Abstract
<title>Abstract</title> <p> <bold>Objectives</bold> : Composed image retrieval (CIR) identifies a target image by combining a reference image with a textual modification, enabling fine-grained visual search such as preserving an object while changing its attributes. Existing methods often rely on manually annotated triplets or unconstrained generated descriptions, limiting adaptation to target domains where labelled composed queries are scarce. <bold>Methods</bold> : This paper investigates zero-shot target-domain adaptation for CIR through structured semantic supervision. The proposed framework uses an offline vision-language and language-model agent to convert unlabeled images into structured semantic anchors containing global descriptions, attributes, and entities. These anchors replace free-form caption supervision with a low-entropy, field-constrained interface for edit-related semantics. The framework then infers target-side anchors from reference anchors and modification texts, and learns a query-slot-driven multi-vector representation that separates subject preservation from attribute variation. A multi-view masked learning strategy aligns gallery indexing, attribute-aware supervision, and online query modeling under different information-visibility conditions. <bold>Results</bold> : Experiments on FashionIQ and CIRR show that the approach improves zero-shot retrieval performance in fine-grained fashion scenarios while maintaining competitive open-domain retrieval ability. On FashionIQ, it achieves Avg R@10 = 31.81 and Avg R@50 = 53.07, outperforming the strongest compared zero-shot baseline by 2.42 and 2.19 percentage points, respectively. On CIRR, it obtains R@1 = 24.20, R@5 = 55.18, R@10 = 66.71, and R@50 = 90.30. <bold>Conclusion</bold> : The results suggest that structured semantic supervision offers a controllable route for adapting CIR models without manually annotated triplets, while revealing remaining challenges in relational reasoning, viewpoint change, and reproducible generated supervision. </p>