Cepheid: Decoupling Scale from Structure for Cross-Family Motion Completion

Chih-Hai Su, Tsung-Wei Ke, Yu-Lun LiuarXiv 2026 · arXiv

TL;DR

CEPHEID is a novel framework for sparse-to-dense motion completion that can predict the full movement of a 3D mesh from just a handful of tracked points (as few as 6–10). By decoupling the "structure" of motion from its physical "scale," the model achieves state-of-the-art performance across wildly different geometry families—from humans and animals to cloth and robots—even in zero-shot settings where it has never seen the object type before.

What Problem Does This Paper Address?

Recovering the dense motion of a 3D object from a few sparse observations is a classic challenge in computer vision and robotics. While we have specialized models for humans (using skeletons) or specific robots (using kinematic chains), creating a unified model that works across different "families" of objects is incredibly difficult.

The authors identify two major hurdles:

  1. The Scaling Barrier: The magnitude of motion varies by orders of magnitude between, say, a human walking and a flag waving. Standard deep learning models struggle to generalize when the "scale" of the target motion is unknown at test time.
  2. Negative Transfer: Models trained on one geometry (like humans) often perform worse than simple, parameter-free mathematical baselines (like harmonic extension) when applied to a different geometry (like cloth).

Methods

CEPHEID (named after the variable stars used to measure cosmic distances) solves these issues through a "decoupling" strategy:

  • Harmonic Extension as a Base: The model starts with a closed-form geometric solve called discrete harmonic extension. This provides a smooth "best guess" of the motion based purely on the mesh's connectivity.
  • Target-Independent Amplitude Recovery: Since the true scale of motion is unknown at inference, CEPHEID estimates a physical scale ($\hat{s}$) using "observables"—dimensionless features like observation density and the dispersion of the input points.
  • Dimensionless Residual Prior: Instead of predicting raw displacements, the model learns to predict a dimensionless residual (the difference between the harmonic base and the true motion). This is done using a flow-matching objective, which allows the model to capture complex, multi-modal distributions of motion.
  • Backbone: The architecture uses an MGN-T backbone, which combines message-passing (for local interactions) with Transformer attention (to capture long-range dependencies across the mesh).
The CEPHEID pipeline: decoupling residual structure from physical scale to assemble the final completed field.

Datasets

The model was trained and evaluated on a diverse set of data:

  • Training Families: DT4D (animated animals/humanoids), AMASS (human motion capture), and Franka (robotic arm kinematics).
  • Zero-Shot Evaluation: To test generalization, the authors held out Cloth (simulated flags) and Legged Robots (procedurally generated multi-legged embodiments).
  • Real-World Applications: The model was also tested on point tracks from the DAVIS video dataset and proprioceptive data from Isaac Lab simulations.

Key Insights and Ideas

  • Dimensionless is Better: By forcing the model to work with dimensionless units, the authors found it much easier to train a single prior that generalizes across families. The "physics" of how a surface deforms becomes more consistent once you strip away the raw meters or millimeters.
  • The "Safety Net" of Math: Using harmonic extension as a base ensures the model always has a geometrically plausible starting point. If the learned residual is zero, you still get a reasonable (if slightly dampened) result.
  • Scale Recovery is the Key: The authors show that "specialist" models (trained only on one family) often fail on new objects because they either wildly over-correct or under-correct the scale. CEPHEID’s amplitude recovery brings the error within 1–8% of an "oracle" that knows the ground truth scale.
Qualitative comparison on cloth. CEPHEID is the only model that keeps the flag intact and stays below the harmonic baseline.

Key Results

  • Zero-Shot Superiority: CEPHEID outperformed the harmonic baseline on all five evaluated families. It reduced error by 15.6% on unseen cloth and 10.1% on unseen legged robots.
  • Beating Specialists: A single unified CEPHEID model often outperformed "specialist" models trained specifically on the target family when those specialists were tested in sparse-data regimes.
  • Real-World Robustness: The model successfully "densified" sparse tracks from a standard video point tracker (CoTracker3) on animal videos, completing motion even for parts of the animal hidden from the camera.
Demo showing the model densifying video point tracks on a horse, including points on the far side hidden from the camera.

Limitations and Future Work

  • Clean Inputs: The current quantitative evaluation assumes "perfect" vertex displacements. In the real world, sensor noise, occlusions, and depth distortion remain challenges.
  • Mesh Dependency: While the model can work on point sets via k-NN graphs, it performs best when a high-quality surface mesh is available.
  • Fixed Sensor Layouts: The model was trained on random "masks" of points. The authors noted that if a robot always uses the same 3-4 sensors, fine-tuning specifically for that layout can further halve the error.

Overall, CEPHEID represents a significant step toward "foundation models" for 3D motion, moving away from rigid, skeleton-based priors toward a more flexible, geometry-aware understanding of how things move.

Summary updated

Abstract

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. 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. Since this scale depends on unknown target motion, it cannot be measured directly at test time. 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. The same prior applies without retraining to point-tracked animal video and a robot grasping a soft body. Project page: https://su-terry.github.io/cepheid-website/
BibTeX
@misc{su2026cepheid,
  author = {Chih-Hai Su and Tsung-Wei Ke and Yu-Lun Liu},
  title = {{Cepheid: Decoupling Scale from Structure for Cross-Family Motion Completion}},
  year = {2026},
  eprint = {2610.18871},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.18871},
}
Figure