Unpaired Image-to-Image Translation

Unpaired image-to-image translation studies how to change an image’s appearance or modality when corresponding examples from the two domains are unavailable, enabling applications such as synthetic-to-real adaptation, medical imaging, and day-to-night translation. Methods include cycle-consistent GANs, patch-contrastive learning, and newer diffusion, Schrödinger bridge, and flow-matching approaches, often guided by semantic features or anatomical and geometric constraints. A central challenge is balancing target-domain realism with faithful preservation of objects, anatomy, and other task-relevant structure. Image-level metrics alone may not reveal harmful changes, so evaluation increasingly considers downstream tasks, expert assessment, and uncertainty as well as visual quality.