LA-VITON: A Network for Looking-Attractive Virtual Try-on

Abstract

In this paper, we propose an image-based virtual try-on network, LA-VITON, which allows the generation of high fidelity try-on images that preserves both the overall appearance and the characteristics of clothing items. The proposed network consists of two modules: Geometric Matching Module (GMM) and Try-On Module (TOM). To warp in-shop clothing item to the desired image of a person with high accuracy in GMM, grid interval consistency loss and an occlusion handling technique are proposed. Grid interval consistency loss regularizes transformation to prevent distortion of patterns in clothes and an occlusion handling technique encourages proper warping despite target bodies are covered by hair or arms. The following TOM synthesizes the final try-on image of the target person seamlessly with the warped clothes from GMM. Extensive experiments on fashion datasets show that the proposed method outperforms the state-of-the-art methods.