A Study on Region-Aware Fashion Design Based on Probabilistic Diffusion Model

Abstract

To address the issue of image distortion caused by substantial structural discrepancies between original garment images and reference images in the application of image transfer technology for fashion design, this study proposes a region-aware fashion design method based on a probabilistic diffusion model. During the image feature extraction and output stage, the method integrates vision transformer (ViT) with a mask-guided mechanism, enabling the Diffusion model to precisely focus on the transferable regions of the original and reference images, thereby preserving the structural integrity and semantic consistency of the source images effectively. In the image colour and pattern style transfer stage, this study introduces an asymmetric gradient guidance (AGG) strategy to optimise the reverse sampling process of the diffusion model, substantially improving the quality and visual fidelity of the generated images. Experimental results indicate that this method achieves a Fréchet inception distance (FID) score of 103.4, surpassing existing fashion synthesis models. This facilitates the generation of more stable and realistic images for garment design tasks.