Clothing Recommender System Based on Large Language Models and Virtual Try-On

With the growing demand for fashion shopping experiences, the limitations of traditional recommender systems in capturing users' preferences and delivering intuitive try-on effects have become increasingly evident. In this paper, we propose a novel clothing recommender system that combines large language models (LLMs) and virtual try-on technology. The proposed system introduces a recommendation pipeline that integrates bi-encoder and cross-encoder architectures. By incorporating a LLM to recognize user intent and generate recommendation explanations, this framework enables end-to-end conversational recommendations, from efficient candidate recall to accurate ranking. Simultaneously, the virtual try-on module, based on the diffusion model (CatVTON), generates high-fidelity try-on images using the clothing and user-uploaded person images, significantly improving the user experience. This study established four kinds of test scenarios-item recommendation, occasion-based recommendation, item pairing recommendation, and style consistency recommendation-to quantitatively and qualitatively evaluate the system's performance. Experimental results indicate that incorporating a LLM for user requirements analysis enhances recommendation precision. The system exhibited strong performance in recommendation accuracy, explanation rationality, and virtual try-on realism, confirming its technical feasibility and offering valuable insights for advancing clothing recommendation systems.