AI and the World's Hard Problems · Economics & social sciences · arXiv · 2026-10-21
Beauty is in the ELBO of the Beholder: A Variational Account of Processing Fluency in Face Perception
Facial attractiveness has been linked to statistical regularities such as symmetry and averageness, suggesting that beauty may depend on the ease with which a face is perceived. To test this, variational autoencoders were trained on four face datasets without any attractiveness supervision and evaluated on the 597 faces of the Chicago Face Database. Human attractiveness ratings closely aligned with the direction defined by the VAE evidence lower bound (ELBO) in rate-distortion space, independently learned latent spaces contained an attractiveness direction that transferred strongly across initializations and training data, and attractive faces proved more prototypical in both shape and latent space, giving the first explicit link between classic accounts of aesthetics and learned generative models.