Dual-Stage Detail-Preserving Virtual Try-On Network With Geometric Refinement and Multi-Scale Feature Integration

In this paper, we introduces a novel Virtual Try-on model designed to address the core challenges in virtual try-on systems: preserving garment details and accurately modeling clothing-body interactions to generate realistic and consistent results. Our framework features two primary modules: the Garment Geometric Alignment Module (GGAM) and the Garment-Body Fitting Module (GBFM). The GGAM tackles the issues of warping accuracy and detail preservation in existing virtual try-on systems through a progressive refinement warping network. This module combines Thin-Plate Spline (TPS)-based initial warping with an Attention U-Net-based refinement process, achieving high-quality garment warping that retains garment details while seamlessly aligning with the target body structure. The GBFM leverages a newly proposed Trans-Fusion U-Net architecture, incorporating three key innovations: (1) a Groupwise Feature Learning Encoder with self-attention mechanisms, (2) a multi-scale feature fusion strategy, and (3) a cross-attention-based refinement decoder. These architectural advancements ensure the effective preservation of garment details and realistic fitting, while balancing computational efficiency with long-range dependency modeling. Experimental results demonstrate that the proposed method significantly outperforms existing virtual try-on approaches in preserving garment details and maintaining body-garment alignment, producing more realistic and visually coherent try-on outcomes.