Top AI Product

LocateAnything is an NVIDIA vision-language model built around a single bet: stop generating bounding-box coordinates one token at a time. Visual grounding — pointing a model at “the red mug behind the laptop” and getting back exact pixel coordinates — has typically been autoregressive, emitting each coordinate number in sequence. That’s slow, and the speed cost scales with how many objects you’re locating.

## Parallel box decoding

The core technique is Parallel Box Decoding (PBD), a block-wise multi-token prediction scheme. Instead of generating coordinates serially, the model predicts complete bounding boxes and points as parallel structured units, which cuts decoding latency while preserving geometric consistency. The architecture jointly trains next-token prediction and multi-token prediction, so it keeps language-style reasoning while gaining parallel inference. Training runs in four stages, starting from captioning, VQA, and OCR adaptation before grounding-specific fine-tuning.

## Why it matters

Grounding is the bridge between “the model understands the image” and “the model can act on the image” — clicking a UI element, guiding a robot arm, annotating a scene. Those uses need low latency and many boxes per frame, exactly where autoregressive decoding struggles. Making grounding fast without sacrificing accuracy is what moves it from research demo toward real-time perception.