Expressive and Compressive Gan Inversion Network

Youmin Xu,Jian Zhang

DOI: https://doi.org/10.1109/icip42928.2021.9506059

2021-01-01

Abstract:This paper presents a novel encoding framework for the generative adversarial network (GAN), named compressive GAN inverter (Comp-GI), which can effectively compress, restore, and edit the attributes of real-world images. In particular, Comp-GI first transforms the input image into the latent variable space of GAN through an encoder network and then manipulates the latent code vector to enable various semantic image editing operations. StyleGAN [1] and its improved version StyleGANv2 [2] are adopted as the default GANs for image generation. Experiments on benchmark datasets FFHQ demonstrate that the proposed Comp-GI achieves better restoration performance with finer details and fewer artifacts than existing state-of-the-art GAN-inversion methods.