🎯 LocateAnything-3B

Fast & high-quality vision-language grounding with Parallel Box Decoding (PBD) — detect, ground, OCR-locate, and point at any object in an image using natural language. Accepted at ECCV 2026.

Live demo

The real nvidia/LocateAnything-3B running on NVIDIA's ZeroGPU — upload an image (or pick an example below), type what you want to find, and hit Locate. Try the hybrid / fast / slow decoding modes to feel the speed difference.

LocateAnything-3B · official live demo open ↗
What it can do

One generalist model, five localization superpowers. Click “try it” to jump to the demo with a suggested prompt.

Open-vocabulary detection

Dense object detection

dense object detection example

Detect every instance of any category, even in cluttered scenes — no fixed class vocabulary.

person, car, dog, traffic light…
▶ try it
Referring comprehension

Phrase grounding

phrase grounding example

Localize everything that matches a natural-language description, however specific.

“the person in a red shirt”
▶ try it
Scene-text / OCR

Text localization

OCR and layout example

Find every piece of visible text in an image — signs, screens, documents, receipts.

“Detect all the text in box format”
▶ try it
GUI grounding

UI element grounding

GUI grounding example

Describe a UI element and get its exact bounding region — for computer-use agents.

“the search box”
▶ try it
Pointing

Point-based localization

pointing example

Return a precise point instead of a box — for robotics, 3D projection, and fine-grained spatial reasoning.

“the object being pointed at”
▶ try it
Long-tail detection

Open-set & rare objects

LVIS long-tail detection example

Trained on 12M images / 138M queries / 785M boxes — rare and long-tail categories included.

person, surfboard, boat, umbrella
▶ try it
Decoding modes

Parallel Box Decoding predicts each bounding box as one atomic unit — up to 12.7 boxes/s on an H100 (≈10× Qwen3-VL, 2.5× Rex-Omni), with automatic fallback for robustness.

ModeHow it decodesSpeedBest for
hybrid (default) Parallel Box Decoding first, auto-fallback to autoregressive on uncertain boxes ⚡⚡⚡ balanced Everyday use — best accuracy/speed trade-off
fast Pure parallel (MTP) decoding, never falls back ⚡⚡⚡⚡ fastest Simple scenes, on-device / real-time perception
slow Pure autoregressive token-by-token ⚡ most robust High-precision labeling, tricky spatial layouts
Run it yourself

The model is research-only under the NVIDIA non-commercial license. You need a GPU with ≥16 GB VRAM (or ≤8 GB with quantization). Full instructions in the official repo.

# 1. Clone and install
git clone https://github.com/NVlabs/Eagle.git eagle
cd eagle/Embodied
pip install -e .

# 2. Ground anything (HF model weights: nvidia/LocateAnything-3B)
python -c "
from locateanything_worker import LocateAnythingWorker
from PIL import Image

worker = LocateAnythingWorker('nvidia/LocateAnything-3B')
img = Image.open('your_image.jpg').convert('RGB')

result = worker.detect(img, ['person', 'car'])        # detection
result = worker.ground_multi(img, 'people in red')    # grounding
result = worker.detect_text(img)                      # OCR
result = worker.point(img, 'the traffic light')       # pointing
print(result['answer'])
"