SABER contains scenes with up to ten objects and a broad set of attribute types, with hard-negative candidates that re-pair the same objects and attributes.
Text-only finetuning improves attribute-object binding in vision-language models across all retrieval pairings.
The authors introduce SABER, a controlled synthetic benchmark covering all four image-text modality pairings, and show that text-only finetuning transfers binding gains to cross-modal and image-image retrieval.
Big Tech
Netanel Y. Tamir · Michael Green · Issar Tzachor · Or Shimshi · Dvir Samuel · Or Patashnik · +1 more
OriginAI · NVIDIA · Tel Aviv University
Research Digest··2 min read
The authors introduce SABER, a synthetic benchmark that varies attribute-object assignments while holding the object and attribute sets fixed, and evaluate retrieval across all four image-text modality pairings.
§