RSGAN: face swapping and editing using face and hair representation in latent spaces

This abstract introduces a generative neural network for face swapping and editing face images. We refer to this network as "region-separative generative adversarial network (RSGAN)". In existing deep generative models such as Variational autoencoder (VAE) and Generative adversarial network (GAN), training data must represent what the generative models synthesize. For example, image inpainting is achieved by training images with and without holes. However, it is difficult or even impossible to prepare a dataset which includes face images both before and after face swapping because faces of real people cannot be swapped without surgical operations. We tackle this problem by training the network so that it synthesizes synthesize a natural face image from an arbitrary pair of face and hair appearances. In addition to face swapping, the proposed network can be applied to other editing applications, such as visual attribute editing and random face parts synthesis.