Mixing and Matching Elements for Intelligent Fashion Design: A Generative Adversarial Network with Structure and Texture Disentanglement

Mixing and matching design elements from different fashion items to automatically create new textures or structures is desirable for fashion designers in order to facilitate the repetitive drawing process, and this mixing and matching process may also inspire them. In this research, we propose a structure and texture disentanglement generative adversarial network (STD-GAN) to create automated mix-and-match designs. Our model is trained to disentangle fashion items into different style elements and to integrate these elements to create a new design in an unsupervised manner. More specifically, a fashion attribute encoder is developed to disentangle the features of fashion items into two representations based on structure and texture. A texture mapping network is then applied to encode the texture representation in the form of different spatial features. A fashion fusion decoder is also developed that can generate mixed-style fashion items by utilizing the structure representation and the different texture features. In addition, a multi-discriminator framework is proposed to determine the authenticity and texture similarity of the reconstructed and mixed fashion items. Extensive experimental results demonstrate the effectiveness of our STD-GAN and its potential to facilitate the fashion design process by creating different textures and structures in a mix-and-match manner.