SIGGRAPH / TOG

\documentclass[acmtog]{acmart}
\usepackage{graphicx}
\usepackage{amsmath,amssymb}
\usepackage{booktabs}

\acmJournal{TOG}
\acmYear{2026} \acmVolume{45} \acmNumber{4} \acmArticle{123}
\acmMonth{7}
\acmDOI{10.1145/XXXXXXX.XXXXXXX}
\citestyle{acmauthoryear}
\setcopyright{acmlicensed}

\begin{document}

\title{NeuralFabric: Real-Time Cloth Simulation with Learned Collision Response}

\author{First Last}
\affiliation{\institution{University of Example}\country{Country}}
\email{[email protected]}
\author{Jane Doe}
\affiliation{\institution{Example Research Labs}\country{Country}}
\email{[email protected]}
\author{John Smith}
\affiliation{\institution{University of Example}\country{Country}}
\email{[email protected]}
\renewcommand{\shortauthors}{Last, Doe, and Smith}

\begin{abstract}
Real-time cloth simulation requires fast yet accurate collision
response, historically a trade-off between physical fidelity and
interactive rates. We present NeuralFabric, a neural collision module
trained on high-quality reference simulations that plugs into existing
XPBD solvers with a 30$\times$ speedup over traditional implicit
solvers while producing visually indistinguishable results.
\end{abstract}

\begin{CCSXML}
<ccs2012><concept><concept_id>10010147.10010371</concept_id>
<concept_desc>Computing methodologies~Simulation</concept_desc></concept></ccs2012>
\end{CCSXML}
\ccsdesc[500]{Computing methodologies~Simulation}
\keywords{cloth simulation, collision detection, neural networks}

\begin{teaserfigure}
\centering
\rule{0.9\linewidth}{2.5cm}
\caption{Real-time cloth simulation with NeuralFabric.}
\label{fig:teaser}
\end{teaserfigure}

\maketitle

\section{Introduction}
Cloth simulation is a cornerstone of animation, games, and virtual
try-on. Collision response is typically the bottleneck.

\section{Related Work}
Implicit integration, XPBD, projective dynamics, neural physics.

\section{Method}
We train a rotation-equivariant GNN to predict collision velocity
corrections from local mesh and obstacle features.
\begin{equation}
\mathcal{L} = \alpha \mathcal{L}_{\text{pen}} + \beta \mathcal{L}_{\text{fric}} + \gamma \mathcal{L}_{\text{smooth}}.
\end{equation}

\section{Implementation}
Our runtime is in CUDA, integrated with the open-source Warp library.

\section{Results}
\begin{table}[t]
\centering
\begin{tabular}{lrr}
\toprule
Method & Frame time (ms) & Penetration (mm) \\
\midrule
Implicit Euler   & 248 & 0.8 \\
XPBD             &  22 & 3.1 \\
\textbf{Ours}    & \textbf{8.2} & \textbf{0.9} \\
\bottomrule
\end{tabular}
\caption{50k-vertex dress over 120 frames.}
\end{table}

\section{Limitations}
Our method requires retraining for radically different material models.

\section{Conclusion}
Learned collision response is now fast enough and accurate enough for
production real-time pipelines.

\bibliographystyle{ACM-Reference-Format}
\bibliography{refs}
\end{document}