Topology and Detailing in Personalized 3D Garment Modeling Through Neural Implicit Representations and Constraining GANs

Abstract

Current 3D clothing modeling methods often face the problems of discontinuous geometric expression, poor style adaptability and lack of high-quality generation mechanism when dealing with complex topological structure and personalized design requirements. In the aspect of 3D geometric reconstruction, an implicit coding method based on the signed distance function is introduced to realize a high-precision representation of the garment surface in continuous space, and the multi-resolution feature fusion strategy is used to enhance the modeling ability of details of the model. A topology-aware adversarial generation network architecture is constructed, graph convolution is introduced to extract clothing topological structure features, and the generation process is optimized by combining physical constraints to enhance the network's understanding of geometric continuity and physical rationality. In personalized modeling, the mapping mechanism of human morphological parameters to implicit space is systematically constructed, and a conditional generation control method of dynamically deformable topological structure is proposed. The methodology integrates three core components: (1) an Signed Distance Function (SDF)-based implicit encoder with multi-resolution feature fusion, trained on 12,500 garment models to achieve sub-millimeter geometric precision; (2) a topology-aware Generative Adversarial Network (GAN) with graph convolution for structural feature extraction, optimized via 200,000 training iterations; and (3) a conditional control module mapping human parameters (e.g., bust, waist circumference) to garment topology, validated across 34 diverse body shapes. Statistical analysis shows significant improvements (p<0.01) in wrinkle detail capture (67.5% vs. 55.2% for traditional methods) and topological integrity (56.78% recall, 95% CI [52.3%, 61.2%]).