A Conditional GAN-HRNet-DiffCloth Architecture for Personalized Garment Generation and Physics-Based Virtual Fitting

Abstract

This study presents a comprehensive intelligent framework for personalized garment creation and physics-based virtual fitting, incorporating conditional generative adversarial networks, high-resolution neural networks for human pose estimation, spatially-adaptive normalization, and a differentiable physical simulation engine for fabric dynamics. The system addresses critical issues in the fashion sector, including inadequate personalization in design and limited realism in virtual try-on technologies. User preferences are encoded through a pre-trained language representation model, while the system records body posture and simulates fabric deformation to attain high-fidelity virtual fitting. Experimental assessments were conducted using the DeepFashion image dataset, along with a proprietary three-dimensional human body scanning dataset. The proposed system attained a recommendation accuracy of 88 percent, a user satisfaction rate of 92 percent, and a novelty score of 8.9 out of 10. The pose estimation module achieved an accuracy of 96.8 percent according to keypoint localization benchmarks, while the average fabric deformation error decreased to 1.4 millimeters, signifying a 73.6 percent enhancement compared to conventional spring-based physical models. The results illustrate the system's ability to produce manufacturable, customized garment designs while providing realistic, dynamically adaptive virtual try-on experiences. The suggested method provides a scalable solution for innovative fashion design and consumer interaction.