Temporal Calibration measures whether a generated video preserves a stable physical clock. When a file declares a playback frame rate, does the motion visible in the frames imply the same physical-time progression? WorldAtlas applies this as an output-only audit in the physics dimension using the wrapped Pulse-of-Motion / Visual Chronometer predictor [Gao et al., 2026].
This is not the same as optical flow, motion magnitude, or temporal smoothness. A rollout can look smooth frame to frame and still be temporally wrong: falls may drift as if gravity were slowed, or fast periodic motion may appear unnaturally sluggish. Visual Chronometer calls this failure mode chronometric hallucination—visually coherent dynamics with an ambiguous, unstable, or uncontrollable physical time scale.
The predictor estimates Physical FPS (PhyFPS) from overlapping temporal windows instead of trusting container metadata alone. WorldAtlas compares the mean inferred rate with the nominal playback FPS fmeta (OpenCV container FPS, or manifest fallback) and reports three official metrics: physical_fps_error, physical_fps_pct_error, and physical_fps_intra_video_cv. The first two capture global tempo mismatch; the third captures rhythm drift inside a single clip.
For each prediction, the runtime extracts overlapping windows with defaults clip_length = 30, stride = 4, and resolution = 216. Each window yields a segment estimate f^k. The mean inferred physical FPS is
fˉphy=K1k=1∑Kf^k(1)
Absolute and relative calibration errors are
Eabs=fˉphy−fmeta(2)Epct=fmetaEabs(3)
Within-video instability is the coefficient of variation over segment predictions:
Cintra=mean(f^1,…,f^K)std(f^1,…,f^K)(4)
WorldAtlas maps physical_fps_error, physical_fps_pct_error, and physical_fps_intra_video_cv to [0,1] with fixed reverse scaling (physical_fps_error: [0,30]; the other two: [0,1]). Higher normalized values mean better temporal calibration. When the side runtime, checkpoint, or FPS metadata is unavailable, the metric family is reported as not_applicable.
All examples come from image_dynamic, with three pairs each for absolute error, relative error, and within-video variation. The videos move visibly; each pair compares only the named metric.
Audit manifest: docs/assets/metric_examples/curated_same_case_manifest.json.
Both first frames passed geometric matching against the same input. Higher and lower compare physical_fps_error only; scores come from formal reports on the original predictions.
Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。
A freight train rounds a curved mountain track, bringing its locomotive and cargo cars across the frame.
source: curator observation
Both first frames passed geometric matching against the same input. Higher and lower compare physical_fps_error only; scores come from formal reports on the original predictions.
Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。
A passenger boat travels along a canal in a stylized town.
curator observation
Both first frames passed geometric matching against the same input. Higher and lower compare physical_fps_error only; scores come from formal reports on the original predictions.
Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。
Both first frames passed geometric matching against the same input. Higher and lower compare physical_fps_pct_error only; scores come from formal reports on the original predictions.
Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。
A bird repeatedly spreads and beats its wings above the water.
source: curator observation
Both first frames passed geometric matching against the same input. Higher and lower compare physical_fps_pct_error only; scores come from formal reports on the original predictions.
Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。
Water falls continuously into a pool below the rocky cliff.
source: curator observation
Both first frames passed geometric matching against the same input. Higher and lower compare physical_fps_pct_error only; scores come from formal reports on the original predictions.
Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。
Both first frames passed geometric matching against the same input. Higher and lower compare physical_fps_intra_video_cv only; scores come from formal reports on the original predictions.
Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。
Both first frames passed geometric matching against the same input. Higher and lower compare physical_fps_intra_video_cv only; scores come from formal reports on the original predictions.
Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。
A penguin swims through blue water, paddling its flippers as it changes direction.
source: curator observation
Both first frames passed geometric matching against the same input. Higher and lower compare physical_fps_intra_video_cv only; scores come from formal reports on the original predictions.
Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。
