Effective Encoders for the StyleGAN Model

The area of neural networks and deep learning have made significant progress in the last couple of years, in particular the networks have achieved a noticeable quality increase in the tasks for reconstruction and editing of images. One of the most popular methods to do this is using a special type of Generative Adversarial Network model (GAN), called StyleGAN, and its encoder implementations. Firstly, the code for generating a real image is projected into the latent space, secondly this code is edited in a special way using a pretrained algorithm, lastly the output image is synthesized from the edited code using a trained generator. This encoding process has 3 characteristics: speed of execution, quality of reconstruction of the original image and editing capabilities. This work focuses on improving the quality of image editing of the StyleFeatureEditor (SFE) architecture which uses both the high-dimensional feature space F as well as the low-dimensional space W for latent search. Without losing speed and inversion quality, we propose to modify the way with which the direction of the edit is obtained during the training cycle of the Feature Editor. The method is compared to its original and is qualitatively accepted to hold the same clarity in inversion and editing tasks.