MOTIVE framework cuts video training data needs by 90% using motion gradients

MOTIVE (MOTIon attribution for Video gEneration) is a gradient-based data curation framework from NVIDIA, Princeton, and MIT that isolates motion dynamics in video diffusion and flow-matching models. Published on arXiv in January 2026 and honored as an ICML 2026 Outstanding Paper Honorable Mention, the method computes gradients through motion-weighted loss masks, filtering out static background influence and frame-length bias. Structured low-rank projections let it scale to billion-parameter architectures.

Fine-tuning on just 10% of MOTIVE-filtered data matches or beats training on the full dataset. That compression ratio matters for practitioners: training costs drop proportionally when you can discard nine-tenths of your corpus without sacrificing model quality. Existing data attribution methods suffer from what the authors call "kinetic blindness" — they credit static appearance over temporal dynamics, making them poor guides for physics-aware video generation. A clip of a ball rolling across grass might score high on visual diversity but low on motion complexity if the background dominates the gradient signal. MOTIVE addresses that by weighting loss contributions frame-by-frame based on optical flow magnitude, then normalizing by clip length. The result is a scoring function that genuinely reflects how much a training clip teaches the model about movement.

No model weights are released; MOTIVE is a curation tool, not a checkpoint. The implications extend beyond video: the same motion-aware gradient logic could apply to robotics datasets, physics simulations, or any domain where temporal dynamics matter more than static appearance. For open-source video model builders working with limited compute budgets, a 10× data efficiency gain is the difference between a weekend experiment and a multi-week cluster job.