Nonlinear Facial Attribute Editing Method based on Autoencoder in StyleGAN Latent Space

Currently, facial attribute editing based on StyleGAN latent space has achieved great success, but it suffers from problems such as attribute entanglement and identity information loss in real facial attribute editing. To address these issues, this paper suggests utilizing an autoencoder network to extract the high-level identity information from the latent code, and then reconstruct the latent code following the incorporation of conditional semantic information. The embedded conditional semantic information aims to guide the autoencoder to achieve facial attribute editing. In the training process, reconstruction loss and conditional loss are introduced to constrain the autoencoder network, so that it can reconstruct the editing results that keep the identity information unchanged and meet the preset attribute change conditions. The experiments show that this method achieves facial attribute editing with no entanglement, controllability, and retention of identity features in real facial data. In several quantitative evaluation indicators, proposed method is better than the recent methods.