MobileStyleGAN: A Lightweight Convolutional Neural Network for High-Fidelity Image Synthesis

  • 2021-04-10 13:46:49
  • Sergei Belousov
  • 107

Abstract

In recent years, the use of Generative Adversarial Networks (GANs) has becomevery popular in generative image modeling. While style-based GAN architecturesyield state-of-the-art results in high-fidelity image synthesis,computationally, they are highly complex. In our work, we focus on theperformance optimization of style-based generative models. We analyze the mostcomputationally hard parts of StyleGAN2, and propose changes in the generatornetwork to make it possible to deploy style-based generative networks in theedge devices. We introduce MobileStyleGAN architecture, which has x3.5 fewerparameters and is x9.5 less computationally complex than StyleGAN2, whileproviding comparable quality.

 

Quick Read (beta)

loading the full paper ...