Abstract
Large-scale visual search systems can present a trade-off between search speed and accuracy. A technique to address this can include using a fast embedding search to generate potential candidates. Thumbnail images of a query subject and multiple candidates can then be programmatically arranged into a composite image, such as a collage. This collage may be sent with a textual prompt to a VLM for batch verification, for instance, in a single transaction. Such a collage-based approach can amortize the cost and latency of VLM analysis across multiple candidates, which can improve search accuracy by helping to filter false positives and making VLM use potentially more practical in large-scale applications.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
N/A, "Hybrid Visual Search with Collage-Based Batch Verification", Technical Disclosure Commons, (October 12, 2026)
https://www.tdcommons.org/dpubs_series/12095