We improve the image synthesis quality of GANs by modifying the discriminator to take the shape of a popular segmentation network - U-Net, which allows to compute the adversarial loss on per-image and per-pixel basis. Consequently, the discriminator provides both global (image-level) and local (pixel-level) feedback to the generator and thereby enables the generator to produce higher quality images.
The use of the segmentation-based discriminator also opens new possibilities for regularization to further improve image quality, which we demonstrate with our proposed CutMix-based consistency regularization.
Link to the paper: https://arxiv.org/abs/2002.12655
