Cepheid: Decoupling Scale from Structure for Cross-Family Motion Completion
Synthetic voice
The authors address sparse-to-dense motion completion, recovering dense mesh motion from sparse vertex displacements across geometry families. They decouple residual structure from physical scale by learning a dimensionless prior with flow matching and estimating amplitude from observables plus a correction fitted only on training sources. They report that their unified model surpasses parameter-free harmonic extension in pooled error across all evaluated families, cutting it by 15.6% on unseen cloth and 10.1% on unseen legged bodies, while noting it crosses above harmonic extension at 0.5% observation density on legged bodies and that the two zero-shot application demos lack ground truth.
- Per the authors' transfer matrix, seven of twelve single-family transfers fail to improve on harmonic extension, the parameter-free control.
- The unified model cuts pooled error by 15.6% on unseen cloth and 10.1% on unseen procedural legged bodies, but the authors note it crosses above harmonic extension at 0.5% observation density on legged bodies (outperforming it from 1% up).
- On cloth, completion error with recovered amplitude is within 1.4–8.1% of a target-informed oracle (the per-condition optimal scalar that requires ground truth) across observation densities.