A new framework lets researchers trace jerky or implausible motion in AI-generated video directly to the training clips responsible.
Motive — short for MOTIon attribution for Video gEneration — is a gradient-based tool that analyzes which clips in a fine-tuning dataset make motion better or worse. Unlike earlier attribution methods that focus on visual appearance, Motive isolates temporal dynamics specifically, using motion-weighted loss masks to separate how a clip moves from how it looks. The team tested it on text-to-video models and used its findings to curate a higher-quality fine-tuning set. The result was a 74.1% human preference win rate over the base model on VBench, a standard video generation benchmark.
The gap it fills is real. Video generation has improved sharply on visual quality — skin textures, lighting, scene composition — while motion has lagged behind, producing clips where objects teleport slightly or physics behaves strangely. Knowing which training examples cause those failures is the first step toward fixing them systematically rather than guessing.
The authors claim this is the first framework to attribute motion rather than appearance in video models, which is plausible given how recently motion quality became a benchmarked concern — though the field moves fast enough that the "first" label may have a short shelf life.