30th Signal Processing and Communications Applications Conference, SIU 2022, Safranbolu, Turkey, 15 - 18 May 2022
© 2022 IEEE.In this work, we present a self-supervised method to identify the semantic axes in the latent space of a generator in a generative adversarial network (GAN) setting. In this context, a novel Direction-Disentangler (DD) model is proposed. The data that is necessary for the training of this model is generated using the generator in a self-supervised fashion. Using this data, the DD model is trained to identify the direction that encodes the discriminative features between the two reference images in the generator's latent space. The early empirical results using the MNIST dataset show that the identified direction is effective for editing the relevant semantic attributes of different reference images as well.