VariSize: A Dataset and Model for $3\mathrm{D}$ Garment Grading and Editing Based on Latent Space

Garment size is a critical feature of garments. Traditional resizing usually requires grading 2D sewing patterns according to specific rules and then sewing them together. This process is not only unintuitive but also time-consuming. Some methods aim to achieve a better fit by conforming the garment to the body's shape, which alters the garment structure. To simplify this grading process, we propose an automated 3D garment grading method that changes the garment size while maintaining its original structure. We compress garments into latent codes within a high-dimensional latent space. Then we introduce VFNet and SVNet, VFNet employs a multi-class logistic regression model to generate probabilities for attribute labels mapped to latent vectors. By controlling the weights of garment attributes, it enables feature part editing. SVNet learns the size transformation relationships of garments within the latent space, facilitating 3D garment resizing through label-based adjustments. Due to the lack of datasets containing garments of different sizes, we constructed a dataset called VariSize, which includes five types of garments: T-shirts, cardigans, trousers, dresses, and skirts, with each type having four sizes. Each garment's different parts are labeled accordingly. Experimental results demonstrate that our method successfully achieves automatic 3D garment resizing and feature part editing, offering significant value for 3D virtual try -on and garment design.