Analyse et Manipulation de l’espace latent d’un GAN pour la génération d’image

Abstract

In the field of image synthesis and analysis, Generative Adversarial Networks (GANs) have made significant advances. In particular, the StyleGAN model is considered one of the most powerful models for generating images of faces. Its results are of remarkable visual fidelity and quality. Editing a real image requires conversion of the input image into StyleGAN's latent variables. However, it is still difficult to find latent variables that allow meaningful manipulation. In this end-of-study project, the aim is to implement an application capable of controlling image generation by manipulating latent space. The proposed system begins by extracting face attributes from the input image using a CNN via the transfer learning technique, a crucial step in our system. The detected attributes are then presented to the user, so that he can modify those he wishes to change. These modified attributes are then concatenated with a randomly generated vector, to form the latent input vector of the StyleGAN3 generator. This generator was trained on a random selection from the CelebA database. The attribute extraction system was trained on 10% of the CelebA database around 20k images, and obtained a rate of 86%, and the stylegan3 generator was trained on a selected part of the CelebA dataset. Although these results remain insufficient and are encouraging, it is important to emphasize that increasing the training duration will allow a clear improvement in the quality of the Images generated as well as the training method of this Gan