24-GAN: Portrait Generation with Composite Attributes

Abstract

We present a portrait-generation framework that can control composite attributes. Our target attribute covers the three global attributes of age, sex, and race. We built control vectors for the garget attributes and attached them to the latent vector that produced portrait images. Our generator was devised using the StyleGAN generator, and our discriminator has a dual structure for qualities and attributes. Our framework successfully generated 24 faces with the same identity while varying the three attributes. We evaluated our results from three aspects. The identity of the generated faces was estimated using Frechet inception distance, and the attributes of the generated faces were validated using a facial-attribute recognition model. We also performed a user study for further evaluation.