Cepheid: Decoupling Scale from Structure for Cross-Family Motion Completion
Sparse-to-dense motion completion recovers dense mesh motion from sparse vertex displacements.
ProofPaper ↗
Key points
- Residual priors trained on one geometry family frequently fail to beat parameter-free harmonic extension off-family.
- Moreover, cross-family training meets a scaling barrier: residual amplitude varies by more than an order of magnitude across geometry families and collapses as observations become denser.
- We decouple residual structure from physical scale by learning a dimensionless prior, estimating amplitude from observables and a lightweight correction fitted only on training sources.
- Our unified model surpasses harmonic extension across all evaluated families, including unseen cloth and legged bodies, without skeletons, registration, or cross-mesh correspondence.
Sources (2)
- [1]Cepheid: Decoupling Scale from Structure for Cross-Family Motion CompletionHugging Face Daily Papers · Oct 13, 12:00 AM
Sparse-to-dense motion completion recovers dense mesh motion from sparse vertex displacements.
Residual priors trained on one geometry family frequently fail to beat parameter-free harmonic extension off-family.
- [2]Cepheid: Decoupling Scale from Structure for Cross-Family Motion CompletionarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 05:59 PM · same content
Extractive summary: sentences quoted from the sources.
Before this
- Oct 12, 2026The Value of Being Wrong: Self-Mined Visual In-Context Learning from Errors
- Oct 12, 2026AReA: Adaptive Receptive-Field Attention for Efficient High-Resolution Diffusion Transformers
- Oct 11, 2026SpecFit: Not All Draft Tokens Cost the Same in Offloaded MoE Inference
- Oct 11, 2026DC-Flow: Target-Conditioned Cyclic Peptide Design via Delayed-Coupling Flow Matching
- Oct 10, 2026Text-Guided Medical Image Segmentation via Residual Aggregation and Dense Alignment Representations
- Oct 10, 2026CAST: Real-Time Motion Capture for Any Skeleton Topology