14678659, 2003, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/1467-8659.t01-2-00647 by ${individualUser.givenNames} ${individualUser.surname} - <Shibboleth>[email protected] , Wiley Online Library on [09/01/2023].

See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Volume 22 (2003), number 1 pp. 79–85 COMPUTER GRAPHICS for um

Quad/Triangle Subdivision

Jos Stam1 and Charles Loop2

1 Alias wavefront, Seattle, WA, USA

2 Microsoft Research, Redmond, WA, USA

Abstract
In this paper we introduce a new subdivision operator that unifies triangular and quadrilateral subdivision
schemes. Designers often want the added flexibility of having both quads and triangles in their models. It is also
well known that triangle meshes generate poor limit surfaces when using a quad scheme, while quad-only meshes
behave poorly with triangular schemes. Our new scheme is a generalization of the well known Catmull–Clark
and Loop subdivision algorithms. We show that our surfaces are C 1 everywhere and provide a proof that it is
impossible to construct such a C 2 scheme at the quad/triangle boundary. However, we provide rules that produce
surfaces with bounded curvature at the regular quad/triangle boundary and provide optimal masks that minimize
the curvature divergence elsewhere. We demonstrate the visual quality of our surfaces with several examples.
ACM CSS: I.3.5 Computer Graphics—Curve, surface, solid, and object representations

1. Introduction Unfortunately, both Catmull–Clark and Loop surfaces re-

quire that all patches be quadrilateral or triangular, respec-
Subdivision surfaces are currently one of the most powerful tively. In theory this is not a problem, since any triangle
surface representations used to model smooth shapes. Un- can be converted into three quads and any quad can be tes-
like regular surface splines, such as NURBS, subdivision selated. However, as illustrated in Figure 1, Catmull–Clark
surfaces can handle shapes of arbitrary topology in a uni- surfaces behave very poorly on triangle-only base meshes:
fied framework. Also, unlike polygonal meshes, subdivision the resulting surface exhibits annoying undulating artifacts.
surfaces generate smooth surfaces, which is important in de- Similarly, Loop schemes do not perform well on quad-only
signing aesthetically pleasing shapes. Subdivision surfaces meshes. More importantly, designers often want to preserve
were introduced in 1978 by both Catmull and Clark [1] and quad patches on regular areas of the surface where there are
Doo and Sabin [2]. They both generalized tensor product two “natural” directions. Consequently it is often desirable
B-splines of bi-degree three and two, respectively, to arbi- to have surfaces that have a hybrid quad/triangle patch struc-
trary topologies by extending the refinement rules to irregu- ture.
lar parts of the control mesh. Later, in 1987 Loop generalized
triangular Box splines of total degree four to arbitrary trian- To the best of our knowledge only one previous hybrid
gular meshes [3]. These subdivision schemes have the de- quad/triangle subdivision scheme has been proposed in the
sirable property that they admit a polynomial representation literature [7]. This scheme reproduces Catmull–Clark on the
on the regular part of the mesh. Consequently, these surfaces quad region but does not reproduce Loop on the triangular
are curvature continuous almost everywhere [4,5] and can be part. The triangular scheme on the regular part of the
evaluated explicitly anywhere [6]. mesh is actually non-polynomial and is not even curvature
continuous. Our scheme on the other hand reproduces
The visual quality of a subdivision surface depends in a Loop surfaces on the triangular part of the mesh. Also, no
crucial way on the initial, or base, mesh of control vertices. analysis of the limit surface was provided in [7]. Our hybrid
For general shapes designers often want to model certain subdivision algorithm is surprisingly simple and is based on
regions with triangle patches and others with quad patches. the fact that both Loop and Catmull–Clark surfaces can be

c The Eurographics Association and Blackwell Publishing Ltd

2003. Published by Blackwell Publishing, 9600 Garsington Road,
Oxford OX4 2DQ, UK and 350 Main Street, Malden, MA 02148,
USA. 79
 14678659, 2003, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/1467-8659.t01-2-00647 by ${individualUser.givenNames} ${individualUser.surname} - <Shibboleth>[email protected] , Wiley Online Library on [09/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
80 J. Stam and C. Loop / Quad/Triangle Subdivision

e
e d β
b b d γ β
e
b d β
b d β
e e
b a d c β α
b d β

e e γ
b b d d β β
e γ

Figure 1: A regular triangular mesh (left) behaves poorly Figure 4: Smoothing masks for Loop, Catmull–Clark
with Catmull–Clark (middle) and behaves nicely with Loop. and our new scheme. The weights are a = 2(3/8 +
1/4 cos(2π/n e ))2 − 1/4, b = (1 − a)/n e , c = (n e − 3)/n e ,
d = 2/n 2e and e = 1/n 2e . In our example n e = 8 and
n q = 3.

with a linear combination of itself and its direct neighbors

resulting in the mesh shown in Figure 2 (right). When a
vertex is entirely surrounded by triangles or quads we use
the first two masks shown in Figure 4, otherwise when
Figure 2: Our subdivision scheme comprises two steps. A the mask is surrounded by both triangles and quads we
base mesh (left) is first linearly subdivided (middle) and then introduce the new third mask. We first consider the regular
followed by an averaging of the vertices (right). case when a vertex is surrounded by two adjacent quads
and three adjacent triangles. In this case the obvious choice
1 1 1 1 1 1 is the mask shown in Figure 3 (right), which is a simple
16 8 16 8 16 8
average of the regular quad and triangle masks also shown
1
1
1
1
8 1 8 in Figure 3. To derive a general mask at an irregular vertex
8
1
1 4 1
4
1
4 we introduce some notations: let n e be the number of edges
8 8 8
emanating from the vertex and let n q be the number of quads
1 1 1
8 8 8 surrounding the vertex. As shown in Figure 4, we denote by
1 1 1 1 1 1
α the weights associated with the irregular vertex and let β
16 8 16 8 16 8
and γ be the weights associated with the neighboring edge
and face vertices, respectively. A natural generalization of
Figure 3: Averaging masks for regular quads (left), regular the regular case is to let the weights be equal to
triangles (middle) and regular quad-triangles (right).
α α
β= and γ = .
2 4
generated by a linear subdivision step followed by vertex-
smoothing [8]. This algorithm is related to the “repeated The value of α is then determined from the condition that the
averaging” schemes of Zorin and Schröder [9] and Warren rules should be affine invariant:
and Weimer [10] in that two averaging steps are taken at α α
once. α + ne + nq = 1.
2 4

This equation is readily solved for α:

2. Quad/Triangle Subdivision
1
2.1. The Generalization α= n . (1)
1 + n2e + 4q
Our new surfaces are modeled by a base mesh that is formed
of quads and triangles only. We assume that polygons of the
base mesh other than quads and triangles are quadrangulated The choice for our mask is quite arbitrary as there
by placing a vertex at the centroid of the polygon; as in are potentially many other affine invariant generalizations.
step one of the Catmull–Clark algorithm. Optionally, these Ideally we want a mask which gives the most aesthetically
faces could be triangulated. Figure 2 (left) shows a typical pleasing surfaces. One way to formalize this requirement is
base mesh. Our subdivision rules comprise two steps. In to make the curvature well behaved at each vertex. In fact,
the first step we evenly split each edge into two, each quad for n e < 5 our masks are far from optimal. Before we
into four and each triangle into four, as shown in Figure discuss the modification to our masks we have to recall some
2 (middle). Then in a second step we replace each vertex well known results from the theory of subdivision surfaces.

c The Eurographics Association and Blackwell Publishing Ltd 2003
 14678659, 2003, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/1467-8659.t01-2-00647 by ${individualUser.givenNames} ${individualUser.surname} - <Shibboleth>[email protected] , Wiley Online Library on [09/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
J. Stam and C. Loop / Quad/Triangle Subdivision 81
Table 1: Values of the correction factor η for different configura-
tions. Also given is the corresponding value of the ratio ρ.

ne nq η ρ

2 1 −0.20505 1.000
3 1 0.80597 1.227
3 2 0.61539 1.242
4 1 0.34792 1.000
4 2 0.21380 1.000
4 3 0.10550 1.000

technique to achieve this is to follow the averaging step with

a correction step. This idea was first introduced (implicitly)
Figure 5: Surface generated using our new subdivision by Catmull and Clark to improve the behavior of the surface
scheme: without vertex correction (top) after vertex correc- at a vertex where three quads meet: a corner. This fact was
tion (bottom). first pointed out in [7]. The idea behind our correction step is
to translate the smoothed vertex along the direction defined
by the difference between its current position v1 and its
previous position v0 by an amount η:
2.2. Eigen-Analysis of Subdivision
v2 = v1 + η(v1 − v0 ).
Of crucial importance in the theory of subdivision surfaces is
the subdivision matrix S. To every vertex of the control mesh To make the corners more rounded we let η > 0, so that
we associate a one-ring of neighboring vertices as depicted the vertex is drawn more inward. We numerically determine
in Figure 4. The subdivision matrix specifies how this set the optimal value of η which makes the ratio ρ = 1. These
of vertices is transformed into a similar set of vertices after values are reported in Table 1. When n e = 3 the optimal
one iteration of the subdivision rules. This matrix is of size value ρ = 1 cannot be achieved because larger values of η
K × K , where K = 1+n e +n q . As first pointed out in [2] the result in a subdivision matrix having a negative eigenvalue,
eigenstructure of this matrix is important in the analysis of which results in undesirable oscillations. When this happens
the limit behavior of the surface at the central vertex. Due to we compute the value of η which sets the smallest eigenvalue
the property of affine invariance, the matrix S always has a exactly to zero. Figure 5 (bottom) shows the effect of the
maximum eigenvalue equal to one. The next five eigenvalues correction step. Notice how the pinched behavior has almost
in order of magnitude completely disappeared and the surfaces have a rounder
appearance.
1 > λ
λ2 > µ
µ2
µ3 > · · ·
When n e
5 we did not notice any improvement in
are important in characterizing the behavior of the tangent the ratio ρ when varying the correction factor η. Nor did
plane and the behavior of the curvature at the central vertex. we notice any differences in the appearance of the resulting
In particular, the two left eigenvectors corresponding to λ surfaces. Therefore we simply ignore the vertex correction
and λ2 can be used to compute the normal (if it exists) step when n e
5. It is possible to improve on the behavior
of the limit surface at the central vertex. When the surface of the curvature in these regions by modifying the original
curvature is continuous, µ = λ2 . Consequently, to measure mask given by Equation 1. We did not explore this line of
the quality of the curvature we propose the following ratio enquiry further as the surfaces we experimented with had no
ρ = µ/λ2 . Ideally, this ratio should be equal to 1. When apparently bad artifacts in those regions.
ρ < 1 the curvature is zero and the surface has a flat spot,
which is undesirable. On the other hand, when ρ > 1 the
curvature diverges which results in surfaces which appear to 2.4. Creases, Boundaries and Hierarchy
be “pinched” at the central vertex. Creases and boundaries are easily incorporated in our
scheme following [8]. Smoothing along boundaries and
2.3. Vertex Correction creases is achieved using the [ 14 21 14 ] cubic B-spline
averaging mask along the boundary/crease. Alternatively,
Refer to Figure 5 (top) depicting three surfaces created with more clever rules could be devised to handle pathological
our generalized masks. Clearly the corners of these surfaces boundary cases as shown in [11]. On the quad and triangle
appear to be overly pinched. Somehow we want the limit regions of the mesh we can use any of the well known
surface at the corners to be drawn more inward. One simple hierarchical schemes.

c The Eurographics Association and Blackwell Publishing Ltd 2003
 14678659, 2003, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/1467-8659.t01-2-00647 by ${individualUser.givenNames} ${individualUser.surname} - <Shibboleth>[email protected] , Wiley Online Library on [09/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
82 J. Stam and C. Loop / Quad/Triangle Subdivision

2 5 24
26 11 2 -1 -1 11 -6 -4 -2 0 0 8 11 11 11 11 15 15
2 26 3 9 8 8
26 11 2 -1 -1 11 -3 -2 -1 0 0 4 2 2 2 2 3 3
2 26 1 3 0 0
26 11 2 -1 -1 11 0 0 0 0 0 0 -1 -1 -1 -1 -1 -1
2 26 -1 -3 0 0
26 11 2 -1 -1 11 3 2 1 0 0 -4 2 2 2 2 3 3
2 26 -3 -9 8 8
26 11 2 -1 -1 11 6 4 2 0 0 -8 11 11 11 11 15 15
2 -5 24
x 1/3 x 1/3 x 1/2 x 1/3 x 1/4
Figure 6: Control nets defining the characteristic map for
different values of n e and n q . Figure 7: Coefficients of the Box splines corresponding to
the shaded area (top) that reproduce the three quadratics:
x 2 , x y and y 2 (bottom).
3. Smoothness
3.1. C 1 -Continuity 19 17 -9 -1 1 1
696 174 104 26 12 3

Our surfaces are naturally curvature continuous on both the 1 1 1

12 4 6
regular quad and the triangle regions of the mesh since 41
87
1
0
13 -17 13 0
each agree exactly with regular bi-cubic B-splines and 116 52
1 1 -1
triangular Box splines of total degree four, respectively. At 12 4 6

the irregular quad and triangle regions the surface is C 1 but

19 17 -9 -1 -1 -1
not curvature continuous. We refer the reader to the literature 696 174 104 26 12 3

for a proof of this fact [4,5]. The crucial tool used in these
proofs is the characteristic map first introduced by Reif Figure 8: Limit masks for position (left) and normal
[4]. This map is the surface defined by the two-dimensional (middle/right) for a regular quad/triangle vertex.
mesh formed by the two eigenvectors corresponding to λ
and λ2 (see Section 2.2) of the larger subdivision matrix
S , which includes two rings of neighboring vertices.
Figure 6 shows the aforementioned meshes for various
configurations of quads and triangles around a central
vertex. A fundamental theorem of subdivision surfaces
states that when the characteristic map is both regular
and injective, the surface is C 1 at the central vertex. In
Figure 6 we depict several control meshes for characteristic
maps corresponding to different values of n e and n q . It
seems reasonable to infer that the corresponding surfaces are
injective. In fact we can verify injectivity for these meshes
up to machine precision. This argument does not constitute
a formal proof of C 1 for all possible cases. However in
practice what matters most is that we can compute a limit Figure 9: Comparison of Loop (left), Catmull–Clark (mid-
normal from the left two eigenvectors corresponding to λ dle) and our new scheme (right).
and λ2 [12].

half of the patches at the boundary are exact polynomials

3.2. Non-C 2 -Continuity
(shown in grey at the top of Figure 7). On these patches
We now show that our surfaces cannot be C 2 along the we can exactly represent any of the quadratics in terms of
regular quad/triangle boundary. In order for the surface to the Box spline basis functions. This completely constrains
be C 2 it must be able to reproduce the three quadratics x 2 , the values of the control vertices on both sides of the
x y and y 2 at any patch along the boundary (see Theorem quad/triangle boundary. The values of these vertices are
3.1 in [13] for example). After one subdivision step, one readily computed and are given on the bottom of Figure 7.

c The Eurographics Association and Blackwell Publishing Ltd 2003
14678659, 2003, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/1467-8659.t01-2-00647 by ${individualUser.givenNames} ${individualUser.surname} - <Shibboleth>[email protected] , Wiley Online Library on [09/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
83

c The Eurographics Association and Blackwell Publishing Ltd 2003

Figure 10: Different surfaces generated using our new scheme.
J. Stam and C. Loop / Quad/Triangle Subdivision

 14678659, 2003, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/1467-8659.t01-2-00647 by ${individualUser.givenNames} ${individualUser.surname} - <Shibboleth>[email protected] , Wiley Online Library on [09/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
84 J. Stam and C. Loop / Quad/Triangle Subdivision

For the surface to be C 2 these coefficients have to agree curvature. In general, for hybrid quad/triangle meshes our
at the quad/triangle boundary. Unfortunately this is not the new scheme produces nicer surfaces than previous schemes.
case for the coefficients of y 2 and consequently the surface
In the future we intend to find an exact evaluation
cannot possibly be C 2 . Note that our proof carries through
procedure for our surfaces similar to [6]. The evaluation
even if we use a different configuration than the one shown
depends on the ability to evaluate exactly along the regular
in Figure 7. Other configurations change the coefficients we
quad/triangle boundary. Once this is established evaluation
present. These may change the quadratics spanned, but no
everywhere follows directly. We are also searching for a
choice captures all three.
formal proof of the fact that our surfaces are C 1 , perhaps
However, for the rules we have chosen, the surface at the by recasting the problem into the known quad or triangle
quad/triangle boundary has bounded curvature since it has framework or by inventing a new method of proof. A recent
the following regular set of eigenvalues: paper by Levin and Levin proposes such a new framework
which they apply to the regular quad/triangle boundary [14].
1 1 1 1 1 1 3
1, , , , , , , , Finally, even though our schemes work well on hybrid
2 2 4 4 4 8 32
quad/triangle meshes, they still perform poorly when the
such that µ = λ2 . base mesh contains faces with a very bad aspect ratio. This
is a more general problem that plagues uniform stationary
We mention that by increasing the size of the masks and subdivision. A topic for future research could be to come up
allowing negative weights it is possible to construct a C 2 with new rules which can handle any type of base mesh and
scheme [14]. still produce aesthetically pleasing looking shapes.

4. Implementation References

We have implemented our scheme as a MAYA shape plugin 1. E. Catmull and J. Clark. Recursively generated B-spline
node. Our plugin takes as an input any MAYA-modeled surfaces on arbitrary topological meshes. Computer
polygonal mesh comprised of triangles and quads and Aided Design, 10(6):350–355, 1978.
displays a shaded polygonal mesh using the limit positions
and normals of the vertices at a given level of subdivision. 2. D. Doo and M. A. Sabin. Behaviour of recursive sub-
To compute limit positions and limit normals we use the division surfaces near extraordinary points. Computer
left eigenvectors of the corresponding subdivision matrix Aided Design, 10(6):356–360, 1978.
(see Section 2.2). In the interior of the faces these limit
3. C. T. Loop. Smooth subdivision surfaces based on
masks are regular and well known. At irregular quad and
triangles, M.S. Thesis, Department of Mathematics,
triangle vertices we use the masks given in [12] and [3].
University of Utah, August 1987.
At the vertices that share triangles and quads we compute
these eigenvectors numerically only once when reading in 4. U. Reif. A unified approach to subdivision algorithms
the base mesh. Figure 8 provides the limit masks for the near extraordinary vertices. Computer Aided Geometric
regular quad/triangle boundary vertices. Design, 12:153–174, 1995.
Figure 9 demonstrates that our new scheme performs bet- 5. D. N. Zorin. Subdivision and multiresolution surface
ter than Loop and Catmull–Clark on a cylinder-like polyg- representations, PhD thesis, Caltech, Pasadena, Califor-
onal model. Figure 10 depicts different surfaces created us- nia, 1997.
ing our new subdivision scheme. Note that the right-most
pictures show a reflection-line plot of the surface, which 6. J. Stam. Exact evaluation of Catmull–Clark subdivi-
provides an excellent visualization of the curvature. sion surfaces at arbitrary parameter values. In Com-
puter Graphics Proceedings, Annual Conference Series,
pp. 395–404. July 1998.
5. Conclusions and Future Work
7. J. Maillot and J. Stam. A unified subdivision scheme
In this paper we have proposed a novel hybrid quad/triangle for polygonal modeling. Computer Graphics Forum
scheme which unifies Catmull–Clark and Loop surfaces in a (Proceedings of Eurographics 2001), 20(3):471–479,
single framework. We obtained our subdivision rules by de- 2001.
composing the subdivision process into separate steps: linear
subdivision followed by vertex-averaging followed by a ver- 8. J. Stam. On subdivision schemes generalizing uniform
tex correction. We have shown that our surfaces are tangent b-spline surfaces of arbitrary degree. Computer Aided
plane continuous but not curvature continuous. However, we Geometric Design. Special Edition on Subdivision Sur-
have introduced masks which optimize the behavior of the faces, 18:383–396, 2001.

c The Eurographics Association and Blackwell Publishing Ltd 2003
 14678659, 2003, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/1467-8659.t01-2-00647 by ${individualUser.givenNames} ${individualUser.surname} - <Shibboleth>[email protected] , Wiley Online Library on [09/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
J. Stam and C. Loop / Quad/Triangle Subdivision 85

9. D. Zorin and P. Schröder. A unified framework for 12. M. Halstead, M. Kass and T. DeRose. Efficient, fair
primal/dual quadrilateral subdivision schememes. Com- interpolation using Catmull–Clark surfaces. In Proceed-
puter Aided Geometric Design. Special issue on Subdi- ings of SIGGRAPH ’93, Addison-Wesley Publishing
vision Surfaces, 18:429–454, 2001. Company, pp. 35–44. August 1993.
10. J. Warren and H. Weimer. Subdivision Methods For
13. H. Prautzsch and U. Reif. Necessary conditions for sub-
Geometric Design: A Constructive Approach. Morgan
division surfaces. In Tech. report, Sonderforschungs-
Kaufmann, 2001.
bereich 404. Universitĺat Stuttgart, Report 97/04, 1997.
11. H. Biermann, A. Levin and D. Zorin. piecewise smooth
subdivision with normal control. In SIGGRAPH 2000 14. A. Levin and D. Levin. Smoothness analysis of bivari-
Conference Proceedings, Annual Conference Series, ate quasi-uniform subdivision schemes. Unpublished
pp. 113–120. July 2000. manuscript, 2001.

c The Eurographics Association and Blackwell Publishing Ltd 2003