Long-Sequence Metrics

WorldAtlas bundles a native Long-Sequence backend to extract long-horizon motion and quality signals from full prediction videos. These metrics do not only inspect single frames or adjacent frame pairs. They split a generated video into short clips, then analyze longer-range motion continuity and visual quality.

The main question is whether a generated video still preserves motion continuity and image quality over a longer horizon.

These metrics are computed prediction-only. They do not use the benchmark reference rollout or annotations. They are best used to explain long-video degradation modes such as motion collapse, abrupt temporal artifacts, or quality decay over time.

Covered Metrics

Metric name	Wrapped Long-Sequence dimension	Role in WorldAtlas
long_sequence_motion_smoothness	motion_smoothness	motion smoothness based on video-frame-interpolation error
long_sequence_dynamic_degree	dynamic_degree	optical-flow-based probe for whether enough motion occurs
long_sequence_aesthetic_quality	aesthetic_quality	no-reference aesthetic quality
long_sequence_imaging_quality	imaging_quality	MUSIQ-based image clarity and degradation quality

Principle

Short-range metrics can miss failures that appear only after a longer rollout. The Long-Sequence backend first splits the full video into short clips, then applies native long-video routines to motion and quality dimensions.

motion_smoothness uses AMT-S to interpolate middle frames and compares the interpolation with real middle frames. dynamic_degree uses RAFT flow strength to detect whether motion is sufficiently present. aesthetic_quality and imaging_quality use no-reference quality models over clip frames.

Detailed Algorithm

1. WorldAtlas reads the full prediction video and prepares a Long-Sequence input workspace.
2. worldatlas.benchmark.long_video splits the video into split_clip/ inputs and builds the minimal manifest required by each dimension.
3. A dispatcher routes each requested dimension to the corresponding implementation:
   • motion_smoothness: AMT-S interpolation, comparing predicted middle frames with real middle frames.
   • dynamic_degree: RAFT optical flow, counting whether high-motion transitions exceed the threshold.
   • aesthetic_quality: no-reference aesthetic scorer over clip frames.
   • imaging_quality: MUSIQ over clip frames.
4. Each WorldAtlas metric reads only its own wrapped dimension.
5. The wrapper records runtime metadata, including device, source root, clip count, per-video scores, and dimension-specific details.
6. Each metric is independently normalized according to the [0, 1] range in config/benchmark.yaml; all four metrics are higher-is-better.

Key Formulas

For motion_smoothness, AMT-S interpolates middle frames from evenly sampled frames and compares interpolated and real middle frames using mean absolute difference:

Motion smoothness $$s_{\mathrm{smooth}} = \operatorname{clamp} \left( \frac{255-\operatorname{mean}(|I^{real}_{mid}-I^{interp}_{mid}|)}{255}, 0, 1 \right)$$

For dynamic_degree, RAFT estimates optical flow between adjacent sampled frames and uses the mean magnitude of the top 5% highest-motion pixels as the transition motion score. If enough transitions exceed the motion threshold, the clip is marked dynamic and outputs 1.0; otherwise it outputs 0.0.

For aesthetic_quality and imaging_quality, the backend scores frames in each clip with no-reference quality models, then averages over clips, videos, and the sample batch.

Same-input 60-second examples now live on the dedicated pages: motion smoothness, dynamic degree, aesthetic quality, and imaging quality.