WBench - Consistency: leaderboard

The consistency score of WBench (2026), a multi-turn benchmark for interactive video world models: text-driven, camera-controlled and action-conditioned video world models extend a scene turn by turn under their own controls. The score is the mean of eight background, spatial, segment, perspective, subject, geometric and photometric consistency metrics, on a 0-100 scale; higher is better.

Metric: Score (0-100): the mean of eight background, spatial, segment, perspective, subject, geometric and photometric consistency metrics, on the 158 navigation cases, multi-turn generation, each model under its own text, camera or action control; higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 20 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1LingBot-World89.9
2InSpatio-World88.4
3HY-Video 1.587.4
4LongCat-Video87.1
5HY-World 1.586.9
6Cosmos 2.586.5
7Fantasy-World86.4
8Happy Oyster84.3
9Kling 3.083.7
10Kairos 3.082.6
11Genie 382.6
12Wan2.781.6
13Seedance 1.581.3
14YUME 1.580.1
15Infinite-World79.9

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