LocateAnything: Fast and High-Quality Vision-Language Grounding with Parallel Box Decoding

Researchers introduced LocateAnything, a unified grounding and detection framework that decodes boxes and points in parallel.

It uses Parallel Box Decoding instead of token-by-token generation, improving inference speed and throughput.

The team also released LocateAnything-Data, a dataset of more than 138 million samples for large-scale training.

Across benchmarks, the system delivers stronger high-IoU localization while keeping fast performance.