LocateAnything

LocateAnything is a unified generative grounding and detection framework from Nvidia Research. It uses Parallel Box Decoding (PBD), which treats each bounding box or point as one atomic unit and predicts the full coordinate set (x1, y1, x2, y2) in a single step. This replaces token-by-token coordinate generation, which the authors say mismatches box geometry and limits inference speed.\n\nThe architecture combines a Moon-ViT vision encoder and a Qwen2.5 language decoder, joined by an MLP projector, and processes images at native resolution. It offers three inference modes: Fast (parallel prediction), Slow (autoregressive coordinate decoding for stability), and Hybrid (fast by default, falling back to slow when format irregularity or spatial ambiguity is detected). On a single H100 in Hybrid Mode, the project page reports 12.7 boxes per second, versus 1.1 for Qwen3-VL and 5.0 for Rex-Omni.\n\nIt is trained on LocateAnything-Data, which covers about 138M queries and 785M boxes over 12M images. The data spans general object detection, GUI element grounding, referring comprehension, OCR text localization, layout grounding and point localization. Creator coverage describes a 3B-parameter model with released code and training. Intended uses include dataset labeling, document and GUI understanding, and latency-constrained settings such as robotics. The license and exact release links were not given in the provided material.