The paper addresses a problem in self supervised GAN, where the classes strictly have disjoint support. This is mitigated by introducing a new class for generated samples. Experiments are convincing.