Disentangled face editing via individual walk in personalized facial semantic field

Recent generative adversarial networks (GANs) can synthesize high-fidelity faces and the closely followed works show the existence of facial semantic field in the latent spaces. This motivates several latest works to edit faces via finding semantic directions in the universal facial semantic field of GAN to walk along. However, several challenges still exist during editing: identity loss, attribute entanglement and background variation. In this work, we first propose a personalized facial semantic field (PFSF) for each instead of a universal facial semantic field for all instances. The PFSF is built via portrait-masked retraining of the generator of StyleGAN together with the inversion model, which can preserve identity details for real faces. Furthermore, we propose an individual walk in the learned PFSF to perform disentangled face editing. Finally, the edited portrait is fused back into the original image with the constraint of the portrait mask, which can preserve the background. Extensive experimental results validate that our method performs well in identity preservation, background maintenance and disentangled editing, significantly surpassing related state-of-the-art methods.