GroundAnything: Reconciling Parallel Decoding with Precise Visual Grounding at Flash Speed
GroundAnything turns visual grounding into parallel denoising, trading the usual token-by-token latency for faster localization.
The paper presents a 4B-parameter grounding model built around blockwise bidirectional diffusion. Its autoregressive variant reports 72.42% across 30 grounding benchmarks, above similarly sized models and competitive with GPT-6 Astra at 71.35%. The diffusion version averages 61.75%, ahead of the cited fast MTP-based LocateAnything result of 53.32%. An optional self-speculative mode is reported at 4.51x faster than the AR counterpart, with a 0.74-point COCO F1mIoU drop. Source: HF Daily Papers' note
The paper presents a 4B-parameter grounding model built around blockwise bidirectional diffusion. Its autoregressive variant reports 72.42% across 30 grounding benchmarks, above similarly sized models and competitive with GPT-6 Astra at 71.35%. The diffusion version averages 61.75%, ahead of the cited fast MTP-based LocateAnything result of 53.32%. An optional self-speculative mode is reported at 4.51x faster than the AR counterpart, with a 0.74-point COCO F1mIoU drop. Source: HF Daily Papers' note
score 6