Learning Controllable Face Generator from Disjoint Datasets

Jing Li, Yongkang Wong, Terence Sim

Abstract: Recently, GANs have become popular for synthesizing photorealistic facial images with desired facial attributes. However, crucial to the success of such networks is the availability of large-scale datasets that are fully-attributed, i.e., datasets in which the Cartesian product of all attribute values is present, as otherwise the learning becomes skewed. Such fully-attributed datasets are impractically expensive to collect. Many existing datasets are only partially-attributed, and do not have any subjects in common. It thus becomes important to be able to jointly learn from such datasets. In this paper, we propose a GAN-based facial image generator that can be trained on partially-attributed disjoint datasets. The key idea is to use a smaller, fully-attributed dataset to bridge the learning. Our generator (i) provides independent control of multiple attributes, and (ii) renders photorealistic facial images with target attributes.