A Neural-Network-Based Approach for Loose-Fitting Clothing

TL;DR

The paper tackles real-time animation of loose-fitting garments by merging a low-DOF ballistic rope-chain physics core with neural skinning and quasistatic shape inference to deliver high-resolution cloth shapes without the typical stability issues. It introduces vertical rope chains with inextensibility, augmented by neural networks that infer PCA-based skinning and high-frequency shape from the chain DOFs, and uses PINN-style collision losses with analytic or neural SDFs to maintain non-interpenetration. The approach achieves substantial DOF reduction (roughly 80–90x) while producing realistic motion and avoiding common real-time artifacts, though interpenetration remains more challenging for out-of-distribution inputs. The work points to potential extensions in differentiable rope chains, broader garment generalization, and improved collision handling for real-time, in-distribution, and out-of-distribution scenarios alike.

Abstract

Since loose-fitting clothing contains dynamic modes that have proven to be difficult to predict via neural networks, we first illustrate how to coarsely approximate these modes with a real-time numerical algorithm specifically designed to mimic the most important ballistic features of a classical numerical simulation. Although there is some flexibility in the choice of the numerical algorithm used as a proxy for full simulation, it is essential that the stability and accuracy be independent from any time step restriction or similar requirements in order to facilitate real-time performance. In order to reduce the number of degrees of freedom that require approximations to their dynamics, we simulate rigid frames and use skinning to reconstruct a rough approximation to a desirable mesh; as one might expect, neural-network-based skinning seems to perform better than linear blend skinning in this scenario. Improved high frequency deformations are subsequently added to the skinned mesh via a quasistatic neural network (QNN). In contrast to recurrent neural networks that require a plethora of training data in order to adequately generalize to new examples, QNNs perform well with significantly less training data.

Figures (16)

• Figure 1: The loose-fitting garments that we use for numerical experiments. The cape mesh consists of 12690 vertices, and we define 10 rope chains with 13-15 virtual bones in each chain (reducing the total DOFs by a factor of approximately 90). The skirt mesh consists of 18546 vertices, and we define 26 rope chains with 9 virtual bones in each chain (reducing the total DOFs by a factor of approximately 80). Importantly, the large reduction in the number of DOFs is highly beneficial for both RAM and cache performance, not just CPU performance. Note that the unattached virtual bones near the top of both the cape and the skirt are not simulated, but they will be used for skinning (see Section \ref{['sec:skinning']}).
• Figure 2: The far left subfigure shows a a coarsely discretized mesh with position constraints on the five red nodes. The next three subfigures show the results of steady state simulations using a decreasing spring stiffness (from left to right). The final subfigure shows a rope chain simulation (the rope chains are shaded green) of the same degrees of freedom. The mass-spring simulations lock with stiffer springs and overstrech with weaker springs. The spring stiffness in the middle subfigure was chosen to approximately match the downward stretching extent of the rope chain simulation, which is what one would expect without overstretching; however, locking occurs since the springs are still too stiff.
• Figure 3: In this figure, we simulate two virtual bones connected by a single rope with the top virtual bone fixed and the bottom virtual bone free to rock back and forth as a pendulum. The results compare well to the analytic solution for pendulum motion for both shorter times (left) and longer times (right), illustrating the efficacy of our numerical approach.
• Figure 4: Motion of a single swinging rope chain, showcasing varying numbers of Gauss-Seidel iterations. Left to right: 1, 5, 10 iterations, and iterating until the relative error is smaller than a tolerance of $10^{-6}$. With more iterations, the rope chain is less damped (as expected). The same number of iterations (or the same tolerance) was used for both the tension and impulse computations. Note that it might be more efficient (depending on the application) to use a different number of iterations on the tension and impulse computations.
• Figure 5: For the sake of computational efficiency, we represent the volumetric body (left) with a number of analytically defined SDFs (right). Although other representations (three dimensional grid discretizations, neural representations, etc.) are compatible with our approach, they incur higher computational costs.