GroundAnything claims 4.51× grounding speedup via diffusion

GroundAnything, a 4B-parameter visual-grounding foundation model, posts a 4.51× decoding speedup over its own autoregressive counterpart with a 0.74 percentage-point drop in COCO F1mIoU, according to a paper on arXiv.

Grounding, which maps natural-language phrases to specific image regions, is usually serialized token by token, which bakes in sequential latency. The paper reframes the task as "visual evidence extraction" and argues that spatial dependencies "do not imply an intrinsic left-to-right generation order." That reasoning opens the door to bidirectional diffusion, which lets spatial hypotheses "emerge in parallel and be jointly refined through iterative denoising."

The authors train the 4B model through grounding pretraining, AR-to-diffusion conversion with joint AR and diffusion objectives, supervised fine-tuning, and GRPO-based reinforcement post-training. Across 30 grounding benchmarks, their autoregressive variant GroundAnything-VLM averages 72.42%, which they call "a new overall state of the art among similarly sized models" and competitive with GPT-6 Astra at 71.35%. With entropy-guided decoding, GroundAnything averages 61.75% versus 53.32% for the MTP-based LocateAnything.

The 4.51× figure is a self-speculative mode measured against the paper's own AR baseline, not a wall-clock benchmark against production rivals. The abstract reports that "progressive inference optimizations translate parallel decoding into practical speedups" but publishes no absolute latency numbers.