Contrastive Unpaired Translation

Contrastive Unpaired Translation (CUT) is an image-to-image translation framework that replaces the cycle-consistency loss used in CycleGAN with a contrastive learning objective. It computes mutual information between corresponding input and output patches, pulling them together in a learned feature space while pushing away negative samples from other patches within the same input image. This approach enables one-sided translation, meaning only a single generator and discriminator are needed, cutting training time roughly in half compared to CycleGAN while often improving output quality. The method supports standard unpaired translation between two domains, and even the extreme case where each domain consists of a single image. It is designed for researchers and practitioners in computer vision who need efficient, high-quality domain transfer without paired data. CUT is published at ECCV 2020 and provides open-source PyTorch code with pretrained models for common tasks like horse↔zebra, cat↔dog, and photo↔style transfer. The implementation includes a command-line interface for training and testing, with logging and visualization via Visdom. Compared to CycleGAN, CUT offers faster training and fewer artifacts, but it remains a research-oriented toolkit rather than a plug-and-play application.