CA-VTON: Correlation-Aware Image-Based Virtual Try-On Network

Image-based virtual try-on technology aims to transfer a given clothing item onto the corresponding region of a human body image. Existing methods tend to suffer from generating significant amounts of artifacts, particularly when dealing with complex human poses or pose intersections. This limitation severely constrains the poses of input person image and consequently impacts the practicality of virtual try-on networks. To address this predicament, we propose a novel Correlation-Aware image-based Virtual Try-on Network (CA-VTON). Our method mainly consists of two components: a softmax-correlation-guided Optical Flow Warping Module (OFWM) to warp the clothing according to the pose of reference person, and a Try-on Image Synthesis Module (TISM) guided by pose correlation, which supervises the generation process of try-on results. Extensive qualitative and quantitative experiments validate the effectiveness of our proposed method.