Generalizable Attribute-Guided Prompting for Universal Cross-Domain Retrieval

Universal Cross-Domain Retrieval (UCDR) aims to retrieve semantically related images across unseen domains and categories, reflecting realistic open-world scenarios. While recent methods leverage vision-language models like CLIP, they typically align images with coarse class names and overlook fine-grained semantic attributes, leading to limited generalization under domain and semantic shifts. In this paper, we propose Generalizable Attribute-guided Prompting (GAP), a novel end-to-end framework that exploits fine-grained and transferable visual attributes to bridge the gap between seen and unseen domains and categories. First, we incorporate an attribute bank construction process that uses large language models to automatically generate descriptive visual attributes for each class, where class names are collected from multiple datasets, enabling rich and diverse semantic representations. Then, we introduce an attribute-guided learning process that injects attribute-level semantics into visual-text alignment via attribute-aware visual prompts, attribute-guided enhancer (AGE), attribute-level alignment loss, and lightweight attribute-aware adaptation (LAA), all jointly designed to improve generalization across diverse domains and categories. Extensive experiments across multiple benchmarks demonstrate that GAP achieves competitive or superior performance compared to existing methods across various UCDR settings. Our code is available here .