CUNIT-GAN: Constraining Latent Space for Unsupervised Multi-domain Image-to-Image Translation via Generative Adversarial Network

Abstract

Image-to-image translation has gained significant interest due to the success of deep learning models that enforce cycle consistency constraints. However, the recent studies are particularly limited to a subset of domains with significant constraints on style or texture variations. Also, these models show limited performance in multi-domain settings where one image is translated to numerous domains. We propose a Constrained Unsupervised Image-to-Image Generative Adversarial Network (CUNIT-GAN) to address the above problems. It consists of an asymmetric Auto-encoder (AE) based Generator network and a dual-purpose Discriminator network that detects real and fake samples and classifies the input image. This study focuses on enhancing the explainability and representation power of the multidomain latent space through our novel latent contrastive loss, which leads to the clustering of class-level feature embeddings and decoupling of latent space. The effectiveness of CUNIT-GAN is demonstrated through a comprehensive qualitative and quantitative analysis conducted on benchmark multi-domain image datasets. © 2025 Elsevier B.V., All rights reserved.