Augmentation-Aware Self-Supervision for Data-Efficient GAN Training

We further encourage the generator to adversarially learn from the self-supervised discriminator by generating augmentation-predictable real and not fake data.