Skip to content

Motion Smoothness

Motion Smoothness measures temporal interpolatability of a generated video under a frame-interpolation prior. VFIMamba reconstructs discarded odd frames from neighboring even frames; low reconstruction error indicates smooth, continuous motion—not correct semantics or magnitude.

For window kk, I^2k+1=Φ(I2k,I2k+2)\hat{I}_{2k+1} = \Phi(I_{2k}, I_{2k+2}) is compared to ground-truth odd frame I2k+1I_{2k+1} via MSE, SSIM, and LPIPS. Raw components average over KK windows:

Emse=1K∑kMSE(I^2k+1,I2k+1),Sssim=1K∑kSSIM(I^2k+1,I2k+1),Elpips=1K∑kLPIPS(I^2k+1,I2k+1)(1)E_{\mathrm{mse}} = \frac{1}{K}\sum_k \mathrm{MSE}(\hat{I}_{2k+1}, I_{2k+1}),\quad S_{\mathrm{ssim}} = \frac{1}{K}\sum_k \mathrm{SSIM}(\hat{I}_{2k+1}, I_{2k+1}),\quad E_{\mathrm{lpips}} = \frac{1}{K}\sum_k \mathrm{LPIPS}(\hat{I}_{2k+1}, I_{2k+1}) \tag{1}

Decoded uint8 frames are converted to float32 before MSE subtraction and squaring, so negative pixel differences cannot wrap modulo 256.

Each component is normalized independently; the final score is their arithmetic mean. Higher SsmoothS_{\mathrm{smooth}} indicates smoother generated motion. The metric is official when computation succeeds and requires no reference video.

The configured MSE empirical bound predates v2_float32_mse. Corrected raw reports are versioned, but their normalized score remains provisional until that bound is recalibrated on the corrected benchmark corpus.

The higher-scoring side of each example moves visibly. Motion Magnitude only screened out nearly static videos; the displayed scores come from formal Motion Smoothness reports and retain the provisional normalization noted above. Audit manifest: docs/assets/metric_examples/curated_same_case_manifest.json.

Input image: image_dynamic_photorealistic_articulated_0136
Input image
Input verified · same benchmark case · gap 0.5793

image_dynamic_photorealistic_articulated_0136

Suite: image_dynamicScene: photorealistic / articulated
Visible motion in this comparison

A cheetah runs across the savannah; its legs and body change position throughout the clip.

curator observation

The higher-scoring output also moves visibly. Motion Magnitude was used only to screen for visible motion; scores shown are Motion Smoothness. Both first frames passed geometric matching against the same input.

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 0.8239

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 0.2446
Input image: image_dynamic_photorealistic_rigid_0049
Input image
Input verified · same benchmark case · gap 0.4492

image_dynamic_photorealistic_rigid_0049

Suite: image_dynamicScene: photorealistic / rigid
Visible motion in this comparison

Aircraft and the airport view move visibly across the frame.

curator observation

The higher-scoring output also moves visibly. Motion Magnitude was used only to screen for visible motion; scores shown are Motion Smoothness. Both first frames passed geometric matching against the same input.

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 0.8947

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 0.4455
Input image: image_dynamic_photorealistic_fluid_0069
Input image
Input verified · same benchmark case · gap 0.7375

image_dynamic_photorealistic_fluid_0069

Suite: image_dynamicScene: photorealistic / fluid
Visible motion in this comparison

The waterfall flows while the view moves closer to the falling water.

curator observation

The higher-scoring output also moves visibly. Motion Magnitude was used only to screen for visible motion; scores shown are Motion Smoothness. Both first frames passed geometric matching against the same input.

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 0.7818

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 0.0443