A Latent Transformer for Disentangled Face Editing in Images and Videos

Abstract

We project real images to the latent space of a StyleGAN generator and achieve sequential disentangled attribute editing on the encoded latent codes. From the original and the projected image, we can edit sequentially a list of attributes such as: 'smile', 'bangs', 'arched eyebrows', 'age', 'beard' and 'eyeglasses'. All results are obtained at resolution 1024 × 1024.