Implementation of Fashion Reference Model for Sketch-based Image Generation and Unstructured Search

This paper proposes an integrated fashion system that unifies sketch-based fashion image generation, multi-attribute classification, and similar item retrieval into a single end-to-end pipeline. This study adopts a conditional generation model that combines Stable Diffusion XL with ControlNet, enabling the synthesis of garment images that match the user’s sketch with realism comparable to actual fashion products. The generated images are processed by a DenseNet121-based attribute classifier to produce probability vectors for detailed category-specific attributes. Finally, an L2-distance-based similarity function is applied to these attribute probability vectors, allowing quantitative measurement of similarity between the generated image and online retail products and enabling real-world item recommendation. The proposed system not only transforms user sketches into realistic garment images but also provides an effective mechanism for retrieving actual fashion items based on predicted attributes, demonstrating its potential for enhanced communication, design support, and automated fashion search applications.