VISTA (Video Grounding): leaderboard

VISTA (2026) tests open-set spatio-temporal video grounding: from a trimmed video and a natural-language query, a vision-language model must box the queried subject in every frame, over about 12,000 video-query pairs from six re-annotated datasets. Scored by mean spatio-temporal IoU in percent, pooled over referral and freeform queries; higher is better.

Metric: Mean spatio-temporal IoU (m_vIoU, %): per-frame box IoU summed over the frames where predicted and true tubes overlap in time, divided by their temporal union, pooled over referral and freeform queries (R&F), over the 11,814 video-query pairs of VISTA (HCSTVG-v1 and v2, VidVRD, VidSTG, MeViS and RVOS videos re-annotated with an interaction taxonomy), zero-shot on sub-sampled frames; the model must localize the queried subject with a box in every frame; higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 9 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3 VL 8B63.96#917
2VISTA CogVLM Grounding (checkpoint unspecified)54.7
3InternVL2.5-8B49.73#2370
4VISTA Sphinx-v2 (checkpoint unspecified)45.82
5MiniGPT-v245.78
6Qwen-VL-Chat45.49#7860
7Shikra-7B31.21
8LLaVA-Grounding-7B27.11
9Ferret-7B20.53

Interactive version: theaggregate.ai/benchmark?slug=vista-video-grounding · How It Works · Data refreshed daily, snapshot 2026-11-21.