Unsupervised Multimodal Image-to-Image Translation: Generate What You Want

Abstract

Unsupervised image-to-image translation is one of the most important research topics in the field of computer vision. Traditional unsupervised image-to-image translation methods can only learn one-to-one mapping relation between two domains. However, the target domain is multimodal in general. Therefore, some existing methods generate diverse and multimodal outputs for a given source image by introducing random style codes or noise vectors. But they can only generate images with random modal of target domain. In some scenarios, we expect the generated images have desired modal. To solve this problem, we propose a novel framework for unsupervised multimodal image-to-image translation, called MuGAN, which can learn one-to-many mapping relations from source domain to target domain. And by controlling the modal label of target domain, our approach can generate images with desired modal. We verify the effectiveness and superiority of our framework by qualitative and quantitative evaluations on several image datasets.