Rethinking Composed Image Retrieval Evaluation: A Fine-Grained Benchmark from Image Editing

TL;DR

This paper introduces EDIR, a fine-grained benchmark for Composed Image Retrieval using image editing to generate diverse queries, revealing significant performance gaps in current models and exposing limitations of existing benchmarks.

Contribution

We propose EDIR, a novel, comprehensive CIR benchmark created via image editing, to evaluate model performance across diverse, fine-grained categories and identify current limitations.

Findings

State-of-the-art models struggle across all subcategories

Existing benchmarks have modality biases and limited coverage

In-domain training improves performance on solvable categories

Abstract

Composed Image Retrieval (CIR) is a pivotal and complex task in multimodal understanding. Current CIR benchmarks typically feature limited query categories and fail to capture the diverse requirements of real-world scenarios. To bridge this evaluation gap, we leverage image editing to achieve precise control over modification types and content, enabling a pipeline for synthesizing queries across a broad spectrum of categories. Using this pipeline, we construct EDIR, a novel fine-grained CIR benchmark. EDIR encompasses 5,000 high-quality queries structured across five main categories and fifteen subcategories. Our comprehensive evaluation of 13 multimodal embedding models reveals a significant capability gap; even state-of-the-art models (e.g., RzenEmbed and GME) struggle to perform consistently across all subcategories, highlighting the rigorous nature of our benchmark. Through comparative analysis, we further uncover inherent limitations in existing benchmarks, such as modality biases and insufficient categorical coverage. Furthermore, an in-domain training experiment demonstrates the feasibility of our benchmark. This experiment clarifies the task challenges by distinguishing between categories that are solvable with targeted data and those that expose intrinsic limitations of current model architectures.