SkinCells: Sparse Skinning using Voronoi Cells

TL;DR

SkinCells introduces a fully automatic, field-based approach to computing sparse skinning weights by optimizing a continuous weight field in canonical space. It uses a novel skin cell formulation, based on overlapping, parameterized Voronoi-like regions, to enforce sparsity and enable seamless Level-of-Detail weight generation. The method combines a DeltaMush-inspired smoothness loss, a location loss via skeleton-attached springs, and a sparsity loss to ensure per-vertex bone influences are limited, while remaining robust across complex geometries and garment configurations. A Mahalanobis-distance-based distance field with multiple per-joint sites yields topology-agnostic weight fields that transfer across similar meshes and resolutions, enabling fast, automated rigging suitable for mobile and real-time contexts. Empirical results show high-quality deformations, strong sparsity control, and faster convergence compared to traditional biharmonic or proximity-based methods, with successful application to cloth and skirt skinning and cross-LOD weight transfer.

Abstract

For decades, real-time skinning has been the cornerstone of character animation in visual effects and games. Despite its importance, the creation of animatable digital assets remains a labor-intensive manual process. Existing automated tools frequently struggle with intricate geometries, often necessitating significant manual refinement to reach production standards. We present a robust, fully automated method for generating high-quality skinning weights from a standard mesh and skeleton in a canonical A- or T-pose. Unlike traditional approaches, our framework offers direct sparsity controls to limit bone influences per vertex -- a critical requirement for maintaining performance in large-scale mobile environments. Furthermore, we address the challenge of Level-of-Detail (LoD) management by optimizing weights within a continuous spatial volume rather than on discrete vertices. This allows a single optimization pass to be applied seamlessly across multiple asset resolutions and variations. Central to our approach is a novel parameterized family of functions, we call SkinCells. We demonstrate that our method consistently produces stable, high-quality results even in complex scenarios where standard biharmonic weight computations fail.

Figures (13)

• Figure 1: A distance field (black at 0) for 11 regions (colored for clarity) marked with $\times$. (a) shows the field with uniformly shaped cells. (b) shows cells with randomly sampled $\mathbf{s}$ and $\mathbf{R}$ per site. In (c), each region is split into multiple sites sampled along the bones of the outlined skeleton.
• Figure 2: (a) shows the unnormalized weight field $\mathbf{w}$ for $l=1$ where each region $j$ is uniquely colored and white indicates a zero value. (b) plots the normalized Voronoi field with $l=3$ indicating that there are at most 3 nonzero weights for each point $\mathbf{x}$. The outline of the source skeleton is shown in black for clarity.
• Figure 3: Three examples of uniformly sampled poses used for further optimizing $\mathbf{w}$ by updating $S$ to minimize $\mathcal{L}$.
• Figure 4: The Voronoi field produced by our method (left) does not suffer from the same numerical instability resulting from softmax used by prior methods (right) where black represents NaN regions feng2023.
• Figure 5: The parameter $\lambda_{\text{loc}}$ is increased (middle) on the lion character (model in rest pose on the left) to produce deformations more faithful to the underlying skeleton. Decreasing $\lambda_{\text{loc}}$ (right) produces a smoother fold where the leg bends, albeit with more volume loss.