GroundAnything: Reconciling Parallel Decoding with Precise Visual Grounding at Flash Speed

arXiv · Qize Yu, Lianrui Fan, Bowen Ping, Xini Ding, Zetian Song, Junbo Niu, Kaixuan Wang, Tianxing Chen, Yue Chen, Minghua He, Yuran Wang, Jie Huang, Haojun Zhang, Min Chen, Hao Li, Wenxuan Song, Ruihai Wu, Xianming Liu, Shilong Liu, Shuchang Zhou, Ping Luo, Shiyu Huang · 2026-09-30

This item is no longer on the current shelf. Its page remains available, and editorial corrections may be made. See what is featured now.

GroundAnything studies distinct autoregressive and diffusion grounding variants, with measured accuracy/latency tradeoffs.

What it does

GroundAnything is a 4B grounding framework evaluated on 30 benchmarks. Its autoregressive GroundAnything-VLM variant reaches 72.42%, compared with GPT-6 Astra at 71.35%; this is not evidence that GPT-6 Astra is similarly sized. Entropy-guided diffusion reaches 61.75% versus LocateAnything at 53.32%. Optional self-speculation yields a 4.51-fold speedup over the AR counterpart with a 0.74-percentage-point drop in COCO F1mIoU.

Why it matters

The paper measures grounding latency and accuracy under its evaluated variants and inference configurations. These results motivate testing faster grounding in latency-sensitive applications; they do not certify robotics or AR production suitability.

How it applies

Compare the intended variant against the current grounding stack on the target hardware and dataset. Include accuracy, latency and the self-speculative tradeoff rather than treating a decoder swap as a guaranteed lossless improvement.

The short version

What if your visual grounding model could answer a query faster than you can blink?

My take

The useful result is a comparison of decoding modes under measured conditions. It supports a targeted grounding experiment, not a first-proof claim about every diffusion model.

How it connects

• Autoregressive and diffusion variants have different measured outcomes.
• Speculation should be evaluated with both latency and quality metrics.
• Downstream physical use needs its own closed-loop evaluation.

Bottom line

Test the relevant variant and report both latency and accuracy; self-speculation has a measured accuracy loss.

Takeaways

• The 72.42% result is for the autoregressive variant, not diffusion.
• Entropy-guided diffusion averages 61.75% versus LocateAnything at 53.32% in the reported comparison.
• Optional self-speculation gives 4.51-fold speedup with a 0.74-percentage-point COCO F1mIoU loss.
• Variant, benchmark and inference configuration must stay attached to every result.

Topics: visual grounding, parallel decoding, diffusion models, efficiency