Handcrafting Generative Adversarial Networks for Fashion Design Generation

Arun Patro, Vishnu Vardhan Makkapati, Jayanta Mukhopadhyay

Keywords: Generative Adversarial Networks, Regularization, Symmetry, Periodicity, Tiling, Generative Models, Generator, Discriminator

TL;DR: We propose methods to generate fashion images by exploiting the symmetry and periodicity of the patterns.

Abstract: Natural images, particularly, those related to fashion exhibit structural coherence including symmetry. This is also true for natural textures which comprise of a repeating periodic pattern. In this paper, we build on existing work and demonstrate that it is possible to generate images of desired characteristics. We model losses based on symmetry and periodicity to train Generative Adversarial Networks. We extend the notion of a fictional generator that modifies the generated images to those with expected properties. The discriminator is trained to classify them equivalently as real or fake. We also propose a regularizer on the generator to produce images with specified properties. We use data from fashion catalog to train the models and show that generating high resolution images is feasible. We present numerical results to quantitatively evaluate the methods in comparison to existing ones and include images to qualitatively show they produce images of superior quality.