Information-based Boundary Equilibrium Generative Adversarial Networks with Disentangled Representation Learning

Abstract

This paper describes a new image generation algorithm, developed from BEGAN. With an information-theoretic extension to BEGAN, this new algorithm is able to learn disentangled representations from image data sets. The information-theoretic extension was inspired by InfoGAN, a generative adversarial network that maximizes the mutual information between small subset of the latent variables and the observation. By using the proposed method of this paper, one can manipulate the generated images as desired by controlling the latent codes of input variables. Also the visual qualities of produced images are highly maintained.