Image-to-Image Translation

Image-to-image translation studies how to convert an input image into another style, condition, or modality while retaining the information needed for people or downstream systems. Applications include MRI-to-PET synthesis, SAR-to-optical conversion, day-to-night scene translation, and image restoration, evaluated with paired or unpaired data and benchmarks such as BraTS and remote-sensing datasets. Methods range from paired U-Net regression and GANs to diffusion bridges, Brownian bridges, and rectified flow, with foundation-model features, physical priors, and semantic guidance increasingly used to preserve important structure. A central tension is balancing realism and diversity against faithful anatomy, identity, and object semantics; diffusion and flow methods can help model ambiguity but may cost more or hallucinate details, while simpler models can be faster and more faithful in some paired settings.