News

StrideWorld Benchmark Evaluates Scene Rigidity in Interactive Video Models

StrideWorld introduces a benchmark to test scene rigidity and action consistency in interactive video world models, highlighting key performance gaps.

StrideWorld Benchmark Evaluates Scene Rigidity in Interactive Video Models

Researchers have introduced StrideWorld, a new benchmark designed to evaluate scene rigidity and action consistency in interactive video world models. The study, led by Chuang Wang and collaborators, was published on November 18, 2026, on arXiv and is accompanied by an open-source implementation available on GitHub.

Addressing Scene Deformation and Action Inconsistency

Interactive video world models enable users to explore generated scenes through commands, requiring the preservation of 3D scene structure and stable action effects. Existing benchmarks often use references derived from the video under test, which can mask errors such as scene deformation and inconsistent action responses. StrideWorld addresses these limitations by using fixed references established before motion generation. Specifically, it evaluates scene rigidity against the geometry of the initial view and action consistency based on the stride of each command, measured during the model's initial steps.

Key Findings and Model Performance

StrideWorld tested 15 open models across 140 start views, generating 54,600 videos. The results revealed significant performance gaps:

  • All but one model exhibited scene deformation during camera movement, with offsets reaching up to 7.3 times those of a rigid rendering.
  • Most models maintained a steady pace for single actions, but several experienced mid-route stalls during composed routes.
  • Only two models that scaled commands by estimated scene depth achieved accurate route predictions, aligning closely with their own stride-based estimates.

Fine-tuning two models on a small set of rendered clips effectively reduced mid-route stalls, demonstrating the potential for targeted improvements.

Open-Source Availability

The StrideWorld benchmark, along with its code, data, and leaderboard, is publicly accessible on GitHub. The dataset includes 3,640 requests per model, covering 104 rendered scenes and 36 ETH3D photographs. Metrics such as rigidity offsets, pace scores, and composition errors are provided to facilitate comprehensive evaluations.

StrideWorld represents a significant step forward in benchmarking interactive video world models, offering a robust framework to identify and address critical shortcomings in scene and action consistency.

Sign in to save this article Free account: build your research collection and return to read it anytime.