Designing an encoder for StyleGAN image manipulation

Recently, there has been a surge of diverse methods for performing image editing by employing pre-trained unconditional generators. Applying these on real images, however, remains challenge, as it necessarily requires the inversion images into their latent space. To successfully invert image, one needs to find code that reconstructs input accurately, and more importantly, allows its meaningful manipulation. In this paper, we carefully study space StyleGAN, state-of-the-art generator. We identify analyze existence distortion-editability tradeoff distortion-perception within StyleGAN then suggest two principles designing encoders in manner control proximity inversions regions was originally trained on. present an encoder based our is specifically designed facilitating balancing tradeoffs. By evaluating performance qualitatively quantitatively numerous challenging domains, including cars horses, show method, followed common techniques, achieves superior real-image quality, with only small reconstruction accuracy drop.