Visual Comput (2005) 21: 506–519

DOI 10.1007/s00371-005-0347-6 INVITED PAPER

Nadia Magnenat-Thalmann From early draping to haute couture models:

Pascal Volino
20 years of research

Published online: 1 September 2005 Abstract Simulating the complex prototyping of garments along drape
© Springer-Verlag 2005 fashion garments of haute couture can motion and comfortability tests on
only be reached through an optimal animated postures. These techniques
combination of modeling techniques have been successfully used to bring
and numerical methods that combines haute couture garments from early
high computation efficiency with the draping of fashion designers, to
versatility required for simulating be simulated and visualized in the
N. Magnenat-Thalmann (u) · P. Volino intricate garment designs. Here we virtual world.
MIRALab, University of Geneva, CH-1211 describe optimal choices illustrated
Switzerland by their integration into a design and Keywords Cloth simulation · Virtual
thalmann@miralab.unige.ch simulation tool that allow interactive garments · Fashion design

1 Introduction of efficient mechanical simulation models, which can ac-

curately reproduce the specific mechanical properties of
Traditional fashion design involves substantial work from cloth. However, cloth is by nature highly deformable, and
garment stylists and fashion designers to create garment specific simulation problems arise from this fact. First, the
patterns, then build garment models on mannequins to mechanical representation should be accurate enough to
assess the fitting and drape, possibly involving correc- deal with the nonlinearities and large deformations occur-
tions of the design. Creativity is quite limited by the large ring at any place in the cloth, such as folds and wrinkles.
amount of work required to build garment prototypes. Vir- Moreover, the garment cloth interacts strongly with the
tual simulation would greatly help fashion designers not body that wears it, as well as with the other garments
only to speed up their creative work, but also to bring their of the apparel. This requires advanced methods for ef-
creations to life through high-quality mechanical simu- ficiently detecting the geometrical contacts constraining
lation on animated characters (Fig. 1). Old heritage gar- the behavior of the cloth and integrating them into the
ments may also be brought to life through simulation of mechanical model (collision detection and response). All
ancient cloth materials. these methods require advanced and complex computa-
tional methods where most important key issues remain
computation speed and efficiency. For real-time applica-
1.1 The problematic of garment simulation tions, however, only specific approximation and simplifi-
cation methods allow the computation of garment anima-
Virtual garment simulation is the result of a large combi- tion, trading off some of the mechanical accuracy of the
nation of techniques that have also dramatically evolved result for visual realism.
during the last decade. Unlike the mechanical models
used for existing mechanical engineering for simulating 1.2 Early developments in garment simulation
deformable structures, many new challenges arise from
the highly versatile nature of cloth. The central pillar of Garment simulation, which started in the late 1980s with
garment simulation obviously remains the development very simple models such as Weil’s approach [28], has ben-
 From early draping to haute couture models: 20 years of research 507

Fig. 1. Recreating animated haute couture garments from sketches

efitted considerably from the increasing performance of

computer hardware and tools as well as the development
of specific simulation technologies that have led to im-
pressive applications not only in the field of simulation of
virtual worlds but also as design tools for the garment and
fashion industry.
In the field of computer graphics, the first applica-
tions for mechanical cloth simulation appeared in 1987
with the work of Terzopoulos et al. [20, 21] in the form of
a simulation system relying on the Lagrange equations of
motion and elastic surface energy. Solutions were obtained
through finite difference schemes on regular grids. This
allowed simple scenes involving cloth to be simulated,
such as the accurate simulation of a flag or the draping
of a rectangular cloth. However, the first really accurate
garment simulation applications started to appear in 1990
(Fig. 2) with the consideration of many other technolo-
gies complementing cloth simulation [2, 14], such as body Fig. 2. “FlashBack”: Early virtual garments used context-dependent
modeling and animation and collision detection and re- simulation of simplified cloth models
sponse [29]. These applications were innovative in that
they provided the first virtual system allowing virtual gar-
ment patterns to be sewn together around a model. * The size of garments, which can be 1 m long and yet
Since then, most developments have been aimed at op- need to be simulated at millimeter accuracy.
timizing the accuracy and efficiency of cloth simulation * The intricate and highly variable shape of garments,
methods, along the lines of developing actual applications which interact through complex contact patterns with
and commercial products. the body (which is itself a complex deformable entity)
as well as with other garments.
1.3 Challenges * The highly deformable nature of cloth, which trans-
lates very subtle mechanical variations into large drap-
Facing the stringent requirements of garment designers ing and motion variations that modify completely the
not only in terms of realism but also quantitative mechan- visual appearance of garment models.
ical accuracy of draping and motion, the complexity and * The highly intricate anisotropic and nonlinear mechan-
diversity of real garment models make it a real challenge ical behavior of garments, requiring accurate measure-
to design a garment simulation system [26, 27]. Among ment, modeling, and complex numerical methods for
the main difficulties are: their resolution.
508 N. Magnenat-Thalmann, P. Volino

In the following section, we review how state-of-the- the deformation of the surface for given values of the in-
art techniques are combined to address this challenging terpolation parameters. An equation system based on the
goal. energy variation is then constructed with these degrees
of freedom. Surface continuity between adjacent elements
imposes additional constraint relationships. A large sparse
linear system is built by assembling successively the con-
2 Garment simulation techniques tributions of all the elements of the surface and then solved
using optimized iterative techniques, such as the conjugate
At the core of a garment design and simulation system,
gradient method.
various components aim at providing an accurate repre-
Finite elements have only had a marginal role in cloth
sentation of a virtual cloth, computational features needed
simulation. The main attempts are described in [5, 10, 11].
to reproduce accurately its mechanical behaviors, and, fi-
Most implementations focus on the accurate reproduction
nally, the design features necessary to create complex gar-
of mechanical properties of fabrics but restrict the appli-
ments. The following sections detail these features.
cation field to the simulation of simple garment samples
under elementary mechanical contexts, mostly because
2.1 Mechanical simulation of complex cloth materials of the huge computational requirements of these models.
Furthermore, accurate modeling of highly variable con-
The accurate reproduction of the mechanical behavior of straints (large nonlinear deformations, highly variable col-
cloth has always been a key issue in garment simulation. lisions) is difficult to integrate into the formalism of finite
The mechanical behavior of cloth is usually measured elements, and this sharply reduces the ability of the model
using standardized protocols, such as the Kawabata Eval- to cope with the very complicated geometrical contexts
uation System (KES) or the simpler FAST method, which that can arise in real-world garment simulation on virtual
are based on the experimental measurement of strain– models.
stress curves for elongation, shearing, and bending on nor-
malized samples of fabric (Fig. 3). Different representa- 2.1.2 Particle systems
tions of the cloth surface mechanics then allow the virtual
reproduction of the behavior of cloth. An easier and more pragmatic way to perform cloth simu-
lation is to use particle systems. Particle systems consider
2.1.1 Finite elements the cloth to be represented only by the set of vertices that
constitute the polygonal mesh of the surface. These par-
Well known in mechanical engineering, the finite element ticles are moved through the action of forces that represent
method considers the cloth surface as being discretized in the mechanical behavior of the cloth, which are computed
interpolation patches for a given order (bilinear, trilinear, from the geometric relationships between the particles that
quadrilinear) and an associated set of parameters (degrees measure the deformation of the virtual cloth. Among the
of freedom) that give the actual shape to the interpolation different variations of particle systems, the spring-mass
surface over the element. From the mechanical properties scheme is the simplest and most widely used. It consid-
of the material the mechanical energy is computed from ers the distance between neighboring particle pairs as the

Fig. 3. Kawabata curves for tensile, shear, and bending

 From early draping to haute couture models: 20 years of research 509

Fig. 5. Advanced garment design requires simulating anisotropic

bending stiffness, with a possible rest curvature defined on the sur-
face (left). Rest curvature may also be defined along precise lines
(center). Lines may also carry additional stiffness with their own
Fig. 4. Drape accuracy between a simple spring-mass system along custom rest length (right). All these features bring lots of possibil-
the edges of a triangle mesh (left) and an accurate particle system ities for designing complex garment models
model computed on triangle surfaces (center). Color scale shows
deformation. The spring-mass model exhibits inaccurate local de-
formations, along with an excessive “Poisson” behavior. This is elements in weft, warp, and shear modes according to the
not the case with the accurate model, which may still model the current 3D position of the triangle vertices. Then, the ac-
“Poisson” effect if needed (right, with a Poisson coefficient of 0.5). tual strain-stress curves (modeled as piecewise polynomial
The spring-mass model is also unable to simulate anisotropic or
nonlinear models accurately splines) are used to compute the resulting strain on the
element, and equivalent forces are applied on its vertices.
While a particle system, this scheme offers the accuracy of
only deformation measurement and interaction source rep- first-order finite elements. Bending is also included in the
resenting the internal elasticity of the cloth. model as a torque effect created between elements. One
Particle systems are among the simplest and most effi- major advantage of this scheme is that it does not rely on
cient ways to define rough models that compute highly de- the use of regular grids for modeling the cloth surface, and
formable mechanical systems such as cloth with computa- this allows a good degree of freedom in the design of the
tion times small enough to integrate them into systems for pattern shapes of the garment.
simulating complete garments on virtual bodies. Among The mechanical representation of cloth also has to be
the main contributions on particle system models, early further extended to support advanced mechanical features
works considered simple viscoelastic models on regular that are needed in complex garment design (Fig. 5).
grids with applications for draping problems with sim- While the choice of particle system scheme dictates the
ple numerical integration schemes [19]. Accurate models accuracy of the mechanical representation of the proper-
started with Breen et al. [2] on modeling the microstruc- ties of the cloth, any of the resulting equations have to
ture of cloth using parameters derived from KES behav- be integrated efficiently with advanced numerical methods
ior curves and integration based on energy minimiza- that determine the computation performance.
tion. However, such accurate models required consider-
able computation for solving problems that were restricted 2.2 Numerical integration
to draping. On the other hand, more recent models trade
accuracy for speed, such as the grid model detailed by While various models can be used to compute the force
Provot et al. [18], which additionally includes geometric applied on each particle given its position and speed, these
constraints for limiting large deformations of cloth. Ad- forces have then to be integrated in time to obtain the pos-
ditional contributions from Eberhardt et al. [8], who use ition and speed of the particle for the following timesteps
the simulation of KES parameters and comparison of the using methods related to the integration of ordinary dif-
efficiency of several integration methods. Advanced sur- ferential equation systems. Most recent approaches, how-
face representations are used in [6], where the simulation ever, focus on improvements of the numerical integration
model and collision detection take advantage of the hier- methods to improve the efficiency of the simulation.
archical structure of subdivision surfaces. Modeling ani- Explicit integration methods are the simplest methods
mated garments on virtual models is the specific aim of the available for solving first-order ordinary differential sys-
work described by Volino et al. [22, 23], which investigate tems. They consider the prediction of the future system
improved spring-mass representations for better accuracy state directly from the value of the derivatives. The best
of surface elasticity modeling on irregular meshes. known techniques are the Runge–Kutta methods. Among
An accurate particle system scheme is proposed them, the fast but unstable and inaccurate first-order Euler
by [22]. It computes precisely the stresses of triangle method, used in many early implementations, considers
510 N. Magnenat-Thalmann, P. Volino

Fig. 6. The numerical accuracy of integration methods can be evaluated through the energy dissipation they produce on undamped me-
chanical systems (a balancing square of cloth attached along one edge). Runge–Kutta is the most accurate at the expense of very high
computation time, whereas implicit methods allow one to trade off numerical accuracy for computation speed

Fig. 7. Intricate garments require robust and accurate collision processing to identify and handle contact points between multilayer cloth
and body

the future state as a direct extrapolation from the current “backward” Euler computation would return the initial
state and the derivative. Higher-order and more accurate state. It performs the computation using not the deriva-
methods also exist, such as the second-order midpoint tive at the current timestep but the predicted derivative for
method, used for instance in early models by Volino et the next timestep. Besides the inverse Euler method, other,
al. [22], and the very accurate fourth-order Runge–Kutta more accurate higher-order implicit methods exist, such as
method, used for instance by Eberhardt et al. [8]. the inverse midpoint method, which remains quite simple
Besides considerations of accuracy, stability, and ro- but exhibits some instability problems. A simple solution
bustness, there are other key factors to consider. For most is to interpolate between the equations of the Euler and
situations encountered in cloth simulation, the numerical midpoint methods, as proposed by Volino et al. [24–26].
stiffness of the equations (stiff elastic forces, small surface Higher-order methods, such as the Rosenbrook method,
elements) require that the simulation timesteps be small however, do not exhibit convincing efficiencies in the field
enough to ensure the stability of the system, and this limits of cloth simulation. Multistep methods, which perform
the computation speed much more than accuracy consid- a single-step iteration using a linear combination of sev-
erations. Adequate timestep control is therefore essential eral previous states, are other good candidates for a good
for an optimal simulation. A common solution is to use the accuracy–stability compromise. Among them, the second-
fifth-order Runge–Kutta algorithm detailed in [17], which order backward differential formula (BDF-2) has shown
embeds integration error evaluation used for tuning the some interesting performances, as used by Eberhardt,
timestep adaptively [23]. Hauth et al. [9, 12] and Choi et al. [4].
In order to circumvent the problem of instability, im- Whatever variation chosen, the major difficulty in
plicit numerical methods are being used. For cloth simula- using implicit integration methods is that they involve
tion, this was first outlined by Baraff et al. [1]. The most the resolution of a large and sparse linear equation sys-
basic implementation of an implicit method is the Euler tem for each iteration, constructed from the Jacobian
step, which considers finding the future state for which matrix of the particle forces against their position and
 From early draping to haute couture models: 20 years of research 511

Fig. 8. Robust collision detection and response are necessary for simulating intricate cloth behaviors

speed. A commonly used simplification involves lineariza- on the simulation context, considerations might be the fol-
tion of the mechanical model so as to obtain a linear lowing:
approximation of the matrix that does not evolve along
time and on which initial construction and preprocess- * For draping applications, the only matter is for the
ing allows efficient resolution methods to be used, as for method to converge as quickly as possible to the rest
example in Kang et al. [13], or even the matrix inverse state, without bothering about the realism of the cloth
to be precomputed as done by Desbrun et al. [7]. A fur- motion. Using the backward Euler method with fairly
ther simplification is to suppress completely the need to large timesteps seems to be the fastest and most robust
compute the matrix using an adapted approximation em- approach for this, particularly when using nonlinear
bedded directly in an explicit iteration. A big drawback mechanical models. This does not require the model-
of all these methods results from the approximation of ing of any damping. However, adding some viscosity
the matrix that cannot take into account the nonlineari- to the model may speed up convergence.
ties of the model (mostly those resulting from the change * When realistic motion of the cloth is expected, the
of orientation of the surface elements during the simula- second-order implicit midpoint method or the third-
tion). While this is acceptable for draping applications, order accurate BDF-2 variation will certainly help in
animations behave usually poorly because of excessive obtaining accurate cloth motion. As demonstrated in
numerical damping, which also increases as the timestep this work, these methods perform better through the
becomes large. addition of a moderate stabilization viscosity than the
The best numerical method for actually resolving the inverse Euler method or regular BDF-2. The overall
linear system seems to be the conjugate gradient method, computational gain might be twofold in some cases.
as suggested by Baraff et al. [1], with various variations * If perfect reproduction of the dynamic behavior of
and preconditioning schemes depending on how the me- the cloth is expected, the explicit fifth-order Runge–
chanical model is formulated and geometrical constraints Kutta method with adaptive timestep is still the high-
of the cloth integrated. accuracy method to consider. It does require including
There is, however, no universal solution for integrating some damping in the mechanical model. However, this
the equations of a mechanical model of cloth. Depending comes at the high computational cost needed for com-
512 N. Magnenat-Thalmann, P. Volino

puting exactly all the elements of the mesh accurately, The collision processing is therefore general enough for
which may become very high for stiff materials and handling contacts between the several garments of a com-
refined meshes. This is usually out of reach for applica- plex dress style as well as interactions between complex
tions that require the accurate simulation of complete fold patterns when animating gestures of wide amplitude.
garments with actual mechanical properties of fabrics. The model is also accurate enough for reproducing friction
behavior with precision, allowing, for example, pants to
While viscous damping is fairly easy to integrate into hold to the waist with friction alone during character mo-
a mechanical model for improving the numerical behav- tion, without “cheating” using geometrical attachments.
ior of the integration, its actual measurement on real fabric Good stability allows the simulation of complete multi-
materials still remains a challenge that needs to be ad- layer garments with millimeter collision thickness despite
dressed. This is made necessary for tackling the next step large cloth speed and tension produced by complex char-
of garment simulation, in its move from draping applica- acter motion.
tions to accurate simulation of the cloth motion, a main
factor in the beauty of garments and the people who wear
2.4 A system for interactive garment design
them.
Powerful cloth simulation methods are not sufficient for
2.3 Collision detection and response designing beautiful fashion garments, which are usually
made through a combination of highly complex patterns.
Although there are other approaches to removing the col- It is also important to give the pattern designer a tool for
lision detection process in animation including generating performing garment creation in a way that allows high-
a unified mesh for garments and the body [16], it is in- level interaction, for testing the fitting and drape, dynami-
deed one of the most time-consuming tasks when it comes cally adjusting the shape of the patterns in a trial-and-error
to accurately simulating virtual characters wearing whole scheme, just like a real designer would do with a real man-
garments [15]. The goal is to detect the contacts between nequin.
regions of the cloth with the body or other parts of the
cloth to simulate reaction and friction forces. Without col- 2.4.1 Interactive pattern editing
lision processing, all objects of the scene would behave
independently and interpenetrate. This is particularly true The high level of interactivity required by these features
when the goal is to simulate the intricate multilayer design necessitates simultaneous computation of the 3D garment
of complex fashion garments. updated immediately to each design modification done to
Many different methods exist for addressing the prob- the patterns. The solution is to provide a dual view of the
lem of collisions. In state-of-the-art systems [26], this garment, featuring both the 2D view of the pattern shapes
task is usually performed through an adapted bounding- cut on the fabric and the 3D view of the garment worn
volume hierarchy algorithm, which uses a constant discrete- by a virtual character, with tight synchronization (Fig. 9).
orientation-polytope hierarchy constructed on the mesh, Any editing task carried out in one view is directly dis-
and optimization for self-collision detection using curva- played in the other view.
ture evaluation on the surface hierarchy. This algorithm The system features a fast constrained Delaunay trian-
is fast enough to allow for full collision and self-collision gulation scheme that allows the discretization of complex
detection between all objects of a scene with accept- patterns described as polygonal lines of control points
able impact on the processing time (which rarely exceeds (2D locations on the fabric). The system allows variable
20% of the total time). Thus, body and cloth meshes discretization densities over the mesh, as well as size
are handled totally symetrically by the collision detec- anisotropy (elements elongated in a given direction), for
tion process, ensuring perfect versatility of the collision representing adaptively complete garments from large sur-
handling between the body and the several layers of gar- faces to intricate details.
ments [23]. The interactivity of the system is based on two main
Collision response can be handled using a geometrical features:
scheme based on the correction of mesh position, speed,
and acceleration [24]. This scheme ensures good accu- * Mesh mapping update: The 2D displacement of any
racy and stability without the need for large nonlinear control point of the pattern shape on the cloth surface
forces that alter the numerical resolution of the mechanical immediately updates the mesh of that pattern on the
model. This model simulates contact forces through a per- cloth, while leaving the 3D drape position of the cloth
fectly damped reaction model associated to a Coulombian constant. To obtain this, each vertex of the mesh keeps
(solid) friction model. track of a weighted sum of the pattern control points,
This collision model ensures full mesh-to-mesh col- which is computed during the triangulation process.
lision response, which can handle very complex multi- This allows any measurement or shape editing to be
layer collision configurations involving several surfaces. directly taken into account by mechanical simulation
 From early draping to haute couture models: 20 years of research 513

* Preview of fabric deformations and tensions along any

weave orientation.
* Preview of pressure forces of the garment on the body
skin.
* Immediate update of these evaluations according to
pattern reshaping and sizing, fabric material change,
and body measurement and posture changes.

Dynamic surface remeshing allows the best compro-

mise between accuracy and computation speed to be se-
lected adaptively according to the needs of the garment
designer. For instance, while the garment assembly pro-
cess can be carried out in a matter of seconds using an ap-
proximate mechanical model on a rough garment surface
mesh, the garment designer may then switch to a more
accurate model for tasks such as accurate draping and
Fig. 9. A garment design system should offer high-quality garment comfort evaluation. The model is still efficient enough
simulation, along with highly interactive 2D–3D design patterns to react interactively to design changes with garments
and preview tools allowing for the efficient design of complex gar-
ment models with many features such as seams, buttons, pockets, made of tens of thousands of polygons, an accurate drap-
belts, etc. ing being obtained in a few minutes. Practical geomet-
ric accuracy is roughly limited by using 5-mm elements.
Using time-accurate computation on animated characters,
without heavy recomputation, for immediate feedback a high-quality catwalk is computed in a matter of a few
of any pattern sizing adjustment. hours.
* Mesh topology reconstruction: When the topology of
the pattern mesh is changed (rediscretization, new fea-
tures, etc.), the 3D drape position of the new mesh 3 From sketches to 3D simulation
is automatically recomputed from the drape position
of the old one. During this process, advanced algo- Relating a powerful and versatile cloth simulation system
rithms compute, for each mesh vertex of the new to an efficient interactive pattern design interface allows
mesh, the location of the surface of the old mesh the use of software for numerous applications involving
having identical 2D coordinates on the fabric. Ex- accurate virtual clothing, such as haute couture fashion de-
trapolation methods are used for computing the loca- sign of virtual heritage applications.
tion of vertices located outside the old surface. This
allows pattern design changes (new features, darts, 3.1 The Robert Piguet exhibition in Yverdon
seams, etc.) to be added and modified without having
to reassemble and redrape the garment on the virtual For the Robert Piguet exhibition taking place in Yverdon
body. les Bains this summer in Switzerland, MIRALab has cre-
ated 18 haute couture garments virtually from sketches
Taken together, these techniques greatly enhance the by Marc Bohan, Serge Guérin, and Hubert de Givenchy,
workflow of garment prototyping. For instance, an ini- former assistants of the famous couturier (Fig. 11). This
tial garment could be quickly draped over a character exhibition has been organized by the Swiss Fashion Mu-
using a rough mesh. Then, the designer could enhance seum with the collaboration of MIRALab.
the pattern shapes and, while updating the mesh mapping, Today’s fashion industry illustrates new designs with
automatically alter the mechanical state of the draped technical sketches and detailed description sheets to com-
garment, changing the draping shape. Once the garment municate easily with overseas production locations. The
design is ready, a high-accuracy drape is automatically meaning of an aesthetic haute couture drawing is a dis-
produced using topology reconstruction with a refined tinctive one: it is not only an information vehicle but can
mesh. be seen as a work of art, visualizing cultural aspects of
a garment corresponding to certain époques. Haute cou-
2.4.2 Interactive garment prototyping and fitting ture designers are characterized by their particular draw-
ing styles, visible in the different types of sketches of
Combined with the accuracy and speed of the proposed Bohan, Givenchy, and Guérin. Shown in two dimensions
mechanical simulation engine, tasks such as comfortabil- from only one side, the sketch leaves considerable space
ity evaluations are open to the garment designer (Fig. 10) for individual interpretation of the garment in three dimen-
through the addition of several visualization tools, such as: sions.
514 N. Magnenat-Thalmann, P. Volino

Fig. 10. Virtual prototyping: Displaying weft constraints on an animated body (from standing to sitting). Element size is 5 mm

Fig. 11. Sketches from Marc Bohan (1946), Hubert de Givenchy (1946), Serge Guérin (1950)

The virtual garments for the Robert Piguet exhibition ing it as a detailed 3D garment. With the desired 3D shape
were composed using MIRALab’s virtual garment cre- in mind, the 2D pattern can be designed according to pat-
ation software Fashionizer, which imitates the real-world tern construction rules. If the original 2D patterns of the
tailoring process: the patterns are initially designed and sketches are available, they canbe digitized on a digital-
cut in 2D space, placed around a virtual mannequin in ization board and the obtained electronic data directly
3D space, sewn together to make a completed article of imported into Fashionizer from the CAD system. The di-
clothing, and finally simulated according to the physical mensions of the 2D patterns are also determined by the
properties of the fabric and its environment. During the an- size of the body to be dressed.
imation, the garment follows the movement of the virtual Over time, body shapes have changed due to vary-
mannequin. ing lifestyles (including practice of sports, changes in
diet, etc.). Hence, the body shape of an average woman
3.2 Design of 2D patterns of the 1940s was different from today’s body. The look
of a body silhouette in sketches is in addition a product
Taking a 2D pattern as a base is the simplest way to of the beauty ideal of a given time period. The post-
obtain a precise, exact, and measurable description of war period was characterized by the typical wasp waist.
a 2D surface, which is the representation of the vir- In virtual space it becomes possible to model the body
tual fabric. One garment is composed of several 2D according to overstated temporal tendencies since the
surfaces, the single-pattern pieces. Fashionizer is ba- body is merely a hull of polygons. In the range between
sically an editor of polygons representing clothing in a very realistic and a more abstract body, a combina-
a number of 2D polygons connected by seam lines tion of both was chosen for the virtual mannequin for
(Fig. 12). the Robert Piguet exhibition. The waist is overly slim,
To derive the exact 2D pattern shape from the 2D mixed with a typical feminine curved body from the
sketch, first the drawing needs to be interpreted by imag- 1940s.
 From early draping to haute couture models: 20 years of research 515

Fig. 12. Creation of 2D patterns

Today’s pattern construction methods are affected by

new developments in the field of textiles and textile pro-
cessing methods. For instance, elastic materials such as
Lycra, only introduced into the textile market in 1962, sud-
denly allowed tight-fitting and comfortable garments. This
progress of technology over the past 60 years needed to be
considered in the creation of the 2D pattern for the Robert
Piguet gowns as well, as it influences the global fit and
appearance of a garment.

3.3 3D pattern placement on the body

The patterns obtained in the preceding part are displayed
in Fashionizer on a grid, the surface. The planar patterns
are placed around the virtual body (Fig. 13). A manual
placement is implemented with an automatic function to
bring the pattern to a position closest to the body surface.
Considering that the seams will gather the edges of each
pattern together, an approximate initial positioning is ne-
cessary. The space between two seamlines should be as
small as possible in order to accelerate the process and
to obtain a precise final garment. Through collision detec-
tion, small initial problems can be automatically solved. It
is preferable that the patterns do not interpenetrate them-
selves and the body initially.
Furthermore, Fashionizer contains a fully automatic
placement method. It works according to a placement file,
which is created from a previous, similar garment posi-
tioning. However, the automatic placement is not recom-
mended for use with haute couture garments. Many unique
pattern pieces are enclosed, and thus no predefined place-
ment file would be available.
After the placement of the patterns around the virtual
mannequin, the seaming can be executed. In Fashionizer
the seams are indicated with red lines (Fig. 12). The seams Fig. 13. 3D pattern placement
force the patterns to pull and bring the matching pattern
borders together during simulation. adjusted through the garment material panel. It features
two categories: environment (global parameters) and ob-
3.4 Specifying physical parameters and textures ject (local parameters). Among the global simulation pa-
rameters we find gravity, collision distance, and detection
The garment parameters, and particularly the physical pa- modes. The local parameters include elasticity, surface
rameters of a texture (such as cotton, linen, silk, etc.), are density, bending rigidity, friction values, and Poisson coef-
516 N. Magnenat-Thalmann, P. Volino

Fig. 14. The seaming lines perform the assembly of the 3D garment

ficient, as well as viscosity and nonlinear elasticity values, The simulation must be executed until the fabric is dy-
which are the mechanical properties of the garment. namically stabilized. The resulting position is a suitable
Fabric information for the Robert Piguet garments was starting point for the simulation of animation.
based on small fabric swatches attached to the sketches. On the screen, the virtual garment appears with an im-
With a size of around 4 × 6 cm, they were too small to pressive effect of realism thanks to the mathematical and
be measured with the Kawabata Evaluation System (KES) physical models working “behind the scenes,” visualizing
or a similar method. As replacement for the swatch, cor- the behavior of the different kinds of fabric.
relative and already tested fabrics were selected for the
extraction of the physical parameters.
Textile color and quality information was additionally 3.6 Garment animation
found in handwritten form on the original sketches. Sur-
The realistic clothing animation is simulated according to
prisingly, this information did not correspond in some
the movement of the virtual actor (Fig. 15). This is pos-
cases to the fabric attached to the drawing. The small
sible thanks to the collision detection and friction with the
fabric swatches are almost 60 years old, and the original
surface of the body. The simulation parameter adjustments
shade had discolored over time. Priority was then given to
can be different from those used during the process of
the handwritten information on the sketch, and similar fab-
seaming and assembling the garment. Mechanical simu-
rics were searched from a fabric library. The chosen fab-
lation lends realism to the animation of clothing on the
rics were photographed and prepared in a repetitive way
virtual mannequin.
for texture mapping. In addition, typical buttons and buck-
In general, for the animation of clothing on virtual
les from the 1940s were photographed to be mapped as
models, the Vicon Motion Tracking System is used to
accessories.
record realistic body postures and fashion gaits. For the
Robert Piguet exhibition, poses have been recorded ac-
3.5 Garment fitting cording to the sketches, allowing a direct association be-
tween the virtual garment and the drawings. In addition,
Once texturing and garment property setting are com-
natural ways of walking have been recorded.
pleted the “fitting” of a garment can be started by perform-
ing a mechanical simulation, which brings the surfaces
together along the seam lines (Fig. 14). 3.7 Results
The surface deforms according to the shape of the
body. The simulation engine first uses a simplified me- The system described above makes it possible to re-
chanical model, which is optimized for speed by leaving produce haute couture garments from the 1940s based
the physical parameters and environment parameters out on sketches. The system successfully simulates garments
of the calculation. After this first simulation, where the close to the body as well as loose and complex garments
garment conforms to the shape of the body, a second me- with realistic folds or localized rigidities (Fig. 16). The re-
chanical model is made for the actual simulation. In the sult obtained shows the versatility and the robustness of
second mechanical model all parameters are taken into ac- the Fashionizer software for the creation and simulation of
count. fabric in various contexts.
 From early draping to haute couture models: 20 years of research 517

Fig. 15. Computation of garment animation

Fig. 16. Creating a 3D garment based on sketch of Serge Guérin

Fig. 17. 3D garment designed by Serge Guérin

518 N. Magnenat-Thalmann, P. Volino

4 Conclusion will be ready in the near future for implementation in daily

tasks. This will have an immense effect on the different
modules in the clothing industry with its related branches.
The system described above is versatile with respect to It can be seen as a way to move a very traditional industry
creating any kind of virtual 3D clothing. It has been tested to a higher level.
for visualization of ancient historical clothing and haute For the future, one of our main goals will be to com-
couture garments of different decades. Very challenging pletely automate the above described processes for mak-
tests were performed for the purpose of prototyping en- ing virtual clothes. In the ideal application, a technical
tire garment collections, with the goal of replacing costly sketch or drawing would be given to a computer and the
physical prototypes. Accurate, fast, and flexible, the sys- computer would manufacture, assisted by robots, the real
tem gives the designer the possibility of a 3D visualiza- and virtual cloth automatically.
tion early in the design process. Furthermore, it will serve
as a new language for communicating ideas and improv-
Acknowledgement We are grateful to Christiane Luible for her
ing understanding between people in the entire production work on the Piguet haute couture garments. We also would like to
chain. thank all the fashion designers and all the people who contributed
Thanks to these latest developments, this new technol- the images presented in this paper. Our special thanks go to Nedjma
ogy, sure to cause upheavals in the fashion design world, Cadi. This work is funded by the LeapFrog CA European project.

References
1. Baraff, D., Witkin, A.: Large steps in cloth in Textiles and Apparel (IEEE Computer Graph. (Proceedings of SIGGRAPH’97),
simulation. Comput. Graph. (Proceedings Graphics and Applications), IEEE Press, 21, 205–214 (1987)
of SIGGRAPH’98), 32, 106–117 (1998) pp. 71–80, (1996) 21. Terzopoulos, D., Fleischer, K.: Modeling
2. Breen, D.E., House, D.H., Wozny, M.J.: 11. Gan, L. et al.: A study of fabric inelastic deformation: viscoelasticity,
Predicting the drape of woven cloth using deformation using non-linear finite plasticity, fracture. Comput. Graph.
interacting particles. In: Comput. Graph. elements. Textile Res. J. 65(11), 660–668 (Proceedings of SIGGRAPH’88), 22,
(Proceedings of SIGGRAPH’94), 32(4), (1995) 269–278 (1988)
365–372 (1994) 12. Hauth, M., Etzmuss, O.: A high 22. Volino, P., Courchesne, M.,
3. Carignan, M., Yang, Y., performance solver for the animation of Magnenat-Thalmann, N.: Versatile and
Magnenat-Thalmann, N., Thalmann, D.: deformable objects using advanced efficient techniques for simulating cloth and
Dressing animated synthetic actors with numerical methods. In: Proceedings of other deformable objects. Comput. Graph.
complex deformable clothes. Comput. Eurographics (2001) (Proceedings of SIGGRAPH’95), 29(4),
Graph. (Proceedings of SIGGRAPH’92), 13. Kang, Y.M., Choi, J.H., Cho, H.G., Lee, 137–144 (1995)
26(2), 99–104 (1992) D.H., Park, C.J.: Real-time animation 23. Volino, P., Magnenat-Thalmann, N.:
4. Choi, K.J., Ko, H.S.: Stable but responsive technique for flexible and thin objects. In: Developing simulation techniques for an
cloth. Comput. Graph. (Proceedings of Proceedings of WSCG, pp. 322–329 (2000) interactive clothing system. In: Proceedings
SIGGRAPH’02), 21(3), 604–611 (2002) 14. Lafleur, B., Magnenat-Thalmann, N., of Virtual Systems and Multimedia
5. Collier, J.R., Collier, B.J., O’Toole, G., Thalmann, D.: Cloth animation with (VSMM’97). IEEE Press, Geneva,
Sargand, S.M.: Drape prediction by means self-collision detection. In: Proceedings of pp. 109–118 (1997)
of finite-element analysis. J. Textile Inst. the IFIP Conference on Modeling in 24. Volino, P., Magnenat-Thalmann, N.:
82(1), 96–107 (1991) Computer Graphics, pp. 179–197 (1991) Implementing fast cloth simulation with
6. Derose, T., Kass, M., Truong, T.: 15. Metzger, J., Kimmerle, S., Etzmuss, O.: collision response. In: Proceedings of
Subdivision surfaces in character Hierarchical techniques in collision Computer Graphics International,
animation. Comput. Graph. (Proceedings of detection for cloth animation. J. WSCG pp. 257–266. IEEE Press, New York (2000)
SIGGRAPH’98), 32, 148–157 (1998) 11(2), 322–329 (2003) 25. Volino, P., Magnenat-Thalmann, N.:
7. Desbrun, M., Schröder, P., Barr, A.: 16. Oh, S., Kim, H., Magnenat-Thalmann, N., Comparing efficiency of integration
Interactive animation of structured Wohn, K.: Generating unified model for methods for cloth simulation. In:
deformable objects. In: Proceedings of dressed virtual humans. In: Pacific Proceedings of Computer Graphics
Graphics Interface, AK Peters, Warriewood, Graphics (2005) (in press) International. IEEE Press, New York (2001)
NSW, Australia (1999) 17. Press, W.H., Vetterling, W.T., Teukolsky, 26. Volino, P., Magnenat-Thalmann, N.:
8. Eberhardt, B., Weber, A., Strasser, W.: S.A., Flannery, B.P.: Numerical recipes in Accurate garment prototyping and
A fast, flexible, particle-system model for C, 2nd edn. Cambridge University Press, simulation. Comput.-Aided Des. Appl.
cloth draping. In: Computer Graphics in Cambridge, UK (1992) 2(5), 645–654 (2005)
Textiles and Apparel (IEEE Computer 18. Provot, X.: Deformation constraints in 27. Volino, P., Cordier, F., Magnenat-Thalmann,
Graphics and Applications), IEEE Press, a mass-spring model to describe rigid cloth N.: From early virtual garment simulation
pp. 52–59 (1996) behavior. In: Proceedings of Graphics to interactive fashion design.
9. Eberhardt, B., Etzmuss, O., Hauth, M.: Interface’95, pp. 147–154. AK Peters, Comput.-Aided Des. 37, 793–608 (2005)
Implicit-explicit schemes for fast animation Warriewood, NSW, Australia (1995) 28. Weil, J.: The synthesis of cloth objects.
with particles systems. In: Proceedings of 19. Sakagushi, Y., Minoh, M., Ikeda, K.: Comput. Graph. (Proceedings of
the Eurographics Workshop on Computer A dynamically deformable model of dress. SIGGRAPH’86), 20, 49–54 (1986)
Animation and Simulation. Springer, Berlin Trans. Soc. Electron. Inf. Commun. 54, 29. Yang, Y., Magnenat-Thalmann, Y.: An
Heidelberg New York, pp. 137–151 (2000) 25–32 (1991) improved algorithm for collision detection
10. Eischen, J.W., Deng, S., Clapp, T.G.: 20. Terzopoulos, D., Platt, J.C., Barr, H.: in cloth animation with human body.
Finite-element modeling and control of Elastically deformable models. Comput. Comput. Graph. Appl. (Proceedings of
flexible fabric parts. In: Computer Graphics Pacific Graphics’93), 1, 237–251 (1993)
 From early draping to haute couture models: 20 years of research 519

NADIA M AGNENAT-T HALMANN has pioneered PASCAL VOLINO is a computer scientist and
research into virtual humans over the last 25 researcher. He studied engineering sciences in
years. She obtained several Bachelor’s and Lyon, France and graduated in 1992. He then
Master’s degrees in various disciplines (psy- joined MIRALab, University of Geneva, and has
chology, biology, and chemistry) and a Ph.D. in since been working in various areas, including
quantum physics from the University of Geneva. techniques for collision detection, mechanical
From 1977 to 1989 she was a professor at the simulation, and user interaction. He obtained his
University of Montreal and led the research lab Ph.D. in 1998. His dissertation was on garment
MIRALab in Canada. She moved to the Univer- simulation. He is currently involved in several
sity of Geneva in 1989, where she founded the European projects related to cloth simulation,
Swiss MIRALab, an internationally interdisci- virtual garments, and hair animation.
plinary lab composed of about 30 researchers.
She is author or coauthor of a very large number
of research papers and books in the fields virtual
human modeling and augmented life. She has
received several scientific and artistic awards
for her work. Recently, in 1997, she was elected
to the Swiss Academy of Technical Sciences
and has been nominated as a Swiss personality
who has contributed to the advance of science in
the 150-year history of the Swiss Confederation
Parliament. She is editor-in-chief of the Visual
Computer Journal published by Springer and
coeditor-in-chief of the Computer Animation &
Virtual Worlds journal published by John Wiley
& Sons.