New Framework Improves Zero-Shot Composed Image Retrieval.

Composed image retrieval, which involves finding an image based on a reference image and a textual modification, is challenging in a training-free, zero-shot setting. Existing methods typically generate a single, unified textual query, often leading to semantic distortions or omissions because preserving reference attributes and integrating textual modifications can conflict. This research introduces PEC-CIR, a novel training-free framework designed to make zero-shot composed image retrieval more robust. PEC-CIR reframes query construction as a multi-stage reasoning process, moving beyond single-pass generation. It employs a Planner-Executor-Critic architecture: the Planner identifies explicit constraints, the Executor generates multiple potential target descriptions, and the Critic then evaluates these candidates for compliance with the identified constraints. By explicitly evaluating candidate queries before retrieval, PEC-CIR significantly reduces the propagation of generative errors, leading to improved retrieval precision and stability. This strategic planning and self-criticism approach offers a more reliable way to integrate visual and textual information for complex image search tasks.