Latent space representation and manipulation of StyleGANs

Abstract

StyleGAN models are a new paradigm in artificial image generation. Initially proposed to generate fake facial images, these deep generative models can be used to edit real photographs with the aid of GAN Inversion and Latent Manipulation algorithms. GAN Inversion techniques embed real images into StyleGAN’s latent space, yielding a latent point used to generate an artificial image as close as possible to the original one. Latent Manipulation operations work on top of the resulting latent point to do semantic-oriented edits, aiming to preserve the remaining image’s characteristics. Recent research initiatives attempt to better understand and model the rich latent space from StyleGANs, focusing on how to invert and edit real images with minimum distortion and maximum editing efficiency. Although significant advances have been made in the past few years, GAN In version and Latent Manipulation methods still face difficulties when trying to disentangle semantic features in latent spaces. A better understanding of basic latent-space arith metic is needed to assess the entanglement of StyleGAN’s semantic features. This under graduate thesis describes the state-of-the-art in this field, defines basic latent arithmetic operations, and performs a variety of latent arithmetic experiments. The experimental results are used to develop a better understanding of StyleGAN’s latent space, setting a theoretical basis for future research directions in GAN Inversion and Latent Manipulation.