Virtual Measurement Garment for Per-Garment Virtual Try-On

The popularity of virtual try-on methods has increased in recent years as they allow users to preview the appearance of garments on themselves without physically wearing them. However, existing image-based methods for general virtual try-on provide limited support to synthesize realistic and consistent garment images under different poses, due to two main difficulties: 1) the dataset used to train these methods contains a vast collection of garments, but they lack fine details of each garment; 2) they synthesize results by warping the front-view image of the target garment in a rest pose, which results in poor quality and detail for other viewpoints and poses. To overcome these drawbacks, per-garment virtual try-on methods train garment-specific networks that can produce high-quality results with fine-grained details for a particular target garment. However, existing per-garment virtual try-on methods require the use of a physical measurement garment, which limits their applicability. In this paper, we propose a novel per-garment virtual try-on method that leverages a virtual measurement garment, which eliminates the need for the physical measurement garment, to guide the synthesis of high-quality and temporally consistent garment images under various poses. Furthermore, we introduce a gap-filling module that effectively fills the gap between the synthesized garment and body parts. We conduct qualitative and quantitative evaluations against a state-of-the-art image-based virtual try-on method and ablation studies to demonstrate that our method achieves superior performance in terms of realism and consistency of the generated garment images.

Recommended citation: Zaiqiang Wu^*, Jingyuan Liu^* (equal contribution), Toby Chong, I-Chao Shen, Takeo Igarashi. Proceedings of the 50th Graphics Interface Conference . Graphic Interface 2024.
Download Paper