In this paper, we propose a strong recipe for transferring image-text models to open-vocabulary object detection.