Apppublic

nielsr/locateanything-demo

sourceHugging Faceotherupdated 3mo agoView on Hugging Face

0likes

App README

LocateAnything 3B — Interactive Visual Grounding Space

This Space hosts [nvidia/LocateAnything-3B](https://huggingface.co/nvidia/LocateAnything-3B), the generalist vision-language grounding model from *LocateAnything: Fast and High-Quality Vision-Language Grounding with Parallel Box Decoding*.

LocateAnything predicts whole bounding boxes (or points) as atomic blocks in a single forward pass via Parallel Box Decoding (PBD / "fast" mode), while still supporting token-by-token autoregressive decoding ("slow" mode) and an adaptive "hybrid" mode that falls back to slow decoding when the fast output looks ambiguous.

What you can do here

Upload an image and pick a task:

Task	Output	Example query
Detect (multi-category)	boxes	`person, car, bus`
Ground single	1 box	`a red umbrella`
Ground multi	boxes	`people wearing red shirts`
Text detection	boxes	(no query — all scene text)
Point	points	`the traffic light`
GUI grounding	box or point	`the search button`

Pick a decoding mode — fast (MTP), slow (NTP), or hybrid — and the predicted boxes/points are drawn directly on the image, with raw model output and timing stats shown below.

Requires a GPU Space. The 3B model runs in bf16 on a single A10G (22 GB).