Paper 2601.08828
Motion Attribution for Video Generation
- Published
- Jan 2026
- Research lab
- NVIDIA
- Citations
- 2
- GitHub
- Not linked
01 In brief
Summary
The paper introduces Motive, a gradient-based data attribution framework for video generation models that isolates motion from static appearance.
It uses motion-weighted loss masks, computed via AllTracker optical flow, to focus influence scores on dynamic regions.
Motive scales via single-timestep estimation, common randomness, frame-length normalization, and Fastfood projections.
Experiments on Wan2.1-T2V-1.3B and Wan2.2-TI2V-5B show that fine-tuning with top-10% Motive-selected data improves VBench dynamic degree (47.6% and 48.3%) over random selection (41.3% and 41.6%) and full fine-tuning (42.0% and 45.3%), while maintaining motion smoothness.
Human evaluation shows a 74.1% win rate over the base model.
Ablations confirm that projection dimension 512 preserves rankings (Spearman 74.7%), single timestep suffices, and frame-length normalization reduces spurious length correlations by 54%.
Motive is the first motion-specific attribution framework for video generative models, enabling targeted data curation for improved temporal dynamics.
02 From the paper
Abstract
Despite the rapid progress of video generation models, the role of data in influencing motion is poorly understood. We present Motive (MOTIon attribution for Video gEneration), a motion-centric, gradient-based data attribution framework that scales to modern, large, high-quality video datasets and models. We use this to study which fine-tuning clips improve or degrade temporal dynamics. Motive isolates temporal dynamics from static appearance via motion-weighted loss masks, yielding efficient and scalable motion-specific influence computation. On text-to-video models, Motive identifies clips that strongly affect motion and guides data curation that improves temporal consistency and physical plausibility. With Motive-selected high-influence data, our method improves both motion smoothness and dynamic degree on VBench, achieving a 74.1% human preference win rate compared with the pretrained base model. To our knowledge, this is the first framework to attribute motion rather than visual appearance in video generative models and to use it to curate fine-tuning data.