Controlling GANs Latent Space [in Russian]

Speaker: Vadim Titov, MIPT, AIRI Slides: Modern GAN architectures generate highly realistic images in a variety of domains. Much recent works have focused on understanding how its latent space is connected with generated image semantic. It is discovered that there exist meaningful latent manipulations that allow to semantically edit image. Many proposed methods include supervision from pretrained models which is a strong limitation. This weakness could be eliminated by unsupervised methods that have their own disadvantages. In this talk we will describe the main directions (supervised, unsupervised, text-guided) and current state-of-the-art methods of semantic image manipulation through GAN latent space.