Unpaired Image-to-Image Translation Using Cycle-Consistent Adversarial Networks

The paper targets image-to-image translation when aligned image pairs are not available for training. It learns a mapping from a source domain X to a target domain Y using an adversarial loss so that the distribution of translated images is indistinguishable from the target distribution. Because this mapping alone is highly under-constrained, the authors couple it with an inverse mapping from Y back to X and introduce a cycle-consistency loss that pushes the round-trip translation to reconstruct the original image.

The approach enables translation across many settings where paired data does not exist, including collection style transfer, object transfiguration, season transfer, and photo enhancement. Qualitative results demonstrate compelling translations, and quantitative comparisons against several prior methods show the superiority of the approach, demonstrating the value of cycle consistency for unpaired image translation.