Event
GroundAnything Visual Grounding Model Paper Released
1 reports · 1 sources
Qize Yu et al. released a paper on arXiv proposing GroundAnything, a 4-billion-parameter visual grounding foundation model that combines parallel decoding with precise object localization via block-wise denoising, and whose training integrates pretraining on public datasets, direct conversion from autoregressive to diffusion, supervised fine-tuning, and GRPO-based reinforcement post-training. The paper claims that its autoregressive variant GroundAnything-VLM achieves a new overall best result among models of the same scale on 30 grounding benchmarks with 72.42%; the optional self-speculative mode achieves a 4.51x speedup over the autoregressive version at the cost of a 0.74 percentage point drop in COCO F1mIoU; when using entropy-guided decoding, its average score is 61.75%, exceeding the 53.32% of the fast MTP-based LocateAnything.
Report timeline
- GroundAnything: Balancing Parallel Decoding and Precise Visual Grounding with Blockwise Denoising
GroundAnything is a 4-billion-parameter visual grounding foundation model that combines parallel decoding with precise object localization through blockwise denoising. Its training integrates pretraining on public datasets, direct conversion from autoregressive to diffusion, supervised fine-tuning, and GRPO-based reinforcement post-training.
arXiv cs.RO ·