MOTIVE: Teaching AI Video Models to Move Better
AI-generated videos often look great frame by frame—but the motion between frames can jitter, drift, or break physics. A new study introduces Motive, a motion-first way to find which training clips actually drive how a model moves.
Instead of focusing on static appearance, Motive uses gradient-based influence signals weighted by motion to pinpoint clips that help—or hurt—temporal dynamics in text-to-video systems.
- Curates better fine-tuning data by selecting high-impact motion clips
- Improves temporal consistency and physical plausibility
- Scales to large, modern video datasets and models
With Motive-picked data, the team improved motion smoothness and dynamic degree, achieving a 74.1% human preference win rate on VBench versus the base model.
This is the first framework to attribute motion (not just looks) in video generators—paving the way for fewer jittery, more believable AI videos.
Paper: https://arxiv.org/abs/2601.08828v1
Paper: https://arxiv.org/abs/2601.08828v1
Register: https://www.AiFeta.com
AI VideoGeneration ComputerVision MachineLearning GenerativeAI Research VBench
