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

Researchers introduce LocateAnything, a new vision-language model framework that uses Parallel Box Decoding to detect and localize objects simultaneously rather than sequentially, improving both inference speed and accuracy. The team curated a 138-million-sample dataset and demonstrated significant performance improvements across multiple benchmarks.

LocateAnything addresses a fundamental inefficiency in how current vision-language models perform visual grounding and object detection. Traditional approaches serialize 2D bounding boxes into sequential tokens, requiring independent decoding of each coordinate—a process that creates both computational bottlenecks and geometric inconsistencies. By treating bounding boxes as atomic units decoded in parallel, the framework preserves the structural relationships between geometric elements while dramatically reducing inference latency.

This advancement builds on years of research into unified vision-language models, which have struggled to balance speed with precision in localization tasks. The introduction of Parallel Box Decoding represents a meaningful architectural shift rather than incremental optimization. The team's complementary effort to build LocateAnything-Data with 138 million training samples reflects industry-wide recognition that large-scale, diverse datasets drive performance across computer vision tasks. This data-centric approach mirrors successful strategies in large language models.

The implications extend across multiple sectors relying on real-time object detection: autonomous systems, robotics, augmented reality, and industrial inspection all benefit from faster, more accurate localization. Higher-quality bounding boxes at higher-IoU thresholds directly improve downstream application reliability. For AI researchers and practitioners, this work demonstrates that algorithmic efficiency and training data scale are complementary forces rather than trade-offs.

The research establishes new benchmarks that competitors will likely target, potentially accelerating improvements in vision-language model efficiency. As vision-language systems become increasingly deployed in production environments, throughput gains measurably reduce infrastructure costs while accuracy improvements expand viable use cases.

• →Parallel Box Decoding replaces sequential token generation with simultaneous box decoding, reducing inference bottlenecks.
• →LocateAnything-Data containing 138 million samples substantially increases training diversity for visual localization tasks.
• →The framework achieves better high-IoU localization accuracy while improving throughput on diverse benchmarks.
• →Atomic unit decoding preserves geometric coherence within bounding boxes, improving consistency and reliability.
• →The approach addresses a core architectural limitation affecting real-time deployment of vision-language models.