Generating Fashion Sketches Using Generative Adversarial Networks

Generative Adversarial Networks (GANs) are an active area of research in deep learning and computer vision. They are concerned with generating and manipulating photorealistic synthetic images as well as manipulating real ones. This work used two types of GANs to generate new fashion sketches. The first type was Deep Convolutional GAN (DCGAN), which is one of the earliest GAN models. The results were acceptable compared with the size of the data and the basic types of networks used. For the second model, fashion sketches were generated using StyleGAN-ADA, a powerful model developed mainly for limited datasets. The results were much better visually, with an Fréchet Inception Distance (FID) score of 15.69 and a Kernel Inception Distance (KID) score of 3.2e-3 for 50k generated images against the entire dataset. Besides its ability to be used for data augmentation in projects that use fashion sketches, our research can be a source of inspiration for fashion designers.