness assessments based on their relationship to objective properties, an experiment was conducted (in accordance with ethical guidelines of the Declaration of Helsinki) using art portraits as stimuli. The database consists of 192 digital copies of art portraits (longest side 800 pixels) balanced for style (Baroque, Impressionism and Expressionism). Art portraits from various styles were used, as style may influence the viewer due to varying degrees of abstraction (see Figure 1 for examples). In total, 300 participants (18–71 years, M = 33.6 years, 101 female) took part in an online experiment. Before the actual rating experiment, the difference between attractiveness (a quality that causes an interest, (sexual) desire in, or gravitation to someone and is a property of the person.) and beauty (A quality related to the composition, coloring, painting technique, chosen perspective and other features that is a property of the image itself.) was explained to the participants using two example images. Then, each participant rated 70 randomly selected images on two 9-point Likert scales for attractiveness and beauty without time restrictions.

Figure 1.

Spearman ρ correlation between beauty and attractiveness ratings for portraits of different styles. ***p < .001.

Analysis showed high correlations between attractiveness ratings and beauty ratings (Spearman ρ = 0.658, p < .001) over all paintings and for single styles (see Figure 1).

A linear mixed-effects model was conducted to examine the effects of artistic style and rating type on hedonic ratings, including their interaction, with random intercepts for participants and stimuli. Results revealed significant main effects of style, F(2, 187) = 89.90, p < .001, and rating type, F(1, 42105) = 374.57, p < .001, qualified by a significant style × rating interaction, F(2, 42105) = 764.52, p < .001 (see Figure 2). A second model additionally included quantitative image properties (QIPs). The effects of style and the style × rating interaction remained highly significant, indicating that stylistic differences were not predicted by QIPs.

Figure 2.

Post hoc comparisons on the effect of style on hedonic ratings. Error bars represent 95% confidence intervals. ***p < .001.

To analyze the portrait paintings, the QIPs toolbox (Redies et al., 2025) was used. To reduce dimensionality among the 41 QIPs, separate LASSO-regularized linear regression models were fitted for beauty and attractiveness ratings. In both analyses, the regularization parameter λ was selected automatically by minimizing the mean squared error (MSE) on a validation dataset. The data were randomly split into training (n = 13,440), validation (n = 3,360), and test sets (n = 4,200). The beauty model used a smaller regularization parameter (λ = 5.28 × 10⁻⁵) than the attractiveness model (λ = 1.71 × 10⁻⁴), resulting in a less strongly regularized coefficient structure. Predictive performance was comparable across validation and test sets for both models (beauty: validation MSE = 2.966, test MSE = 3.008; attractiveness: validation MSE = 3.304, test MSE = 3.168), indicating stable generalization. The beauty model showed relatively stronger and more interpretable weights for PHOG-based self-similarity, mirror symmetry, and complexity- and color-related measures, whereas the attractiveness model showed no dominant QIPs. Based on these results, four QIPs showing comparatively stronger contributions in the beauty model were selected for subsequent inferential analyses:

Mirror Symmetry: Assessment of how similar an image is to its mirror reflection along major axes. Higher scores indicate greater symmetry.

HOG Complexity: Overall strength of luminance-based edges in an image. Strong, frequent intensity changes result in higher complexity.

PHOG-Based Self-similarity: Measure of how similar edge and gradient patterns are across different spatial scales of an image. Higher values indicate repeated structural patterns at multiple levels.

Color Entropy: Quantification of how evenly different color hues are distributed in an image. Images with many equally frequent hues have high entropy, while single-hue images have low entropy.

Portraits of different styles differ in terms of their average beauty and attractiveness ratings (e.g., Baroque portraits were rated as most beautiful) and in terms of QIPs (Figure 3).

Figure 3.

Examples for art portraits from Baroque (B), impressionism (I), and expressionism (E). The diagrams show mean values for z-transformed aesthetic ratings and quantitative image properties (QIPs). ***p < .001.

To assess how specific QIPs (as determined by the LASSO model) predict ratings, eight additional linear mixed-effects models were fitted separately for beauty and attractiveness ratings (four conditions: overall, Baroque, impressionism, expressionism), including z-transformed QIPs as fixed effects and random intercepts for participants and stimuli. The overall beauty model predicted substantially more variance (marginal R² = .205) than the overall attractiveness model (marginal R² = .077). In the beauty model, PHOG-based self-similarity showed a negative effect, F(1, 442.43) = 33.98, p < .001, b = −0.309, and mirror symmetry was positively associated with beauty, F(1, 184.02) = 7.66, p < .01, b = 0.212. In contrast, attractiveness ratings were not predicted by any specific QIPs (see Figure 4 for results). The results indicate that PHOG self-similarity and mirror symmetry were significantly associated with beauty evaluations, whereas they showed no reliable associations with attractiveness ratings.

Figure 4.

Standardized fixed-effect estimates (β) from linear mixed models predicting beauty and attractiveness ratings from quantitative image properties (QIPs). Positive values indicate positive associations with ratings. Error bars represent 95% confidence intervals. *p < .05, **p < .01, ***p < .001.

The effect is likely driven by Baroque portraits: Firstly, Baroque portraits are on average less self-similar and more symmetrical than portraits from Impressionism and Expressionism (lower self-similarity and higher symmetry were associated with higher beauty ratings). Secondly, only within Baroque portraits, PHOG self-similarity has a negative influence on beauty ratings (Figure 3).

In this study, two closely related hedonic values were examined, only one of which showed systematic relationships with self-similarity and mirror symmetry. Overall, beauty evaluations are not only more influenced by cognitive processes than attractiveness evaluations (Schulz & Hayn-Leichsenring, 2017), but are also more strongly associated with specific QIPs, which was not observed for attractiveness ratings.

Supplemental Material

sj-docx-1-ipe-10.1177_20416695261458506 - Supplemental material for Disentangling beauty and attractiveness in art portraits: Differential links to quantitative image properties

Supplemental material, sj-docx-1-ipe-10.1177_20416695261458506 for Disentangling beauty and attractiveness in art portraits: Differential links to quantitative image properties by Gregor U. Hayn-Leichsenring in i-Perception