LocateAnything: Universal Visual Grounding & Spatial Localization

Model Overview

LocateAnything is a high-speed, high-accuracy vision-language model developed by NVIDIA Learning and Perception Research (LPR) Lab (Shihao Wang, Shilong Liu, et al., accepted to ECCV 2026). It introduces Parallel Box Decoding (PBD) to predict bounding box coordinates and keypoints as atomic multi-token units in a single forward pass, achieving up to 10× higher throughput compared to traditional coordinate token models.

---

Key Features

• Parallel Box Decoding (PBD): Decodes bounding box coordinates atomically in a single pass rather than sequentially token-by-token.
• High Throughput & Speed: Delivers up to 10× higher decoding throughput (up to 12.7 boxes/sec on an NVIDIA H100 GPU) supported by custom la_flash CUDA kernels.
• Hybrid Inference Mechanism: Uses Fast Mode (PBD) by default with automatic fallback to Slow Mode for complex edge cases.
• LocateAnything-Data Corpus: Pretrained on 138+ million training samples and 785+ million bounding box annotations across natural scenes, GUI screenshots, and OCR.
• Universal Multi-Task Localization: Supports open-set object detection, GUI element grounding, scene text OCR localization, and visual prompt tuning via LoRA.

---

Verified Project Links

• Project Website: https://research.nvidia.com/labs/lpr/locateanything/
• arXiv Paper: https://arxiv.org/abs/2605.27365
• GitHub Repository: https://github.com/NVlabs/Eagle
• Hugging Face Model: https://huggingface.co/nvidia/LocateAnything-3B

---

Performance & Benchmarks

• COCO Object Detection: 54.7 F1@mIoU
• LVIS Long-Tailed Object Detection: 50.7 F1@mIoU
• RefCOCOg Referring Expression Grounding: 78.7 (val) / 76.7 (test-dev) / 77.6 (test-std)