Visual-Centric Region Decoupling for Zero-Shot Composed Image Retrieval
Abstract
Zero-shot composed image retrieval (ZS-CIR) aims to retrieve a target image from a large gallery that satisfies a user-specified modification intent, given only a reference image and a textual instruction, without any task-specific training. Existing methods suffer from a language-dominant bias: by converting the reference image into textual surrogates such as pseudo-tokens or generated captions, they often discard fine-grained visual details and fail to preserve consistency in unmodified regions, especially under local edits. To address this limitation, we propose VCRD (Visual-Centric Region Decoupling), a visual-centric ZS-CIR framework that explicitly decouples editable and preserved regions at the feature level. The editing intent is first parsed by a multimodal model and grounded to produce a spatial binary mask. With all encoders frozen, lightweight test-time optimization is applied exclusively to the editable visual features, while the preserved features remain unchanged. In parallel, a text-guided global visual query is optimized to balance target semantics with the reference image's global visual representation. To support fine-grained retrieval, we further devise an adaptive multi-path scoring mechanism that jointly integrates preserved-content consistency, editable-region similarity, and global visual alignment, with final matching performed entirely in the visual embedding space. Extensive experiments on three prominent ZS-CIR benchmarks show that VCRD achieves strong performance among zero-shot methods, with particularly notable gains on local attribute modifications and complex compositional scenarios while improving visual consistency in preserved regions.