LocateAnything
research.nvidia.com- Category
- AI Tools
- Rank
- No. 645Tools index
- Pricing
- Open Source
- Platform
- web
- Type
- TOOL
- Added
- Jul 2, 2026
About
LocateAnything is a fast, high-quality vision-language grounding framework that uses Parallel Box Decoding (PBD) to predict bounding boxes as atomic units in a single forward pass. It supports diverse localization tasks including document understanding, GUI grounding, dense object detection, and OCR localization. By decoding geometric elements in parallel rather than sequentially, it achieves up to 2.5× faster throughput while improving localization accuracy.
What it can do
Locate objects in an image using natural language queries
Image and a text description of a target object → Bounding box coordinates around the described object
Detect and localize UI elements in GUI screenshots
Screenshot of a graphical user interface and a text query → Bounding boxes identifying the specified GUI components
Extract and localize text regions in documents
Document image and OCR localization query → Bounding boxes around detected text regions or specific text content
Perform dense object detection across an entire image
Image with multiple objects → Bounding boxes for all detected object instances with labels
Ground language references to regions in complex documents
Document image and a natural language reference to a specific region or element → Bounding box pinpointing the referenced document region
Run high-throughput batch visual grounding
Multiple images paired with language queries → Bounding box predictions for all queries processed in parallel at up to 2.5× faster speed
Why it made the leaderboard
NVIDIA's grounding framework that decodes bounding boxes in parallel — each box an atomic unit in a single forward pass — instead of dribbling out coordinate tokens sequentially. Trained on 138M queries and 785M boxes, it covers GUI grounding, document layout, OCR localization, and dense detection in one model.
Tags
Media
Comments (0)
No comments yet
Indexed by a proprietary survey. Corrections welcome.