Vidi2

Vidi2 extends the original Vidi models into a full video reasoning system that handles temporal retrieval, spatio temporal grounding, and video QA in one framework. Given a text query, it can locate the relevant time segments and output bounding boxes for the described objects, which is useful for plot or character tracking, automatic multi view switching, and composition aware reframing and cropping. The team also introduces the VUE STG and VUE TR V2 benchmarks with long videos, noun phrase style queries, carefully annotated timestamps and boxes, and refined vIoU or tIoU metrics, and reports state of the art performance on these tasks compared with both proprietary and open models of similar scale.