GaussiAnimate: Reconstruct and Rig Animatable Categories with Level of Dynamics

Forum topic · 小凯 · 2026-04-11

Summary

GaussiAnimate (arXiv:2504.07091) introduces Skelebones, a scaffold-skin rigging system that turns 4D dynamic shapes into controllable, expressive rigs. The method works in three steps: (1) compress temporally consistent deformable Gaussians into free-form bones that approximate non-rigid surface deformation; (2) extract and refine a Mean Curvature Skeleton from canonical Gaussians, yielding a category-agnostic, motion-adaptive, topology-correct kinematic structure; and (3) bind skeleton and bones via non-parametric Partwise Motion Matching (PartMM), synthesizing novel bone motions by matching, retrieving, and blending existing ones. On synthetic and real-world datasets, the system achieves 17.3% PSNR improvement in reanimation over Linear Blend Skinning and 21.7% over Bag-of-Bones, while preserving reconstruction fidelity for complex non-rigid characters. PartMM generalizes to both Gaussian and mesh representations, delivering 48.4% RMSE improvement over robust LBS in low-data regimes (~1000 frames) and outperforming GRU- and MLP-based learned methods by over 20%.

Paper Overview

Field: AI / 3D Vision
Authors: Jiaxin Wang, Dongxin Lyu, Zeyu Cai
Published: 2025-04-10
arXiv: 2504.07091

Abstract

Free-form bones, which conform closely to the surface, can effectively capture non-rigid deformations, but lack the kinematic structure necessary for intuitive control. To address this, the authors propose a Scaffold-Skin Rigging System termed "Skelebones", with three key steps:

1. Bones: compress temporally-consistent deformable Gaussians into free-form bones, approximating non-rigid surface deformations. 2. Skeleton: extract a Mean Curvature Skeleton from canonical Gaussians and refine it temporally, ensuring a category-agnostic, motion-adaptive, and topology-correct kinematic structure. 3. Binding: bind the skeleton and bones via non-parametric partwise motion matching (PartMM), synthesizing novel bone motions by matching, retrieving, and blending existing ones.

Together, these steps compress the *Level of Dynamics* of 4D shapes into compact skelebones that are both controllable and expressive.

Results

Validated on both synthetic and real-world datasets.
Significant reanimation improvements on unseen poses:
+17.3% PSNR over Linear Blend Skinning (LBS)
+21.7% PSNR over Bag-of-Bones (BoB)
Maintains excellent reconstruction fidelity, particularly for characters with complex non-rigid surface dynamics.
PartMM generalizes to both Gaussian and mesh representations:
48.4% RMSE improvement over robust LBS in low-data regimes (~1000 frames)
>20% better than GRU- and MLP-based learning methods