Enhancing Representation Inversion and Alignment for Zero-Shot Composed Image Retrieval

Composed Image Retrieval (CIR) retrieves a target image by combining a reference image with a textual modification. Existing supervised methods have demonstrated strong performance; however, their dependence on costly, manually annotated triplets limits their scalability and generalizability across diverse applications. Zero-Shot Composed Image Retrieval (ZS-CIR) addresses this by mapping visual inputs to textual representations, enabling annotation-free composed image retrieval. However, accurately capturing both fine-grained details and high-level semantics remains a key challenge. To address this, we introduce an enhanced visual representation inversion and alignment model designed to bridge the gap between visual and textual modalities. Specifically, we propose a Multi-Level Visual Prompting (MLVP) module that integrates local image information into global representations through both instance-level and feature-level prompting, ensuring a more comprehensive capture of fundamental visual content. We further introduce a Dynamic Alignment (DA) module that complements contrastive learning, ensuring semantic flow is effectively transferred and aligned across different modalities without adding extra trainable parameters, thereby narrowing the semantic gap between the visual and textual representations. Extensive experiments across multiple datasets validate the superiority of our method over existing state-of-the-art approaches, demonstrating robust generalization and adaptability for zero-shot composed image retrieval.