Gradient Adjusting Networks for Domain Inversion

StyleGAN2 was demonstrated to be a powerful image generation engine that supports semantic editing. However, in order manipulate real-world image, one first needs able retrieve its corresponding latent representation StyleGAN’s space is decoded an as close possible the desired image. For many images, does not exist, which necessitates tuning of generator network. We present per-image optimization method tunes such it achieves local edit generator’s weights, resulting almost perfect inversion, while still allowing editing, by keeping rest mapping between input tensor and output relatively intact. The based on one-shot training set shallow update networks (aka. Gradient Modification Modules) modify layers generator. After Modules, modified obtained single application these original parameters, previous editing capabilities are maintained. Our experiments show sizable gap performance over current state art this very active domain. code available at https://github.com/sheffier/gani .