Motion Magnitude

Motion Magnitude quantifies how much internal motion a generated video contains, independent of motion correctness. It detects conservative generation—visually stable but nearly static outputs—without evaluating direction, semantics, or object behavior.

For adjacent frames $(I_t, I_{t+1})$, SEA-RAFT yields flow magnitude $F_t(x,y) = \sqrt{u_t^2 + v_t^2}$. The frame-pair statistic is the pixel median; the raw score averages over pairs:

$$S_{\mathrm{mag}}^{\mathrm{raw}} = \frac{1}{T-1}\sum_{t=0}^{T-2} \mathrm{median}_{x,y}\, F_t(x,y) \tag{1}$$

WorldAtlas z-score calibrates with $\mu = 3.2425$, $\sigma = 3.4505$, clip $[-0.8638, 2.7498]$, mapping to $[0.25, 0.75]$. Higher normalized scores indicate stronger visible motion. The metric is official when computation succeeds.

For video evaluation, all models use the same aspect-preserving input cap of 1920 pixels on the longest side, without upscaling smaller videos. SEA-RAFT’s existing half-scale inference then uses at most 960 pixels on its longest side. All adjacent frame pairs remain included. Before computing the magnitude and pixel median, flow components are converted back to source-video pixel units: $u=u_e W_s/W_e$, $v=v_e H_s/H_e$. Reports record both resolutions and both scale factors. This preserves the displacement units, but inference at a different resolution is not numerically identical to uncapped inference; affected old scores must be recomputed, not mixed into the same aggregate.

Verified Same-Input Examples

Audit manifest: docs/assets/metric_examples/curated_same_case_manifest.json.

Input image

Input verified · same benchmark case · gap 0.6749

image_dynamic_photorealistic_articulated_0004

Suite: image_dynamicScene: photorealistic / articulated

Benchmark prompt target

Camera fixed. A white horse gallops through a snowy field, legs extending and gathering as its hooves kick up loose snow. Its mane and tail flow while the snow-covered trees remain still.

prompt source: prompt_target

Both first frames passed geometric matching against the displayed input; labels compare this metric only.

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。

Higher metric scoreScore 0.9149

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。

Lower metric scoreScore 0.2400

Input image

Input verified · same benchmark case · gap 0.7604

image_dynamic_photorealistic_articulated_0068

Suite: image_dynamicScene: photorealistic / articulated

Benchmark prompt target

Camera fixed. The dog runs quickly across the grassy field, moving forward with alternating legs, open mouth, and perked ears while the fence and trees stay in place.

prompt source: prompt_target

Both first frames passed geometric matching against the displayed input; labels compare this metric only.

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。

Higher metric scoreScore 1.0000

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。

Lower metric scoreScore 0.2396

Input image

Input verified · same benchmark case · gap 0.5764

image_dynamic_photorealistic_articulated_0023

Suite: image_dynamicScene: photorealistic / articulated

Benchmark prompt target

Camera fixed. A large herd of horses gallops across the dry landscape, hooves kicking dust while manes and tails stream backward. The two distant observers and low hills remain stable in the background.

prompt source: prompt_target

Both first frames passed geometric matching against the displayed input; labels compare this metric only.

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。

Higher metric scoreScore 0.8167

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。

Lower metric scoreScore 0.2403

References

SEA-RAFT: Simple, Efficient, Accurate RAFT for Optical Flow

Wang et al., ECCV 2024.