Improving CycleGAN Image-to-Image Translation Results by Refining Cycle Consistency Loss

This paper proposes modifications to the cycle consistency loss in CycleGAN to improve the realism of image-to-image translation, addressing issues like unrealistic artifacts caused by overly strict pixel-level cycle consistency.