SuperFashion boosts fashion image retrieval by 9.35% with superpixel tokens

Existing patch-based Transformer methods for attribute-specific fashion retrieval (ASFR) struggle with irregular attribute regions and background noise, limiting their ability to capture subtle microstructures. To address this, researchers introduce SuperFashion, the first ASFR framework that replaces traditional patches with superpixel tokens. The approach first uses an attribute-guided attention mechanism to extract attribute-related features and crop semantically meaningful image regions. Superpixel segmentation then generates compact, semantically coherent tokens from those regions. By incorporating modality-specific embeddings for both attribute and superpixel tokens, the superpixel token-based Transformer enables adaptive interaction and fusion, significantly improving attribute localization and discrimination.

SuperFashion was evaluated on three major fashion datasets: FashionAI, DARN, and DeepFashion. It achieved relative mean average precision (MAP) improvements of 1.84%, 9.27%, and 9.35% respectively over previous state-of-the-art methods. The paper was presented at the ACM Web Conference 2026 (WWW '26) and demonstrates a new paradigm for fine-grained image retrieval. By moving from rigid patches to flexible superpixel tokens, SuperFashion offers a practical solution for web-scale fashion image search where users need to find items based on specific attributes like collar type or sleeve length.