Demographic Bias Effects on Face Image Synthesis

Abstract

Face image synthesis has shown remarkable progress in recent years. However, the effect that the demographics of the data used to train synthesizers has on the generation of new face images remains an open question. This paper investigates the effects of the training set demographics in the face image synthesis task. To this end, we propose a strategy that allows synthesizing face images for specific groups of people with a high visual quality. The strategy uses an unsupervised learning approach to discover groups of people in the training set based on Bayesian inference via a probabilistic mixture model. If labels are available to define the groups, our strategy can also exploit such information in lieu of unsupervised learning. Once the groups are defined, our strategy trains a Generative Adversarial Network on each group to generate new face images with specific characteristics. Our results show remarkable performance in terms of image quality compared to several state-of-the-art baselines. More importantly, our strategy allows synthesizing face images with reduced demographic biases.