Demographic Differences in Automated Facial Attractiveness Prediction: Evidence from SCUT-FBP5500 and MEBeauty

Abstract

In this work, we examine automated facial beauty prediction (FBP) models to evaluate demographic variation in human attractiveness scores, facial landmark preprocessing, and cross-dataset prediction error. We perform multistage quantitative sociotechnical audits using SCUT-FBP5500 and a 1,962-image MEBeauty subset, including DLib landmark-extraction analysis, intersectional ordinary least squares (OLS) regression, observed-rating inter-rater variability analysis, and a five-seed ResNet-50 cross-dataset evaluation. DLib landmark extraction was successful for 1,929 of 1,962 MEBeauty images, with an overall failure rate of 1.7%. The highest observed subgroup failure rate was among Middle Eastern men (4.2%; 2/48). In the symmetry-adjusted OLS model (N = 1,929, R² = .022), no gender, ethnicity, or gender × ethnicity coefficient was statistically significant at the Benjamini–Hochberg correction. At the same time, blur was the sole significant image quality control (p=.006). Subgroup RMSE estimates across datasets ranged from .470 to .512. Black men had the highest point estimate (RMSE = .512), but prespecified image-level bootstrap comparisons with Asian men (difference = .002, 95% CI [−.09, .09]) and White men (difference = .022, 95% CI [−.05, .09]) both included zero. These results show how disaggregated sociotechnical auditing can be useful and also emphasize the significance of disentangling descriptive subgroup variation from statistically supported demographic disparities.