Nvidia's LocateAnything spots objects at 10x speed

If you have ever tried using vision-language models to pinpoint objects in crowded scenes, you know the frustration. The model easily understands what you want, but marking exact bounding boxes feels like watching paint dry.

That bottleneck comes down to a fundamental design mismatch. Most current VLMs treat 2D bounding boxes as text coordinates, spitting them out token-by-token. Splitting a single box into four sequential numbers destroys geometric coherence and turns dense object detection into a crawling inference nightmare.

Nvidia researchers just introduced LocateAnything, a unified framework designed to fix this pipeline and spot objects from natural language prompts up to 10 times faster in dense scenarios.

Bounding boxes as atomic units

Instead of treating coordinates like words in a sentence, LocateAnything introduces Parallel Box Decoding (PBD). The idea is refreshingly direct: treat geometric elements like bounding boxes and points as atomic units decoded in a single step.

This single-step parallel decoding preserves the geometric relationship of box coordinates while letting modern hardware do what it does best: process tasks concurrently. Under the hood, the system couples Qwen2.5-3B with Moon-ViT, offering flexible execution via fast, slow, and hybrid inference modes.

Speed means little without precision, so the team also curated LocateAnything-Data, a massive dataset spanning over 138 million training samples. The result is a system that not only decodes dramatically faster but also scores higher localization accuracy on strict IoU benchmarks.

Why it matters

We think LocateAnything tackles one of the most annoying gaps between multimodal models and physical-world execution.

For robotics, automated dataset annotation, and industrial inspection, language models have historically been too slow or clumsy to localize hundreds of items in real time. Forcing a neural network to serialize 2D space into sequential 1D tokens was always an awkward workaround.

By packaging coordinates into atomic units and pairing a 3B-parameter base with a 138-million-sample dataset, Nvidia is clearing the runway for practical spatial AI. When vision models can accurately locate and crop any item from natural language prompts without a latency penalty, downstream autonomous agents get significantly closer to real-time viability.