GroundAnything: Reconciling Parallel Decoding with Precise Visual Grounding at Flash Speed
GroundAnything, a 4B model, pairs parallel diffusion decoding with precise visual grounding and large inference speedups.
GroundAnything is a 4B grounding model that treats localization as parallel visual-evidence extraction via blockwise diffusion denoising rather than autoregressive token order. Training includes public and engine data, AR-to-diffusion conversion, supervised fine-tuning, and GRPO. Its autoregressive variant, GroundAnything-VLM, averages 72.42% on 30 benchmarks, competitive with GPT-6 Astra at 71.35%. Entropy-guided decoding reaches 61.75% versus 53.32% for LocateAnything, and an optional self-speculative mode is 4.51 times faster than the AR counterpart with a 0.74-point COCO F1mIoU drop.
- 4B model uses blockwise diffusion denoising instead of left-to-right spatial tokens.
- GroundAnything-VLM averages 72.42% on 30 benchmarks, near GPT-6 Astra at 71.35%.
- Entropy-guided decoding scores 61.75% versus 53.32% for LocateAnything.
- Optional self-speculation is 4.51x faster than autoregressive decoding with a 0.74-point COCO drop.
From the original article
1 min read at huggingface.co
Autoregressive (AR) grounding models serialize spatial predictions, introducing sequential latency and imposing a causal order on output tokens. We view grounding as visual evidence extraction: objects, locations, and spatial relations are jointly constrained by the image and query, yet their dependencies do not imply an intrinsic left-to-right generation order.
By Qize Yu, Lianrui Fan, Bowen Ping · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.