Adaptação da StyleGAN2 para geração bidirecional de faces individuais via transfer learning

Abstract

This work presents an approach to optimize the generation of individual faces using transfer learning based on the StyleGAN2 network. The objective is to reduce training time and avoid the need to start each facial generation from scratch while maintaining high quality in the generated images. An architecture was adapted and coupled to the pretrained generator, aiming to promote specialization, as well as the selective freezing of parts of the original generator and the discriminator to enable processing optimization. Subsequently, the approach was combined with an independent encoder responsible for mapping real images to the latent space, allowing the reconstruction of faces from their latent vectors. The obtained results demonstrated consistent generations and reconstructions, evidencing the potential of the methodology for personalized facial synthesis applications.