NVIDIA and universities launch LocateAnything for fast, accurate object detection

NVIDIA, together with Hong Kong Polytechnic University and Nanjing University, has released LocateAnything, a model designed for fast and accurate object detection. It can quickly find objects in photos or screenshots and mark them with a box, mainly for robot sensing and computer automation.

The model uses parallel box decoding to predict bounding boxes in one step. It offers three modes: Fast Mode for speed, Slow Mode for accuracy, and Hybrid Mode that switches between them. The team also built a dataset with 12 million images and 138 million language queries covering various detection tasks.

On a single H100 GPU, Hybrid Mode processes 12.7 boxes per second, beating Qwen3-VL's 1.1 and Rex-Omni's 5.0. In high-accuracy tasks, it scores 31.1 on LVIS with IoU=0.95 and 60.3 average F1 on ScreenSpot-Pro, outperforming similar models.