FUIT: Improving Semantic Consistency in Unpaired Image Translation via High and Low Frequency Guidance

With the development of generative models, un-paired image translation has made remarkable progress. It aims to translate an image from one domain to another domain while remaining the content information of the input image. However, existing methods still suffer from semantic inconsistency during the translation process. In this paper, we propose a novel unpaired image translation framework to improve semantic consistency via high and low frequency guidance. Our key idea is to divide images into high and low frequency components. Based on frequency representation of images, we design a new high-frequency encoding module that remains content semantic information from the original domain. We also present a novel style consistency regularization on low-frequency components, which enhances style consistency of the output images. Our approach achieves sota performance on multiple unpaired image translation tasks. Extensive experiments prove our method significantly improves semantic consistency that remains content semantics from the original images and generates style-harmonious images.