CycleGAN Based on Relative Loss Functions

Abstract

At the upsurge of deep learning, leap-forward achievements have been made in the field of computer vision, among which the application of CycleGAN to style transfer is eye-catching. The CycleGAN theory argues that concentration on making fake data closer to the real value alone is unfavorable to the stability of the network output. This paper introduces the concept of relativity to CycleGAN so to improve the stability of the network.