Automatic generation of garment designs using generative adversarial networks

Abstract

Generative models like GANs are able to generate realistic samples [1] https://thispersondoesnotexist.com. GANs have been used in the fashion domain as well. Manipulating attributes of a given garment [2], filling a garment sketch using a given fabric [3], generating new clothing on an image of a wearer through text description [4] are a few of such usages. However modeling a distribution of the dresses and manipulating the attributes of the generated dresses are not up to date with the advancement of the GANs. Firstly, the current state of the art GAN models and their applicability for the dress images is analyzed. Then the methods of manipulating the attributes of generated dresses by interpreting the latent space are explored. Finally, the application of the GANs for dress images and a way to interpret the latent code to manipulate the dress attributes successfully are presented with results.