Robust GAN inversion
Keywords: GAN inversion, regularization strategy, optimization
Abstract: Recent advancements in real image editing have been attributed to the exploration of Generative Adversarial Networks (GANs) latent space. However, a key challenge in this process is GAN inversion, which aims to accurately map images to the latent space. Current methods working on the extended latent space $W+$ struggle to achieve low distortion and high editability simultaneously. In response to this challenge, we propose an approach that operates in the native latent space $W$ and fine-tunes the generator network to restore missing image details. This method introduces a novel regularization strategy with learnable coefficients acquired through training a randomized StyleGAN 2 model - WRanGAN, surpassing traditional approaches in terms of reconstruction quality and computational efficiency. It achieves the lowest distortion compared to traditional methods. Furthermore, we observe a slight improvement in the quality of constructing hyperplanes corresponding to binary image attributes. The effectiveness of our approach is validated through experiments on two complex datasets: Flickr-Faces-HQ and LSUN Church
Submission Number: 78
Loading