Skip to main content
Back to timeline
arXivSource publication:

StrideWorld fixes the evaluation reference before the motion: fourteen of fifteen open interactive video world models deform their scenes as the camera moves

Synopsis

Interactive video world models must keep a scene's 3D structure as the camera moves and give each command a stable effect, yet existing benchmarks mostly fit their references to the video under test, which can absorb gradual deformation and pace changes. StrideWorld fixes its references before the motion: scene rigidity is checked against the ground-truth geometry of the start view, and action consistency against the stride each model sets in its own first four steps. Across 54,600 videos from fifteen open models, all but one deform their scenes as the camera moves, most keep a steady pace under a single held action, and composed routes often stall or end away from the positions their own strides predict.

Source-provided article image: StrideWorld: Benchmarking Scene Rigidity and Action Consistency in Interactive Video World Model
arXiv

Interpretation

Evaluation references for interactive video world models are now fixed before the motion they judge, so scene rigidity and action consistency are no longer defined by the video under test. Earlier benchmarks either reconstructed the scene from generated frames with 3D foundation models and checked consistency against that reconstruction, or aligned the recovered camera trajectory to the commanded path with a fitted scale; both references are fitted to the video under test and can absorb gradual deformation and changes in distance per command. Across 54,600 videos from fifteen open models, every model exceeds a rigid rendering of the true scene along the same camera path, with ratios from 1.6 to 7.3, so tracking noise alone does not account for the excess.

Scene rigidity places 3D anchors at the ground-truth depth of the start view and checks whether visible surfaces stay in place as the camera moves. Anchors are split alternately into two sets, one used only to fit the camera pose and the other held out for testing, so the test points do not enter the fit. On UE5 renderings with known camera paths, recovered cameras deviate from the true ones by a median of 1.48 degrees in translation and 0.31 degrees in rotation; after blur, compression, flicker, or moderate wobble the median errors stay within 3.18 and 2.09 degrees, and only the strongest wobble raises them to 8.26 and 4.40 degrees.

Action consistency uses the stride a model sets in its own first four steps as the unit, predicts the camera pose at every later step, then tests whether a held action keeps a steady pace and whether multi-action routes add up. Control interfaces share no unit of motion, so calibrating on each model's own stride lets every model be compared only with itself, without fitting a route-wide scale or assuming opposite commands cancel. Ten raters blind to models and metrics judged randomly sampled videos for whether the camera follows commands, and the per-video score separates videos judged good with an AUC of 0.88, while the swap-pair score separates pairs judged to end at the same place with an AUC of 0.83.

Perspective

The evaluation covers static scenes and planar motion within surfaces visible from the start view, with pitch evaluated separately; long-horizon exploration and revisits are outside its scope. It is intended for developers and evaluators of interactive video world models, and code, data, and a leaderboard are public.

A stride cannot be measured reliably when a model moves too little in its first steps, which leaves AlayaWorld unranked; cameras are recovered from generated frames, so strong deformation also shifts the recovered camera, changing the near-prediction decision for some routes at the deformation level of generated videos while leaving the model ranking unchanged.