ContourCraft: Learning to Resolve Intersections in Neural Multi-Garment Simulations

TL;DR

ContourCraft tackles the challenge of garment intersections in learned multi-layer cloth simulations by introducing an Intersection Contour Loss that penalizes interpenetrations and drives their resolution. It embeds this loss within a GNN-based simulator (extending HOOD) and differentiates repulsive from non-repulsive interactions using open/closed intersection contours, enabling recovery from initial penetrations and handling dynamic collisions. The method is trained in three stages (physics-motivated pretraining, intersection avoidance, and intersection resolution) and validated on AMASS BEDLAM sequences, demonstrating substantial reduction of intersections, competitive perceptual realism, and automatic outfit resizing capabilities. This work advances practical learned garment simulation by enabling robust multi-layer outfits and resizing without requiring intersection-free initial geometries, with potential impact on games, films, and fashion design.

Abstract

Learning-based approaches to cloth simulation have started to show their potential in recent years. However, handling collisions and intersections in neural simulations remains a largely unsolved problem. In this work, we present \moniker{}, a learning-based solution for handling intersections in neural cloth simulations. Unlike conventional approaches that critically rely on intersection-free inputs, \moniker{} robustly recovers from intersections introduced through missed collisions, self-penetrating bodies, or errors in manually designed multi-layer outfits. The technical core of \moniker{} is a novel intersection contour loss that penalizes interpenetrations and encourages rapid resolution thereof. We integrate our intersection loss with a collision-avoiding repulsion objective into a neural cloth simulation method based on graph neural networks (GNNs). We demonstrate our method's ability across a challenging set of diverse multi-layer outfits under dynamic human motions. Our extensive analysis indicates that \moniker{} significantly improves collision handling for learned simulation and produces visually compelling results.

Figures (5)

• Figure 1: We distinguish between two types of garment nodes. Repelled nodes are those that either do not participate in the penetrations or lie outside a closed contour. Non-repelled nodes are those that are either part of an open contour or lie inside a closed one. If a triangle-node correspondence only contains repulsive nodes, we define this correspondence as repulsive and apply a repulsion loss to it, otherwise it is non-repulsive.
• Figure 2: A face (blue triangle) intersects a perpendicular plane (vertical line with intersecting segment shown in red). Green arrows show the negative partial gradients of the contour loss $\mathcal{L}_{IC}$ w.r.t. the triangle nodes. Blue arrows indicate negative partial gradients w.r.t. the coordinate of the intersection point $s_j$. The former gradient (green) squeezes the triangle to decrease the contour length, while the latter (blue) moves it along the plane's normal direction to resolve the intersection. We only use the gradient w.r.t. $s_j$ in our training. $J_A^{s_0}$ is the Jacobian of $s_0$ w.r.t. $A$.
• Figure 3: Since our method can resolve intersections present in the initial geometry, we can model automatically resized outfits without manual resolution of the intersections arising during this process.
• Figure 4: The plot shows the fraction of the triangle-triangle intersections left after each frame (up to 50) relative to the initial geometry. The values are aggregated across the whole validation set (40 sequences). Note that during dynamic movements new intersections may appear, hence the plot is not monotonic.
• Figure 5: For one of the validation sequences, we plot the number of intersecting triangle pairs for all compared ablations. Starting from an intersecting geometry, our method quickly resolves most penetrations. During dynamic and complex motion sequences (for instance, those with body self-intersections), it may miss new penetrations but then is able to recover from them as well.