Breaking

NVIDIA unveils MOTIVE, a method to pinpoint training clips that boost video motion

July 7, 2026 at 15:28 EDT

Media Generation
Research & Papers
Foundation Models

NVIDIA Research has introduced MOTIVE, a method that identifies and re-weights which clips in a video model's training data actually improve motion, or temporal dynamics. The work received an Outstanding Paper Honorable Mention at ICML 2026.

ICML 2026 · NVIDIA Research — Spatial Intelligence Lab

Just 10% of the Data — and Better Motion Than Training on All of It

MOTIVE pinpoints which video clips actually teach a model how things move, then re-weights training toward them — beating full fine-tuning while using a fraction of the data. Awarded an Outstanding Paper Honorable Mention.

10%

of training data used — yet outperforms full fine-tuning

+8 pts

Dynamic Degree gain (39.6% → 47.6%) on Wan2.1-T2V-1.3B

74.1%

human win rate vs. the base model (13.6% losses)

Dynamic Degree on VBench — higher = more motion

Wan2.1-T2V-1.3B. Each block ≈ 5 percentage points.

39.6%

Base model

41.3%

Random 10%

42.0%

Full fine-tune

47.6%

MOTIVE 10%

How the attribution loop works

Mask motion

Motion-weighted loss masks split appearance from movement

→

Score clips

Gradient signals rank each clip's influence on motion

→

Re-weight & train

Top-K or majority-vote selection favors motion-rich data

Why it lands

Beats full fine-tuning using only 10% of data
Motion Smoothness held steady at 96.3%
Scales to billion-parameter video models
+6.3 pts Dynamic Degree on Wan2.2-TI2V-5B (42.0%→48.3%)

Caveats

Clips with strong static backgrounds score as negative influence
Results shift substantially with vs. without motion masking
Released as a research framework, not a product

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.