Learning to Compose: Revisiting Proxy Task Design for Zero-Shot Composed Image Retrieval

This paper introduces FoCo for zero-shot composed image retrieval (ZS-CIR), where a reference image and a textual modification specify the target image. Instead of relying on fixed composition rules such as pseudo-word injection or linear feature arithmetic, FoCo learns composition through two coordinated proxy tasks: text-anchored visual aggregation to focus on modification-relevant content, and context-conditioned semantic completion to form the target representation. Both tasks are jointly trained with a cross-instance contrastive objective intended to encourage semantic diversity and reduce shortcut solutions. The authors report state-of-the-art results and improved generalization across four ZS-CIR benchmarks, although the abstract does not provide benchmark names or numerical gains.