Beauty is in the ELBO of the Beholder: A Variational Account of Processing Fluency in Face Perception

What makes a face beautiful? Classic theories in psychology suggest that beauty is linked to statistical regularities like symmetry and averageness, which may facilitate easier perceptual processing. This paper investigates whether this processing fluency—the ease with which a stimulus is encoded—can be modeled using modern generative AI, specifically variational autoencoders (VAEs).

The authors trained convolutional VAEs on four diverse face datasets (FairFace, FFHQ, CelebA, and UTKFace) without any attractiveness labels. They then evaluated these models on 597 neutral-expression faces from the Chicago Face Database (CFD). By analyzing the VAE's internal objective function—the evidence lower bound (ELBO)—the researchers measured how well each face was reconstructed (distortion) and how much information was required to represent it (rate). They compared these computational metrics against human attractiveness ratings to see if the model's internal "fluency" matched human aesthetic judgments.

The study reveals that human attractiveness ratings are consistently aligned with the VAE's ELBO. Attractive faces are not only reconstructed with higher fidelity (lower distortion) but are also represented more efficiently (lower rate). This relationship holds across all four training datasets, suggesting that the model's unsupervised learning process naturally organizes faces along a dimension of aesthetic preference. Furthermore, the authors found that attractive faces are more prototypical in both physical landmark space and the model's latent space. When they aligned the latent spaces of independently trained models, they discovered a shared "attractiveness direction" that transfers across different models and datasets, indicating that this aesthetic geometry is a robust property of generative face representations.