Abstract
Traditional optimization-based shape editing often requires manual rigging or the placement of static constraints to prevent global "ripple effects" during deformation. We present a novel approach to local shape editing that utilizes a "smoothly clamped $\ell_1$" (SC-L1) regularization term. This sparsity-inducing loss allows for geometry-aware, physics-based deformations that remain localized without manual intervention. By employing a specialized three-block Alternating Direction Method of Multipliers (ADMM) optimization scheme, our method achieves interactive performance, outperforming previous sparsity-based solvers by up to three orders of magnitude.
Introduction
In digital content creation, artists frequently need to modify specific regions of a 3D model while ensuring the rest of the geometry remains unchanged. Current tools generally fall into two categories: direct manipulation brushes and optimization-based solvers. Brushes are fast but "geometry-blind," often affecting nearby but disconnected parts of a mesh. Conversely, optimization-based methods like As-Rigid-As-Possible (ARAP) produce natural, physically plausible results but are inherently global; moving a single vertex can cause subtle shifts across the entire model unless the artist manually "pins" the rest of the mesh in place.
This paper introduces a method that automates locality through a new regularization framework, removing the need for tedious manual rigging while maintaining the physical realism of global optimization.
The Smoothly Clamped $\ell_1$ Regularizer
The primary contribution of this work is the Smoothly Clamped $\ell_1$ (SC-L1) regularization term. To achieve true locality, a solver must encourage zero displacement for the majority of the mesh while allowing large, physically-driven displacements in the edited region.
Standard $\ell_1$ norms are effective at inducing sparsity (forcing values to zero), but they continue to apply a linear penalty as displacement increases. This often results in "dragging" artifacts where the mesh is unnaturally pulled toward its original rest position. The SC-L1 norm addresses this by behaving like a standard $\ell_1$ norm near the origin—effectively "pinning" unedited vertices—but clamping to a constant value once a displacement threshold ($s$) is reached. This clamping allows the underlying elastic energy of the model to dictate the shape of the deformation in the active region without interference from the sparsity term.
High-Performance Optimization
To make this regularization practical for interactive use, the authors developed a three-block ADMM optimization strategy. Sparsity-constrained problems are notoriously difficult to solve quickly due to the non-smooth nature of the $\ell_1$ norm and the non-linearity of elastic energies.
The three-block approach decouples the problem into three distinct sub-steps:
- Physics/Rotation Step: Handles the local rotations and elastic energy (e.g., ARAP).
- Sparsity Step: Applies the SC-L1 shrinkage operator to enforce locality.
- Global Position Step: Solves a linear system to reconcile the vertex positions.
By splitting the variables this way, each step can be solved efficiently using either closed-form shrinkage operators or pre-factored linear systems, enabling the system to maintain high frame rates even on complex meshes.
Results and Discussion
The method was evaluated on a diverse set of models, including 3D meshes (such as the "Octocat," pigs, and robots), high-resolution cloth simulations (up to 47k vertices), and 1D vector graphics.
Performance
The three-block ADMM solver demonstrates significant performance gains over previous state-of-the-art sparsity solvers. For small deformations, the system is up to 1000x faster than the method proposed by Chen et al. (2017), and it maintains a 100x speedup even during large-scale deformations. This efficiency allows for a "rig-free" workflow where artists can interactively manipulate geometry without any pre-processing or weight painting.
Geometry Awareness
Unlike radius-based brushes, the SC-L1 approach is inherently geometry-aware. Because the locality is governed by the mesh's own elastic energy, the deformation respects the connectivity of the model. For example, editing a character's finger will not inadvertently move a nearby thumb if they are not topologically connected, as the energy required to bridge that gap would exceed the sparsity threshold.
Limitations and Future Work
While the current system is highly effective for mesh-based editing, it relies on per-vertex regularization. This makes it less directly applicable to continuous representations like NURBS or splines without prior tessellation. Additionally, while the ADMM solver is fast, its convergence can slow down when extreme precision is required or when paired with highly non-convex energy models. Future research may explore providing artists with more intuitive controls over the scale of locality ($s$ and $w$ parameters) to further refine the interactive experience.
Conclusion
By combining a novel smoothly clamped sparsity norm with an efficient ADMM solver, this work provides a powerful new tool for interactive shape modeling. It bridges the gap between the speed of heuristic brushes and the quality of global optimization, allowing for a more fluid and intuitive creative process.
Summary updated