Wang et al. introduce StrideWorld, a benchmark for interactive video world models that evaluates scene rigidity and action consistency using fixed references established before motion occurs. Unlike prior benchmarks that fit references to test videos—absorbing deformation and pace changes—StrideWorld measures scene rigidity against ground-truth geometry of the start view and action consistency against the stride, defined as the camera motion from a single command measured during the model's initial steps and then applied to subsequent steps. Validated on rendered videos with known camera trajectories, StrideWorld scores correlate with human judgments. Across 140 start views and multiple open models, all but one model deform scenes during camera movement, with offsets up to 7.3× those of rigid rendering along the same path. Most models maintain steady pace under repeated single actions, but along composed routes, several frequently stall and most finish far from positions predicted by their own strides. Only the two models that scale commands by estimated scene depth consistently reach predicted endpoints. Fine-tuning two models on a few rendered clips eliminates most mid-route stalls. Code, data, and a leaderboard are publicly released.
Read original
huggingface/daily-papers