Composed Image Retrieval With No Vision Model: 44% R@1 Zero-Shot
Composed image retrieval ('this jacket, but in red, with a hood') done with images represented purely as text: a caption plus an attribute checklist at ingest, an LLM to merge the modification, plain text retrieval, then an LLM reranker — 44.04% R@1 on CIRR zero-shot.
Where this fits in Mixpeek
Composed queries — an image plus a modification — are a core multimodal retrieval pattern. Representing images as rich text (caption + attribute checklist) and reranking with an LLM is one way to serve them without a specialized vision model, and it slots into Mixpeek's extract-then-retrieve stages: derive structured attributes at ingest, then compose and rerank at query time.
About this research
Composed image retrieval handles queries like 'this jacket, but in red, with a hood.' A University of Tokyo team does it with images represented purely as text: a caption plus an explicit attribute checklist at ingest, an LLM to merge the modification, plain text retrieval, attribute scoring, and an LLM reranker to verify the change happened. Zero-shot: 44.04% R@1 on CIRR, +8.79 over prior methods. Paper: arxiv.org/abs/2607.12621
Frequently asked questions
Put the research to work
Mixpeek turns video, images, audio, and documents in your object storage into searchable, timestamped results through one API — the retrieval stack these papers describe.
Search your own data, free
