Semantic-Aware GAN Manipulations for Human Face Editing

Generative Adversarial Networks (GANs) have greatly advanced image generation, producing high-quality images, close to real. However, manipulation of the semantic properties of generated images remains a challenge. This study focuses on the study of manipulation methods in the intermediate latent space StyleGAN2. A comparison was made of several methods for identifying semantic meaningful directions in the latent space of StyleGAN2, which do not require labeled data. The quality of the received images and the received metrics are analyzed. The results obtained provide an understanding of the capabilities and limitations of existing methods for editing faces using the manipulation of GAN latent spaces. A method based on mapping the latent code z into the extended intermediate latent space \(W+\) and using a pretrained discriminator to control the quality of the edited image was also proposed. The results of experiments and the values of metrics are presented, confirming the improvement in image quality at large shifts in the latent space of StyleGAN2.