Enhancing online clothing retail with generative AI: Innovations in virtual try-on and beyond

Abstract

The trend of customers purchasing clothing products online has seen a significant increase, especially during the Covid-19 pandemic. Customers find it convenient to browse a wide range of items and purchase from anywhere. However, one challenge is that online shopping does not provide the same experience as shopping in physical stores. Customers miss out on the opportunity to try on clothing before making a purchase, leading to potential issues such as dissatisfaction and product returns. This thesis aims to demonstrate how generative AI can address these issues by replicating the physical shopping experience and allowing more interactions. This work contributes to improving existing generative AI models in the fashion context. These models include image-based and multi-pose virtual try-ons. Image-based virtual try-ons allow customers to apply desired clothing to an image of themselves. The proposed model refines the input data to enhance the accuracy of segmentation synthesis and occlusion handling. The virtual try-on model is faster than competitors due to the truncation of the U-Net and the use of affine transformation to perform the geometrical transformation of the clothing. The multi-pose variant enables customers to change the posture of the synthesised image, allowing for a wider range of viewing angles and posture styles. The contribution led to the development of a model combining techniques from virtual try-on and pose transfer. It is demonstrated how multiple discriminators enhance the warping performance for high-resolution images. Additionally, the method for fine-tuning the pose transfer module to adapt it for multi-pose virtual try-on is outlined. This thesis also innovates by proposing an image-to-video synthesis model for creating fashion videos from a single image, a concept not previously explored. This model provides customers with more detailed information about clothing products, showing how they would look while being worn from various angles and how the clothing flows as the person moves. The contribution has led to the development of a video diffusion model. It is shown how the conditioning image is incorporated into the latent video through cross-attention. Using the conditioning image as the initial frame enables subsequent frames to capture detailed clothing characteristics. In order to encourage businesses to employ generative AI, this thesis introduces the MARKGEN framework, which serves as a non-technical comprehensive guide on how to apply generative AI for marketing purposes. This framework aims to simplify the integration of generative AI into business platforms, enabling them to reap its benefits and enhance the overall shopping experience.