TIPSv2: Advancing Vision-Language Pretraining with Enhanced Patch-Text Alignment. A new family of image-text encoder models with strong dense patch-text alignment, evaluated across 9 tasks and 20 datasets.