RigMo: Unifying Rig and Motion Learning for Generative Animation

How to read this

Category

Feed-forward joint rig and motion learning for 4D generative animation

Contributions

• Encodes per-vertex deformations from raw mesh sequences into two latent spaces, a rig latent decoded into explicit Gaussian bones and skinning weights, and a motion latent producing per-frame SE(3) transforms, with no human rig annotations required
• Trains self-supervised on unlabeled 4D data (DeformingThings4D, Objaverse-XL, TrueBones), giving feed-forward rig inference that generalizes across categories, unlike per-sequence SSDR-style optimization
• Introduces Motion-DiT, a diffusion transformer operating in RigMo's learned motion latent, showing the structure-aware latent supports downstream motion generation and interpolation tasks

Context

RigMo sits at the intersection of auto-rigging methods that need artist-annotated skeletons (RigNet-style), inverse-skinning optimization that fits one rig per sequence (SSDR), and motion generation models that assume a ground-truth rig already exists (AnyTop and similar). It tries to unify all three by inferring rig structure and motion jointly in a single feed-forward pass instead of treating them as separate stages.

Correctness

The claims rest on comparisons against auto-rigging and deformation baselines across three datasets, with no manual rig ground truth used in training, so quality is judged by reconstruction fidelity and category generalization rather than agreement with artist rigs. It is a CVPR paper, so it has passed peer review, but the abstract and intro give no indication of testing on production characters with clothing or accessories.

Clarity

Reads like a typical CVPR paper: a clear problem framing and related-work section, then a notation-heavy method section covering the Gaussian bone algebra and dual-branch VAE. Approachable for someone who already knows the auto-rigging and skinning literature, slower going for a pure animator.

How to read it

First pass: abstract, Figure 1 teaser, and Figure 2 architecture diagram to get the two-latent-space idea. Second pass: introduction and related work to place it against RigNet, SSDR, and AnyTop. Third pass: Section 3.1's topology-aware encoder and Gaussian-bone LBS formulation, plus the experiment tables on DeformingThings4D and Objaverse-XL, if you plan to reimplement or compare against it.