Details
- NVIDIA Research unveils MOTIVE, a motion-centric data attribution framework for training generative video models.
- The method identifies which training clips improve or degrade motion, rather than just visual appearance, in video diffusion models.
- MOTIVE uses scalable gradient-based attribution, motion-weighted loss masks, and influential subset selection to focus learning on clips with strong temporal dynamics.
- Applied to text-to-video models, MOTIVE-guided data curation improves temporal consistency and physical plausibility of generated motion.
- On the VBench benchmark, MOTIVE-selected high-influence data yields better motion smoothness and dynamic degree, with a 74.1% human preference win rate over the pretrained base model.
- The work, Motion Attribution for Video Generation, received an Outstanding Paper Honorable Mention at ICML 2026, highlighting its research significance.
- NVIDIA emphasizes that MOTIVE scales to modern, large, high-quality video datasets and models, enabling more principled fine-tuning of motion behavior.
- The project page offers technical details, examples, and evaluation results for researchers looking to apply MOTIVE to their own video generation pipelines.
Impact
By shifting attribution from static appearance to motion quality, MOTIVE gives video model developers a practical tool to systematically curate and fine-tune training data around temporal dynamics. This can narrow the gap between rapid advances in video generation and the still-lagging realism of motion, and may pressure competing labs to adopt similar motion-aware data curation strategies for their text-to-video systems.
