Liu Ming-Yu ; Breuel Thomas ; Kautz Jan

Abstract

Unsupervised image-to-image translation aims at learning a joint distributionof images in different domains by using images from the marginal distributionsin individual domains. Since there exists an infinite set of jointdistributions that can arrive the given marginal distributions, one could infernothing about the joint distribution from the marginal distributions withoutadditional assumptions. To address the problem, we make a shared-latent spaceassumption and propose an unsupervised image-to-image translation frameworkbased on Coupled GANs. We compare the proposed framework with competingapproaches and present high quality image translation results on variouschallenging unsupervised image translation tasks, including street scene imagetranslation, animal image translation, and face image translation. We alsoapply the proposed framework to domain adaptation and achieve state-of-the-artperformance on benchmark datasets. Code and additional results are available inhttps://github.com/mingyuliutw/unit .