NVIDIA's LocateAnything detects objects in photos, UIs, and text at lightning speed

NVIDIA's research team has introduced LocateAnything, a vision-language grounding model that rethinks bounding box prediction for high-speed, high-accuracy object detection. Unlike traditional models that process images sequentially, LocateAnything uses parallel box decoding to detect multiple objects simultaneously, dramatically reducing latency. It is trained on a diverse dataset including photos, application screenshots, documents, and UI elements, giving it broad applicability. The 3B parameter open-source model is available on HuggingFace and achieves state-of-the-art results on fine-grained detection tasks.

In benchmarks, LocateAnything outperforms Qwen3-VL and Rex-Omni on tasks like recognizing individual windows in a building or each piece of wood in a pile, as well as character recognition accuracy. NVIDIA provides a demo on HuggingFace where users can upload an image and specify object names to see real-time detection. Potential applications include robotics (e.g., picking specific objects), automated PC operation (e.g., clicking UI elements), and document analysis. The model's speed and precision make it ideal for AI agents that need to 'see' and act in real time.