The paper presents PluGeN4Faces, a tool that enhances the capabilities of the StyleGAN generative model by disentangling face attributes from a person’s identity. The authors argue that existing models are limited in their ability to modify specific face attributes without altering others because they are typically trained on generated images. PluGeN4Faces addresses this by training on real images, allowing it to maintain a person’s identity while modifying attributes like expression, hairstyle, or age. The approach uses contrastive loss to group images of the same person in similar regions of latent space, making attribute modifications less invasive.
Publication date: 21 Sep 2023
Project Page: https://arxiv.org/abs/2309.12033v1
Paper: https://arxiv.org/pdf/2309.12033
