Motive
Summary
Motive (MOTIon attribution for Video gEneration) is a query-conditioned data-attribution method for video generators. It masks per-location training loss by optical-flow magnitude, projects per-example gradients to a compact vector, and ranks candidate fine-tuning clips by similarity to target motion-query gradients.
Official Artifacts
- Paper: https://arxiv.org/abs/2601.08828v2
- Project page: https://research.nvidia.com/labs/sil/projects/MOTIVE/
- ICML 2026 record: https://openreview.net/forum?id=zAl9heLw4q
- Award: ICML 2026 Outstanding Paper Honorable Mention
- Code: no official repository was linked from the paper or project page as of 2026-07-17.
Role In The Wiki
Motive is evidence that target-conditioned, gradient-based data curation can improve selected temporal dynamics during video-model fine-tuning. Its “10% beats 100%” result means fine-tuning a pretrained model on 1,000 selected clips versus all 10,000 clips in the experimental pool; attribution still processes every candidate, and the paper does not compare foundation-model pretraining corpora or establish a 90% end-to-end compute reduction. Motive is not itself a world model or dynamic curriculum: it has no action or control-input interface, and its paper evaluates one-shot subset selection rather than iterative training-time reweighting.