Generative Adversarial Style Transfer Networks for Face Aging

How somebody looked like when younger? What could a person look like when 10 years older? In this paper we look at the problem of face aging, which relates to processing an image of a face to change its apparent age. This task involves synthesizing images and modeling the aging process, which both are problems that have recently enjoyed much research interest in the field of face and gesture recognition. We propose to look at the problem from the perspective of image style transfer, where we consider the age of the person as the underlying style of the image. We show that for large age differences, convincing face aging can be achieved by formulating the problem with a pairwise training of Cycle-consistent Generative Adversarial Networks (CycleGAN) over age groups. Furthermore, we propose a variant of CycleGAN which directly incorporates a pre-trained age prediction model, which performs better when the desired age difference is smaller. The proposed approaches are complementary in strengths and their fusion performs well for any desired level of aging effect. We quantitatively evaluate our proposed method through a user study and show that it outperforms prior state-of-the-art techniques for face aging.