Abstract: Generating photo-realistic images from labels (e.g., semantic labels or sketch labels) is much more challenging than the general image-to-image translation task, mainly due to the large ...