# StrideWorld Reports Scene Deformation in 14 of 15 Video Models

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/61886
Published: 2026-11-20T08:38:56.000Z
Updated: 2026-11-20T08:38:56.000Z
Section: Technology

> The benchmark tested 15 open models across 140 starting views and generated 54,600 videos to measure scene rigidity and movement consistency.

StrideWorld evaluated how well interactive video models preserve scene geometry and follow movement commands, finding deformation in 14 of 15 open models tested across 54,600 generated videos.

The benchmark used 140 starting views to assess two capabilities: whether a model maintains the geometry of its initial view, and whether its response to movement commands remains consistent. The measured offsets reached up to 7.3 times those of a rigid rendering.

Most models maintained a steady pace when given a single movement command. Several, however, stalled partway through routes that combined multiple commands. Two models that scaled command size using estimated scene depth made accurate route predictions that aligned closely with their own stride-based estimates.

Fine-tuning two models on a small set of rendered clips reduced stalls during composed routes. StrideWorld’s dataset includes 104 rendered scenes and 36 ETH3D photographs, with measures for scene deformation, pace and route-composition errors.

The benchmark, its code, data and leaderboard are publicly accessible on GitHub. StrideWorld provides an evaluation approach for interactive video world models.


