Figure 1. Starting from real fabrics, we construct the matching yarn structures and derive the yarn model parameters. Subsequently, we sample a range of in-plane and out-of-plane deformations, run periodic yarn-level simulations on each deformed surface and record the yarn responses, which we use for optimizing the membrane and bending elasticity parameters of a shell model. We validate our shell parameters by comparing against full yarn-level simulations as well as real fabric captures and measurements. Finally, we simulate full-body garments and compare to reference footage.

Estimating Cloth Elasticity Parameters From Homogenized Yarn-Level Models

Abstract.

Virtual garment simulation has become increasingly important with applications in garment design and virtual try-on. However, reproducing garments faithfully remains a cumbersome process. We propose an end-to-end method for estimating parameters of shell material models corresponding to real fabrics with minimal priors. Our method determines yarn model properties from information directly obtained from real fabrics, unlike methods that require expensive specialized capture systems. We use an extended homogenization method to match yarn-level and shell-level hyperelastic energies with respect to a range of surface deformations represented by the first and second fundamental forms, including bending along the diagonal to warp and weft directions. We optimize the parameters of a shell deformation model involving uncoupled bending and membrane energies. This allows the simulated model to exhibit nonlinearity and anisotropy seen in real cloth. Finally, we validate our results with quantitative and visual comparisons against real world fabrics through stretch tests and drape experiments. Our homogenized shell models not only capture the characteristics of underlying yarn patterns, but also exhibit distinct behaviors for different yarn materials.

1. Introduction

The physical realism of simulated cloth plays a critical role in digital garment design. Cloth simulation, once primarily used in video games and animations, has now become a crucial aspect of this creative process. Increasingly, faithful reproduction of clothing behaviors has become a focus point in garment design software (CLO3D, 2024; Browzwear, 2024; Style3D, 2024) where the designed textiles are simulated to determine the fit on a virtual body. The simulator needs to accurately predict the behaviors of different types of fabrics composed of numerous knitted or woven yarn structures consisting of various materials. While (Wu et al., 2020) demonstrate simulation-assisted design of complex 3D patterns such as a woven shoe, modeling and simulating faithful full body garments that account for the underlying fabric compositions remains challenging.

Textile design software commonly model fabrics as linear or hyperelastic materials (CLO3D, 2024). While it is possible for shell simulations to be computed in real-time on the GPU, they generally require tedious manual processes and expert knowledge in tuning the parameters to reproduce satisfactory dynamics of a desired material type. In addition, the limited degrees of freedom in shell models makes it exceedingly hard for shell simulations to exhibit the rich behaviors that we observe in real fabrics such as edge curling in stockinette patterns and anisotropic bending in weave patterns.

Several approaches have been explored to produce more faithful cloth behaviors. One solution is to model cloth as a collection of interacting yarns (Kaldor et al., 2008; Leaf et al., 2018). This approach is quite accurate in reproducing the complex cloth behaviors but at an excessively high computational cost. Another line of research focuses on improving the shell model and gaining better automatic estimations for the model parameters. A handful of works (Larionov et al., 2022; Miguel et al., 2012; Clyde et al., 2017; Wang et al., 2011; Feng et al., 2022) estimate the cloth elasticity parameters from a range of data gathered from real fabrics including physical experiments of mechanical properties and high resolution images. Although these methods are capable of producing high-quality results, they require potentially expensive specialized equipment. On the other hand, Sperl et al. (2020) learns shell material models from yarn-level deformation responses using numerical homogenization, which is closely related to our work.

We present a method for estimating the shell-level cloth material parameters required to simulate real fabrics given only basic information about the fabric. Using physics-based simulations, we provide parameter estimations that reflect the behaviors of various fabric compositions. Given a real fabric, we derive the yarn model parameters from simple measurements as well as freely available experimental data from textile research, and use numerical homogenization (Sperl et al., 2020) for collecting yarn-level responses to a range of shell deformations including anisotropic bending, while accounting for the nonlinearity in yarn stretching. Our periodic yarn relaxation result (see Fig. 3) can further drive an efficient surface-based cloth appearance model (Zhu et al., 2023), which drastically enhances rendered results.

In contrast to previous work that learns a different material model for every yarn pattern (Sperl et al., 2020), we optimize the thin shell model parameters directly and demonstrate that the parameters are not only determined by the yarn patterns, but also the physical properties such as material type and yarn thickness. Furthermore, we estimate the off-diagonal terms in the bending matrix to capture anisotropic bending effects, thus constructing the entire stiffness matrix through homogenization and eliminating the laborious process of tuning parameters by hand. Finally, we validate our homogenized shell parameters against machine measurements of real fabric swatches, and discover that the simulated fabrics are not only capable of visually establishing the representative behaviors of each material, but also match the real fabrics quantitatively despite the absence of an inverse process in our pipeline.

Our approach generalizes to shell models that are based on the first two fundamental forms (i.e. stretching and bending strains). In addition, we estimate bending parameters along all three out-of-plane directions, providing reliable initial estimates for more complex parameter estimation methods, which can improve over our estimates using specialized machines.

To summarize, we offer the following technical contributions, which allow us to model shell-level cloth starting from only a basic fabric description:

• An end-to-end method to estimate shell simulation parameters corresponding to real fabrics that is compatible with any shell material model based on the first and second fundamental forms.

• A novel yarn parameter estimation method from simple measurements of real fabrics, with a nonlinear stretching model incorporating experimental results of textile research.

• An approach for homogenizing anisotropic bending including the warp-weft coupling term.

2. Related Work

Shell-Level Cloth Simulation. Starting with the pioneering work of Baraff and Witkin (1998), thin shell cloth simulations have seen widespread success over the last decades with applications in animation and special effects (Stuyck, 2022). Since then, many improvements have been made relating to robustness, accuracy and efficiency. Müller et al. (2007) proposed a Position Based Method (PBD) designed for efficiency where constraints are solved in parallel by updating positions directly. To overcome the limitations of PBD, eXtended Position Based Dynamics (XPBD) (Macklin et al., 2016) was proposed to eliminate iteration count dependency of simulation results. Bouaziz et al. (2014) employed an implicit integrator which bridges the gap between continuum mechanics and PBD. In contrast to prior linear element methods, Ni et al. (2023) simulated thin shells using the bicubic Hermite element method.

Yarn-Level Cloth Simulation. Shell-level cloth has seen many advances in recent years which significantly improve realism, yet it remains difficult to reproduce all intricate behaviors of fabrics. To improve on these models, Kaldor et al. (2008) proposed original work to model knits at the yarn-level. Generating yarn-level geometries can be cumbersome. To alleviate this, Yuksel et al. (2012) created full garment yarn-level geometries based on simple triangle models. Albeit more computationally expensive, they were capable of reproducing unique effects such as edge curling under tension. Cirio et al. (2014) extended this work with a focus on computational efficiency by modeling interlaced yarns based on yarn crossings and yarn sliding, with implicit contacts and persistent contacts in follow-up work (Cirio et al., 2015, 2016). A different way to improve efficiency is to model the yarns periodically (Leaf et al., 2018), enabling interactive pattern design. In order to obtain yarn-level like behavior at moderate computational cost, Casafranca et al. (2020) combined both shell-level and yarn-based simulation models in a single framework.

Numerous yarn representations have been proposed. Remion et al. (1999) developed equations for modeling knitted patterns with spline curves, and uses springs between control points for length preservation. Jiang and Chen (2005) modeled the relaxed configuration of woven patterns using spline curves. There has also been a series of work on modeling and simulating thin elastic rods based on Kirchhoff rods (Bergou et al., 2008; Bertails et al., 2006; Spillmann and Teschner, 2007) that uses discrete poly-line representations. Our yarn-level simulation employs the Cosserat rod model (Spillmann and Teschner, 2007) and the periodic boundary conditions (Leaf et al., 2018).

Multiscale Modeling and Homogenization of Cloth. In computer graphics, representing cloth using a mixture of scales has been extensively studied to leverage the efficiency of simulating shell models while preserving the detailed effects produced by yarn models. Martin et al. (2010) defined a unified, high-order integration for elastic responses shared among elastic rods, shells and solids. Casafranca et al. (2020) proposed a kinematic transition between triangle and yarn representations. Fei et al. (2018) produced anisotropic effect in cloth-fluid coupling by modeling the fiber-level porous structures. Our work is most closely related to Schumacher et al. (2018) and Sperl et al. (2020) that derived shell models using elastic responses of underlying thin structures subjecting to shell deformations. Sperl et al. (2022) presented a technique for modeling the yarn-level mechanics of cloth, based on real fabric responses to mechanical forces. Sperl et al. (2021) animated yarn-level cloth geometry on top of a deformed mesh while accounting for the underlying mechanics.

Material Parameter Estimation. Several works focus on estimating material models and parameters to model clothing based on captured data. Early work optimized material parameters from videos of captured cloth (Bhat et al., 2003). Follow up work captured cloth deformations under known loads to estimate the material models (Miguel et al., 2012; Clyde et al., 2017; Wang et al., 2011). Larionov et al. (2022) proposed the use of an FFT-based loss to robustly handle bifurcations in cloth buckling. Differentiable simulation has successfully been applied to material estimation from synthetic (Stuyck, 2022) and real data (Li et al., 2023). Wang et al. (2020) learned a material model directly instead of fitting parameters to existing material models.

3. Background

We briefly review a few key concepts on which our parameter estimation pipeline is based. We introduce the Cosserat rod model for representing yarns, the orthotropic StVK material for modeling thin shells, as well as numerical homogenization detailed by Sperl et al. (2020). We expand on XPBD, a constraint-based simulation framework, that we employ for both the yarn-level and shell-level simulations.

Although we use constraint based dynamics, our approach is compatible with any solver that finds the static equilibrium of yarns. Our technical contribution lies in deriving strain-dependent yarn model parameters from freely available measurements of real yarns and fitting shear bending terms to homogenized yarn-level cloth, which was omitted in previous work.

3.1. Compliant Constraint Formulations of Models

For both yarn and shell simulations, we follow Macklin et al. (2016) and minimize an energy potential of the form

(1) $$U(\bm{x})=\frac{1}{2}\bm{C}(\bm{x})^{\top}\bm{\alpha}^{-1}\bm{C}(\bm{x})$$

where $\bm{\alpha}$ is a compliant matrix that represents the inverse stiffness and $\bm{C}$ are bilateral constraint functions describing the kinematics of the dynamic system (Bender et al., 2015). We describe shell-level constraints with $\overline{\bm{C}}$ to distinguish from yarn-level constraints $\bm{C}$.

3.2. Modeling Yarns With Cosserat Rods

A Cosserat rod is described by a centerline curve and a set of local orthonormal frames $(\bm{d}_{1},\bm{d}_{2},\bm{d}_{3})$ attached to each point on the curve, such that $\bm{d}_{3}$ aligns with the centerline direction and $(\bm{d}_{1},\bm{d}_{2})$ span the cross-sectional direction of the rod.

We represent the yarn centerline as a set of points $\{\bm{p}_{0},\dots,\bm{p}_{N}\}$ and edges $\{\bm{e}_{0},\dots,\bm{e}_{N-1}\}$ where $\bm{e}_{i}=\bm{p}_{i+1}-\bm{p}_{i}$. Each edge has an associating local frame written as a quaternion $\bm{q}$ that consists of a real part $q_{0}=\Re(\bm{q})$ and an imaginary part $\bm{q}_{im}=\Im(\bm{q})$, representing the angle of rotation $\theta=\cos^{-1}(q_{0})$ around an axis $\frac{\Im(\bm{q})}{\sin\theta}$. In our yarn model, $\Im_{0},\Im_{1}$ represent bending and $\Im_{2}$ corresponds to twisting. Contrary to the Discrete Elastic Rod formulation (Bergou et al., 2008) where the bending and twisting are calculated from $(\phi,\theta)$ corresponding to the turning angle between adjacent edges and the rotation angle of local frames, the quaternion is a unified representation of both the bending and twisting, i.e. $\bm{q}^{T}=[q_{0},\bm{q}_{im}^{T}]^{T}$ such that $\bm{R}\left(\bm{q}\right)=[\bm{d}_{1},\bm{d}_{2},\bm{d}_{3}]$, and requires additional geometric constraints to align the local frames with the centerline. The Rodriguez formula $\bm{R}\left(\bm{q}\right)$ computes the rotation from world frame to local frame.

(2) $$\bm{R}\left(\bm{q}\right) =2\bm{q}\bm{q}^{T}+\left(q_{0}^{2}-\bm{q}_{im}^{T}\bm{q}_{im}\right)\bm{1}+2q_{0}\left[\bm{q}_{im}\right]^{\times}$$

Our yarn model follows the formulation by Kugelstadt and Schömer (2016) where we model stretching-shearing and bending-twisting energies. The total energy for a yarn with $N+1$ points is

(3) $$U_{\text{yarn}} =U_{\text{stretch\_shear}}+U_{\text{bend\_twist}}$$

$$=\sum_{i=0}^{N-1}\frac{1}{2}\bm{C}_{s}^{T}\bm{\alpha}_{s}^{-1}\bm{C}_{s}+\sum_{i=0}^{N-2}\frac{1}{2}\bm{C}_{b}^{T}\bm{\alpha}_{b}^{-1}\bm{C}_{b}\ \text{where}$$

$$\bm{\alpha}_{s}^{-1} =\begin{bmatrix}EA&&\\ &EA\\ &&EA\end{bmatrix}l_{0},\ \ \bm{\alpha}_{b}^{-1}=\begin{bmatrix}EI_{1}&&\\ &EI_{2}&\\ &&GI_{3}\end{bmatrix}l_{0}.$$

Here, $E$ is the Young’s modulus, $G$ is the shear modulus where $E=2G(1+\nu)$, $\nu$ is the Poisson’s ratio, and $I_{i}$ is the second area moment of inertia. For a cylindrical yarn with circular cross-section, $A=\pi r^{2}$, $I_{1}=I_{2}=\frac{1}{4}\pi r^{2}$ and $I_{3}=\frac{1}{2}\pi r^{2}$.

Stretching-Shearing

(4) $$\bm{C}_{s}\left(\bm{p}_{i},\bm{p}_{i+1},\bm{q}_{i}\right) =\frac{1}{l_{0}}\left(\bm{p}_{i+1}-\bm{p}_{i}\right)-\bm{d}_{3}(\bm{q}_{i})$$

$$=\frac{1}{l_{0}}\left(\bm{p}_{i+1}-\bm{p}_{i}\right)-\bm{R}(\bm{q}_{i})\hat{\bm{e}}_{3},$$

where $l_{0}$ is the rest length of edge $\bm{e}_{i}$ and $\hat{\bm{e}}_{3}$ is a unit vector in the yarn center-line direction in world frame.
Bending-Twisting

(5) $$\bm{C}_{b}\left(\bm{q}_{i},\bm{q}_{i+1}\right) =\bm{\Psi}-\bm{\Psi}^{0}=\frac{2}{l_{0}}\frac{\Im\left(\bar{\bm{q}}_{i}\bm{q}_{i+1}\right)}{\Re\left(\bar{\bm{q}}_{i}\bm{q}_{i+1}\right)}-\frac{2}{l_{0}}\frac{\Im\left(\bar{\bm{q}}^{0}_{i}\bm{q}^{0}_{i+1}\right)}{\Re\left(\bar{\bm{q}}^{0}_{i}\bm{q}_{i+1}^{0}\right)},$$

where the superscript 0 denotes the rest configuration and $\bm{\Psi}$ is the modified Darboux vector.

3.3. Modeling Shells with Orthotropic StVK

To simulate hyperelastic shells, we require membrane (in-plane) and bending (out-of-plane) energies. We focus on 2D knitted and woven patterns where the yarn patterns are aligned with two orthogonal directions, which we henceforth refer to as weft and warp indicated by subscripts ${}_{t}$ and ${}_{p}$ respectively. We compute the strain following the discrete Kirchhoff-Love shell model (Weischedel et al., 2012; Chen et al., 2018; Wen and Barbič, 2023). The in-plane stretching strain can be represented as

(6) $$\overline{\bm{\epsilon}}_{s}=\begin{bmatrix}\epsilon_{tt}&\epsilon_{tp}\\ \epsilon_{tp}&\epsilon_{pp}\end{bmatrix}\propto\overline{\mathbb{I}}_{\text{rest}}^{-1}(\overline{\mathbb{I}}-\overline{\mathbb{I}}_{\text{rest}})$$

where $\overline{\mathbb{I}}$ and $\overline{\mathbb{I}}_{\text{rest}}$ are the first fundamental forms on the deformed and rest surfaces respectively. The elements $\epsilon_{tt}$, $\epsilon_{pp}$ and $\epsilon_{tp}$ measure weft, warp and shear deformations respectively. Let us define $\tilde{\mathbb{I}}=\overline{\mathbb{I}}_{\text{rest}}^{-1}(\overline{\mathbb{I}}-\overline{\mathbb{I}}_{\text{rest}})$, then the constraint and stiffness matrix are

(7) $$\overline{\bm{C}}_{s} =\sqrt{A}(\tilde{\mathbb{I}}_{tt},\tilde{\mathbb{I}}_{pp},\tilde{\mathbb{I}}_{tp})$$

$$\overline{\bm{\alpha}}^{-1}_{s} =h\begin{bmatrix}s_{00}&s_{01}&\\ s_{01}&s_{11}&\\ &&s_{22}\end{bmatrix}$$

where $A$ is the area of a triangle, and $h$ is the thickness of the shell. The weft and warp stiffness of the fabric are controlled by $s_{00}$ and $s_{11}$ respectively. $s_{01}$ controls the coupling between deformations in warp and weft directions, and $s_{22}$ is the shear stiffness. We can express the stiffness values in terms of the more intuitive Young’s moduli $E_{p},E_{t}$, Poisson’s ratios $\nu_{pt},\nu_{tp}$ and shear modulus $\mu$:

(8) $$\begin{bmatrix}s_{00}&s_{01}&\\ s_{01}&s_{11}&\\ &&s_{22}\end{bmatrix}=\frac{1}{1-\nu_{tp}\nu_{pt}}\begin{bmatrix}E_{t}&\nu_{pt}E_{t}&\\ \nu_{tp}E_{p}&E_{p}&\\ &&\mu(1-\nu_{tp}\nu_{pt})\end{bmatrix}.$$

The bending strain is proportional to the change in the curvature of the surface $\tilde{\mathbb{II}}=\overline{\mathbb{I}}_{\text{rest}}^{-1}(\overline{\mathbb{II}}-\overline{\mathbb{II}}_{\text{rest}})$, where $\overline{\mathbb{II}}$ and $\overline{\mathbb{II}}_{\text{rest}}$ are the second fundamental forms. Using the same formulation, we define the bending constraints as

(9) $$\overline{\bm{C}}_{b}=\sqrt{A}(\tilde{\mathbb{II}}_{tt},\tilde{\mathbb{II}}_{pp},\tilde{\mathbb{II}}_{tp})$$

All our experiments assume a flat rest shape with $\overline{\mathbb{II}}_{\text{rest}}=\bm{0}$. The stiffness matrix is $\overline{\bm{\alpha}}^{-1}_{b}=h^{3}\operatorname*{diag}(b_{00},b_{11},b_{22})$.

Figure 2. A square swatch (oriented vertically along the warp direction) is draped using various bending parameters. Each column indicates which parameter is modified with respect to the others. The right column makes all parameters either soft (top row) or stiff (bottom row).

3.4. Homogenization

Homogenization aims at obtaining macroscopic equations for systems with a fine microscopic structure. Following Sperl et al. (2020), we describe the macroscopic thin shell using reference coordinates $\overline{\bm{X}}$, deformed coordinates $\overline{\bm{x}}$, and deformation gradient $\overline{\bm{F}}=\frac{\overline{\bm{x}}}{\overline{\bm{X}}}$. Correspondingly, the microscopic yarn-level structure has the descriptors $\bm{X}$, $\bm{x}$ and $\bm{F}$.

Given an arbitrary deformation, the deformed thin shell is defined on a mid-surface $\overline{\bm{\phi}}$ corresponding to the representative volume element (RVE) as $\overline{\bm{x}}=\overline{\bm{\phi}}(\overline{\bm{X}})+h\overline{\bm{n}}$. Here, $h$ is the thickness or deviation from the mid-surface and $\bm{n}$ is the surface normal. The underlying yarn structure undergoes the same deformation, i.e. $\mathbb{I}=\overline{\mathbb{I}}$, $\mathbb{II}=\overline{\mathbb{II}}$. The resulting yarn geometry is a subdivision of the macro-scale mid-surface $\bm{\phi}$ plus additional fluctuations $\tilde{\bm{u}}$ induced by the yarn relaxation process expressed as $\bm{x}=\bm{\phi}(\bm{X})+h\bm{n}+\tilde{\bm{u}}(\bm{X})$. Homogenization theory assumes that the macroscale quantities $\overline{\bm{x}}$ and $\overline{\bm{F}}$ are averaged from the microscale quantities $\bm{x}$ and $\bm{F}$ which span the macro region. In addition, the RVE is treated as constant across the micro region. These two assumptions indicate that the yarn-level fluctuations always average out over the shell mid-surface which is expressed as

(10) $$\int_{\Omega}\tilde{\bm{u}}(\bm{X})d\Omega =0,\ \int_{\Omega}\nabla\tilde{\bm{u}}(\bm{X})d\Omega =0$$

4. Method

Figure 3. Overview of our proposed yarn-shell parameter estimation method.

We now elaborate our end-to-end method for estimating shell simulation parameters corresponding to real fabrics. Our method consists of three stages: (i) initial yarn relaxation, (ii) yarn simulation on deformed surfaces and finally (iii) shell model parameter fitting. Figure 3 provides a visual overview.

1. Yarn Initialization and Relaxation. Given a real garment, we first obtain the material type through information readily found on their labels. We then visually determine the yarn structure through close-up captures and sample a fabric swatch to measure the density. With additional information provided by textile research on the tensile strength of materials, we derive the yarn model parameters. We leverage a database of knitted or woven patterns (Leaf et al., 2018) to relax the corresponding yarn pattern using periodic boundary conditions (PBC).

2. Yarn Simulation on Deformed Surfaces. We simulate the relaxed yarns to establish the shape of a deformed mid-surface, required for homogenization. Given a sampled deformation at the macro scale, we compute the resulting shape of the shell. Subsequently, we tile yarns on the deformed shell and record the resulting yarn-level energy by running a yarn-level simulation subject to homogenization constraints. This provides a map** from the deformation space at the macro level to the energy densities at the micro level, which exhibits rich physical responses that are not well represented using the macro level constitutive model.

3. Shell Parameter Estimation. Finally, we fit an orthotropic StVK shell model to the collected yarn-level deformation response data. We validate our homogenized shells for each material against quantitative measurements of real fabrics as well as full yarn simulations (without PBC) on stretching and dra** experiments. In addition, we simulate full-sized garments using our homogenized shell parameters and compare with recorded real-world footage.

4.1. Deriving Yarn Model Parameters from Real Fabrics

Provided with simple measurements of real fabrics, we estimate the stiffness matrices in Eq. 3 using a nonlinear stretching model inspired by (Sperl et al., 2022).

Figure 4. Strain-force plot. Blue: Biphasic piecewise linear stretch proposed by Sperl et al. (2020). Orange: Ours.

(11) $$E=\begin{cases}k_{1}\epsilon_{yarn}&\text{for }\epsilon_{yarn}<0\\ k_{2}\epsilon_{yarn}&\text{for }\epsilon_{yarn}>0\end{cases}\Rightarrow\nabla U_{\text{yarn}}\propto\begin{cases}k_{1}\epsilon_{yarn}^{2}\\ k_{2}\epsilon_{yarn}^{2}\end{cases}$$

The rate of change in the Young’s modulus $k_{1},k_{2}$ are derived from tensile testing on various materials including polyester, yarn and wool fibers (Serra et al., 2019). Note that we only provide a rough estimate for the ranges of the Young’s moduli, which are distinct for the materials in our scope. More accurate estimates can be obtained by conducting physical experiments on the individual yarns composing the respective fabrics.

We initialize the yarn centerlines using the yarn pattern data from Leaf et al. (2018), which contains the geometry of a single periodic repeat. Finally, the yarn radius is estimated as $r=\sqrt{\frac{\rho_{\text{shell}}p_{x}p_{y}}{\rho_{\text{yarn}}L_{0}\pi}}$. Here, $p_{x},p_{y}$ is the width and length of a single repeat and $L_{0}$ is the total rest length of yarns, all of which are calculated from the initialized yarn geometry. $\rho_{\text{shell}}$ is measured by weighing a piece of fabric and dividing the mass by the number of repeats in the swatch.

4.2. Periodic Yarns

We model 2D fabrics as a tiling of a small repeated unit which are strongly connected when tiled periodically. Our homogenization process gathers deformation responses from a large tile by simulating the periodically repeating unit, similar to Sperl et al. (2020). We model the connections of the yarn patch to adjacent ghost tiles using a constraint formulation. We find all pairs of connected yarn ends which are in the form of $(\bm{p}_{i}^{j},\bm{p}_{k}^{l},x_{\textnormal{off}},y_{\textnormal{off}})$ where $\bm{p}_{i}^{j}$ and $\bm{p}_{k}^{l}$ are the connected yarn points (e.g. $j$th sampled point of the $i$th yarn), and $(x_{\textnormal{off}},y_{\textnormal{off}})$ indicate the position of the ghost tile where $v_{k}^{l}$ resides relative to the simulated tile, i.e. $(x_{\textnormal{off}},y_{\textnormal{off}})\in\{(1,0),(-1,0),(0,1),(0,-1)\}$. Each yarn connection introduces two ghost edges $\bm{e}_{i}^{j}$ and $\bm{e}_{k}^{l-1}$ adjacent to $\bm{p}_{i}^{j}$ and $\bm{p}_{k}^{l}$ respectively , corresponding to stretching-shearing constraints $\bm{C}_{s}\left(\bm{p}_{i}^{j},\tilde{\bm{p}}_{k}^{l+1},\bm{q}(\bm{e}_{i}^{j})\right)$ and $\bm{C}_{s}\left(\bm{p}_{l}^{k},\tilde{\bm{p}}_{i}^{j-1},\bm{q}(\bm{e}_{k}^{l-1})\right)$ as well as the bending-twisting constraint $\bm{C}_{b}\left(\bm{q}(\bm{e}_{i}^{j-1}),\bm{q}(\bm{e}_{k}^{l})\right)$ following Equations 4 and 5. Note that $\tilde{\bm{p}}$ indicate positions in the adjacent ghost tiles.

4.3. Yarn Contacts

At each simulation step, yarn self-collisions and inter-object collisions are processed by detecting and culling potential collisions using bounding volume hierarchies that comprise of all edge segments of the discretized yarns and the colliding triangle meshes. Edge-edge and point-triangle contacts are identified by continuous and proximity collision detection. The contacts are then resolved through a unilateral distance constraint with position-based friction to maintain a minimal separation distance equal to the yarn diameter (Macklin et al., 2014).

4.4. Surface Tiling

We introduce additional homogenization constraints such that the periodic yarn tiling matches a deformed surface during simulation.

Mid-Surface Constraints. We build on Eq. 10 and transform all the yarn-level fluctuations to a common frame defined by the mid-surface (Sperl et al., 2020). The transformation matrix is essentially the rotation $\bm{R}$ of the surface normal from the undeformed surface to the deformed surface. For an arbitrary yarn $i$ with $(N+1)$ sampled points, we compute the rotation matrix corresponding to every yarn sampled point by taking the polar decomposition of the local mid-surface deformation $\nabla_{j}\bm{\phi}_{i}=\bm{R}_{ij}\bm{S}_{ij}$ where $\bm{S}_{ij}$ represents the in-plane deformation. The homogenization constraint is

(12) $$\bm{C}_{\text{RVE}}(\tilde{\bm{u}}) =\sum_{j=0}^{N}{\bm{R}_{ij}}^{T}\tilde{\bm{u}}_{ij}=0$$

We also make sure that the boundaries are well-aligned when being tiled. For each yarn connection involving points $(\bm{p}_{i}^{j},\bm{p}_{k}^{l})$,

(13) $$\bm{C}_{\text{RVE}\_\Gamma}(\tilde{\bm{u}})={\bm{R}_{ij}}^{T}\tilde{\bm{u}}_{ij}-{\bm{R}_{kl}}^{T}\tilde{\bm{u}}_{kl}=0$$

Here, $\bm{R}_{ij}$ remains the same across the tiling, given our assumption of constant surface gradient.

Zero-Twist Constraints. We enforce each yarn to have zero net twist and the yarn ends to have matching twist angles, to ensure the twists do not get accumulated when tiling the yarns. The twist angle is proportional to the last entry of the Darboux vector in Eq. 5. For a yarn with $N$ edges:

(14) $$C_{\text{zt}} =\sum_{i=0}^{N-2}\frac{\Im_{2}(\bar{\bm{q}}_{i}\bm{q}_{i+1})}{\Re(\bar{\bm{q}}_{i}\bm{q}_{i+1})}=0,$$

$$C_{\text{zt}\_\Gamma} =\frac{\Im_{2}(\bar{\bm{q}}_{0}\bm{q}_{1})}{\Re(\bar{\bm{q}}_{0}\bm{q}_{1})}-\frac{\Im_{2}(\bar{\bm{q}}_{N-2}\bm{q}_{N-1})}{\Re(\bar{\bm{q}}_{N-2}\bm{q}_{N-1})}=0.$$

4.5. Yarn-Level Simulation

We execute yarn-level simulation in two phases. The first phase brings the yarns to a relaxed state. The second then minimizes the yarn energy density subject to the homogenization constraints.

Yarn Relaxation. We minimize the energy in Eq. 3 with slight modifications to the yarn constraints as shown in Eq. 15. We introduce a small amount of yarn shrinkage by setting the rest lengths of all yarn segments to be shorter than their initial lengths, so that the relaxed yarns are touching while the overall size of the patch remains unchanged. In practice, we set a shrinking factor $r_{s}$ of 0.8 for knitted fabrics and 0.9 for woven fabrics. In addition, we assume that the yarns are naturally straight.

(15) $$\bm{C}_{s}^{\prime}(\bm{p}_{i},\bm{p}_{i+1},\bm{q}_{i}) =\frac{\bm{p}_{i+1}-\bm{p}_{i}}{r_{s}l_{0}}-\bm{d}_{3}(\bm{q}_{i})\ \text{where}\ r_{s}\in(0,1)$$

$$\bm{C}_{b}^{\prime}(\bm{q}_{i},\bm{q}_{i+1}) =\frac{\Im\left(\bar{\bm{q}}_{i}\bm{q}_{i+1}\right)}{\Re\left(\bar{\bm{q}}_{i}\bm{q}_{i+1}\right)}.$$

Homogenization. We use the relaxed yarns from the first phase as the new rest state. From this, we compute the new rest lengths and curvatures. Subsequently, we tile the yarns on a deformed surface following Sec. 4.4 sampled from a range of deformations detailed in Sec. 4.6. In contrast to the compliant yarn constraints, the homogenization constraints are modeled as hard constraints in our simulations. We minimize the energy (Eq. 3) and record the resulting energy density:

(16) $$\overline{U}_{\text{yarn}} =\frac{1}{A}\min_{\tilde{\mathbf{u}}}U_{\text{yarn}}\ \text{where }A=p_{x}p_{y}$$

$$\ s.t.\ \bm{C}_{\text{contact}} =\bm{C}_{\text{RVE}}=\bm{C}_{\text{RVE}\_\Gamma}=\bm{C}_{\text{zt}}=\bm{C}_{\text{zt}\_\Gamma}=\bm{0}.$$

Although we employ a compliant constraint formulation in our implementation, we only obtain the minimized energy from the final quasi-static state. Any alternative quasi-static solver would suffice.

4.6. Fitting

We collect the yarn responses from two distinct sets of experiments where we apply either in-plane or out-of-plane deformations. For membrane deformations, we sample pure weft/warp stretching ($\mathbb{I}_{12}=0$, $\mathbb{I}_{11}\neq 1$ and/or $\mathbb{I}_{22}\neq 1$) and pure shearing ($\mathbb{I}_{11}=\mathbb{I}_{22}=1$, $\mathbb{I}_{12}\neq 0$) to avoid ambiguity in the off-diagonal entries of the transformation matrix. For bending deformations, in addition to sampling orthogonal sets of singly-bent surfaces where only one of $\mathbb{II}_{11}$ and $\mathbb{II}_{22}$ is non-zero, we also sample bending along the bias direction with $|\mathbb{II}_{12}|=|\mathbb{II}_{11}|=|\mathbb{II}_{22}|\neq 0$. Given the minimized energy densities corresponding to all sampled surface deformations, we use the weighted least squares method to optimize the compliance matrices for stretching/shearing and bending for the shell model. Our weights follow a Gaussian distribution centered at the zero strain to produce a better fit at small strains as these are more common due to the limited extensibility of yarns. Given $N$ samples,

(17) $$\epsilon=\min_{\bm{\gamma}} \sum_{i=0}^{N-1}w_{i}(\overline{U}_{\text{shell}}^{i}(\bm{\gamma})-\overline{U}_{\text{yarn}}^{i})^{2}\ \text{where}$$

$$\overline{U}_{\text{shell}}^{i}(\bm{\gamma}) =\frac{1}{A}(U_{\text{stretch}}^{i}+U_{\text{bend}}^{i})$$

$$=\frac{1}{2A}\overline{\bm{C}}_{s}^{T}\overline{\bm{\alpha}}^{-1}_{s}\overline{\bm{C}}_{s}+\frac{1}{2A}\overline{\bm{C}}_{b}^{T}\overline{\bm{\alpha}}^{-1}_{b}\overline{\bm{C}}_{b},$$

$\bm{\gamma}=(s_{00},s_{11},s_{22},s_{01},b_{00},b_{11},b_{22})$ is the set of stiffness coefficients for stretch and bending, and $\overline{\bm{\alpha}}^{-1}_{s},\overline{\bm{\alpha}}^{-1}_{b},\overline{\bm{C}}_{s},\overline{\bm{C}}_{b}$ are defined in Section 3.3. We do not jointly sample membrane and bending deformations, i.e. $\mathbb{I}=\bm{0}$ if $\mathbb{II}\neq\bm{0}$ and vice versa.

5. Results

We estimate the shell parameters for five real fabrics comprised of different yarn structures and materials, as shown in Figure 5. We derive the yarn parameters for three distinct materials: wool, cotton and polyester, following Section 4.1, and provide the exact parameters and simulation setup in a supplementary document. The virtual cloth is visualized using a surface-based shading model (Zhu et al., 2023) with the required normal and tangent information being directly inferred from the relaxed woven and knitted patterns.

Figure 5. The garments used in our results. The bottom row shows close-up images of the fabrics.

Figure 6. Fabric analyzer (FAB) measures force exerted by the fabric as it is stretched. The machine is enhanced with ArUco markers (Bradski, 2000) to determine the ratio between stretch (horizontal length) and orthogonal compression (vertical length drawn in red).

Stretch Test. We validate our results against real world fabric stretch tests, where a piece of rectangular fabric is clamped at two ends, stretched and the resulting forces on the clamps measured. We use the commercially available fabric analyzer (FAB) (Browzwear, 2024) to perform stretch tests with $8\times 5$ cm and $2\times 5$ cm pieces of fabric cut along the warp, weft and bias (at 45 degrees) directions. We augment the FAB with additional markers to track the strain and capture the amount of visual deformation (compression) along the orthogonal direction as pictured in Fig. 6. We then reproduce the same scenario in simulation at the sheet level and plot the resulting boundary forces and compression ratios against real world measurements from the FAB as shown in Fig. 10. This test is critical in confirming that our pipeline can predict real fabric behaviors. Furthermore, we use the same stretch test to demonstrate in Fig. 7 how thicker yarns produce sheets with larger stresses, a phenomenon previously unexplored in homogenized cloth.

Figure 7. The force plot of the stretch test for the estimated cotton jersey material (red) and the same material with 1.5$\times$ thicker yarns (blue). As expected, the thicker material exhibits higher stress during stretching.

Drape Tests. The stretching quality of materials can be verified numerically by comparing force measurement as in Fig. 10 (right column). However, for bending, the measured forces are very small and require very sensitive force sensors. While the FAB produces some force output for bending tests, it is unreliable for some materials as reported by Browzwear’s software. To validate bending parameters, we drape a piece of cloth on a sphere, reconstruct geometries using a standard photogrammetry software (e.g. Agisoft Metashape) and compare the result to the same scenario reproduced in simulation as shown in Fig. 9. We also perform dra** against a wall and compare to captures of real drapes, demonstrated in the same figure.

Full Body Validation. To demonstrate how our models are realized in real world applications, we simulate clothing on a full body. An assembled virtual t-shirt and pants are draped on an animated body and compared to a reference video of the same outfit as shown in Fig. 9. Our simulated garments exhibit realistic folds and creases in response to the body movements.

6. Limitations and Future Work

Edge curling effects. While we use top-view captures for estimating compression ratio, they are insufficient for estimating in-plane compression data due to curling effects at the boundaries of the stretched fabric. The side view close up in Fig. 6 shows fabric at the edges bending out of the view from the top-view camera. To remedy this, we propose to add heterogeneous curved rest shape near the boundary and measure curling amount from a side-view capture. We believe this discrepancy is in part responsible for the mismatch in compression ratio shown in the left column in Fig. 10 for jersey fabrics.

Internal friction. We apply a uniform friction coefficient for all materials in our yarn simulation, while the internal frictions of different materials are vastly divergent in real fabrics. Moreover, we do not estimate frictions between yarn fibres, which causes hysteresis effects in deformed fabrics. Consistent with previous work, we discover a considerable amount of yarn behavior caused by friction, especially in the knitted fabrics. Internal frictions also cause unwanted yarn sliding and lead to mismatches between simulated results and measured data, which explain the difference between compression ratios on cotton and polyester knits, since polyester yarns have fewer dangling fibers as shown in the inset.

Fabric materials. Our yarn-level parameter estimation assumes that the yarns are homogeneous, i.e. the entire fabric is 100% composed of the same yarn material. In practice, however, a blend of organic and synthetic materials is more commonly used in clothing such as pants to improve the durability while preserving comfort. Our experiment uses twill pants made of 98% cotton and 2% spandex, which exhibit higher contraction and flexural rigidity compared to pure cotton (Almetwally and Mourad, 2014), while our estimated shell parameters assume 100% cotton. Furthermore, the fabrication process alters the physical properties of the fabric. For example, manufacturing pants requires a finishing process of pressing and ironing, which changes the stretching/shrinkage properties of the underlying fabric at high temperatures (Islam et al., 2019). Both factors contribute to the significant quantitative mismatch between simulated and real twill pants in the stretch test. Future work entails distinguishing between different yarn materials in a composite via CT scanning and accounting for the post-processing stage in the yarn simulation.

7. Conclusion

We propose an end-to-end method for estimating shell model properties corresponding to real fabrics. Through simple measurements of the fabric swatches and close-up captures, we effectively obtain information about the yarn structures and physical properties such as yarn thickness. Combined with textile research literature that provide approximations for the tensile strengths of different materials, we initialize yarn-level models with properties derived from the real fabrics, which are preserved through our yarn-shell homogenization pipeline. Subsequently, we sample a range of surface deformations given by the first and second fundamental forms, record the responses of periodic yarn patterns on each sampled deformed surface in terms of energy density, and optimize all the stiffness parameters of an orthotropic StVK shell model, including off-diagonal bending terms representing warp-weft coupling. We validate our result quantitatively and visually against measurements and captures of real fabrics on stretch tests and drape experiments, and demonstrate a good quantitative approximation and qualitative match with real fabrics such as reproducing the anisotropic behaviors in satin and twill weaves, despite the lack of an inverse process. Our homogenized shells not only establish the key characteristics of each yarn pattern, but also showcase the effects of yarn material properties. Using only physics-based simulation, our method provides reasonably accurate initial estimates for parameter estimation pipelines requiring more comprehensive knowledge of real fabrics.

Figure 8. We show two different drape setups where we compare simulated results to a real reference, one where the swatch is draped on a sphere and a second one where the top corners are fixed. Given the use of very limited information about the fabrics, we do not expect our drapes to match the reference exactly. However, our estimated materials produce qualitatively similar results to the references such as highly anisotropic bending in satin and twill fabrics, and are noticeably distinct for different material types.

Figure 8. We show two different drape setups where we compare simulated results to a real reference, one where the swatch is draped on a sphere and a second one where the top corners are fixed. Given the use of very limited information about the fabrics, we do not expect our drapes to match the reference exactly. However, our estimated materials produce qualitatively similar results to the references such as highly anisotropic bending in satin and twill fabrics, and are noticeably distinct for different material types.

Figure 9. Full body cloth simulation of the cotton twill pants and polyester jersey t-shirt with parameters estimated using our pipeline. The simulated pants exhibit dramatic creases in poses 3, 4 and 5 matching the reference. The simulated T-shirt exerts fairly natural wrinkles in response to the arm movements and body twists, which are most evident in poses 3 and 5.

Figure 10. Compression ratio (left column) and boundary force (right column) plots for 5 different materials comparing captured data from the FAB (red dots) against shell simulation results (solid blue curves). It is remarkable to see that simulated shell materials capture the key characteristics of yarn structures such as decreased stiffness/increase compression ratio in the bias direction for woven patterns. In some cases like the cotton jersey t-shirt, the compression ratio (determined by the slope of the curve) matched the data remarkably well, in spite of there being no inverse process to match fabric measurements in our pipeline.