Abstract
As a class of generative models, generative adversarial networks (GANs)
are now capable of achieving excellent results, especially relating to
facial images, which could reach 1024*1024 resolution in high-quality
face synthesis. Recently, some studies have tried to map real face
images into the latent space of GANs, for pre-trained GANs to represent
real faces (E.g., optimizing style latent space on StyleGANs). However,
the application of latent space to real face morphing is still in its
infancy. Currently, training costs are high and/or require huge samples
with labels. By adding regularization to the latent optimization, we
propose a novel method to morph real faces based on StyleGAN. We
labelled 12,000 face images with 32 attributes and utilise
logistic-regression models to discover independent attribute vectors in
latent space. Appropriate regularization helps us to regularize both
latent vectors (a face with its attribute) so that they can fall into an
ideal area. Moreover, we use those attribute vectors under different
layer representations to handle real face morphing. Compared to the
existing baseline, our method yields a larger output with higher
quality.