Personalized Clothing Style Generation System Based on StyleGAN3
N. Xu · Advanced Electromagnetics · 2026
In order to solve the core challenges of personalized clothing generation – uncontrollable style, difficult coupling of user preferences and body shape mismatch, this paper proposes a fine-grained automatic error correction system based on Bert, and uses the stylegan3 model. The system begins by parsing user-provided textual descriptions (e.g., style, color preferences) using a pretrained language model to extract structured semantic embeddings. These embeddings, along with body-shape encodings, are then mapped into a unified conditional vector. Our key contributions include a Fine-Grained Style Decoupling Encoder, which employs a frozen ResNet50 backbone with 12 parallel regression heads to independently predict continuous and categorical attributes (e.g., collar type, pattern density). An Attribute Conditional Injection Module adopts a dual-path latent modulation strategy: body conditions are mapped to deformation fields to drive pose-aware latent warping, while style/color embeddings are encoded as AdaIN parameters and injected into StyleGAN3 synthesis layers, ensuring semantic alignment between user intent and the generation process. Differentiable grid sampling enables adaptive adjustments to structural attributes such as shoulder width and waistline while preserving texture details. Experiments on DeepFashion demonstrate superior performance over mainstream methods (e.g., DragGAN, ControlNet+SD), with improved generation quality (FID=12.1), higher attribute control accuracy (AA=89.3%), and better detail preservation after body deformation (LPIPS=0.14). The system achieves high inference efficiency (54ms latency, 18.5 img/s), offering industrial-grade deployability for virtual try-on and personalized e-commerce, thereby supporting personalized e-commerce and virtual try-on. The body-aware garment model is also relevant to wearable product design in which device placement and fit depend on human morphology.