Contrastive Swin-UNet for Unpaired Image-to-Image Scene Translation

Unpaired image-to-image translation is a challenging task that requires mapping between source and target domains without paired examples. Conventional approaches rely on convolutional neural networks (CNNs) and their variants to achieve reliable translation results. However, in applications where structural consistency is key, CNN-based models are unable to capture global and long-range dependencies. In this paper, we introduce Contrastive Swin-UNet, a framework that leverages contrastive learning with shifted window self-attention mechanism, structured in U-shaped encoder-decoder network. Our approach employs a fully transformer-based generation for effective local and global feature learning. Taking advantage of self-attention, we show that integrated contrastive loss significantly improves translation consistency and generative fidelity by maximizing mutual information between source image and translated output. Our findings indicate that Contrastive Swin-UNet sets a new benchmark for unpaired image-to-image translation tasks, particularly in scenarios requiring high precision and regularity.