Abstract Facial Beauty Prediction (FBP) remains a significant challenge in computer vision, primarily due to the inherent subjectivity of human perception and the difficulty in modeling the holistic harmony of facial features. Existing deep learning models, typically based on Convolutional Neural Networks (CNNs), often regress a deterministic score and struggle to surpass a performance plateau. In this work, we propose the Holo-Attentive Relational Network (H-ARN), a novel architecture designed to overcome these fundamental limitations. H-ARN makes two primary contributions: (1) It pioneers the use of distribution learning by predicting the mean and variance of beauty scores, employing a Gaussian Negative Log-Likelihood loss to robustly model the subjectivity and ambiguity in training data. (2) It introduces a relational self-attention module atop a powerful Vision Transformer (ViT) backbone, explicitly learning the complex, long-range relationships between facial components that define aesthetic harmony. Evaluated on the challenging SCUT-FBP5500 benchmark, our proposed H-ARN achieves a new state-of-the-art Pearson Correlation of 0.9333, significantly outperforming previous methods. The code will be made publicly available upon publication https://github.com/DjameleddineBoukhari/H-ARN .
Boukhari et al. (Sun,) studied this question.