Realizing Virtual Try-On with Korean Garment Size Control Using Text-Guided Diffusion Model

This study proposes a diffusion-based virtual try-on model that generates garment try-on images with quantitatively controllable sizes using real-world body and clothing measurements specific to Korean people. While most existing virtual try-on methods rely on qualitative fit categories such as tight, regular, or loose, the proposed approach enables detailed size control by incorporating actual garment measurements into text prompts and fine-tuning a CLIP text encoder accordingly. The model integrates an image prompt adapter and GarmentNet into a text-conditioned Stable Diffusion framework, allowing it to preserve fine-grained visual details and garment identity while naturally adapting to the user's body shape. Trained on the AI-Hub Korean clothing dataset, which includes paired images and real-world measurements of persons and garments, the model can generate try-on results in various sizes based solely on numerical text input. Experimental results demonstrate superior performance compared to state-of-the-art methods across multiple metrics including FID, SSIM, and LPIPS, and show high visual quality and generalization in both paired and unpaired conditions. The proposed method enables efficient and accurate size control without complex preprocessing; also our method is expected to be used for delivering personalized and realistic virtual try-on experiences.