Self-supervised feature matched virtual try-on

Abstract Virtual try-on is a technology that enables users to preview the effect of wearing target garment without actual garment. However, existing image-based virtual methods often require additional human parsing or segmentation operations generate intermediate representations required for deformation and texture fusion. These not only increase computational complexity memory consumption, but also limit real-time portability try-on. Additionally, inaccurate results can lead misleading final generated images. To overcome these challenges, we propose self-supervised feature matched network, which directly high-quality from body images clothing any input. Specifically, design an optical flow warp module, focuses on changes between person image achieve accurate alignment deformation. Furthermore, refine module designed enhance features extracted information original character analysis operations, reducing influence background clutter content, ensuring wrinkles replacement clothes are close clothes. The match developed calculate matching loss converted teacher network student corresponding knowledge distilled passed assist in training. We conduct experiments VITON dataset show our model high quality resolution, proposed method outperforms state-of-the-art both qualitatively quantitatively.