Wang Tiles for Image and Texture Generation

Michael F. Cohen1 Jonathan Shade2,3 Stefan Hiller4 Oliver Deussen4

1 2 3 4
Microsoft Research WildTangent University of Washington Dresden University of Technology

Abstract means to overcome this problem is to create (or capture) a small ex-
ample of complexity and then reuse this example many times. Un-
We present a simple stochastic system for non-periodically tiling the fortunately, when the same example is used many times in a periodic
plane with a small set of Wang Tiles. The tiles may be filled with fashion, the repetition is often apparent and distracting.
texture, patterns, or geometry that when assembled create a contin- We present a new stochastic algorithm to non-periodically tile the
uous representation. The primary advantage of using Wang Tiles is plane with a small set of Wang Tiles [Wang 1961; Wang 1965]. This
that once the tiles are filled, large expanses of non-periodic texture allows Wang Tiles to share the efficiency of reusing example tiles
(or patterns or geometry) can be created as needed very efficiently to create large expanses of complex texture, patterns, or pre-lighted
at runtime. geometry at runtime, while avoiding the obvious visual artifacts of
Wang Tiles are squares in which each edge is assigned a color. periodicity.
A valid tiling requires all shared edges between tiles to have Wang Tiles are a set of squares in which each edge of each tile is
matching colors. We present a new stochastic algorithm to non- colored. Matching colored edges are aligned to tile the plane. We
periodically tile the plane with a small set of Wang Tiles at runtime. demonstrate that as few as eight tiles are needed to non-periodically
Furthermore, we present new methods to fill the tiles with 2D tex- cover the plane. We present new methods to fill the tiles with 2D tex-
ture, 2D Poisson distributions, or 3D geometry to efficiently create ture, 2D Poisson distributions, or 3D geometry to efficiently create
at runtime as much non-periodic texture (or distributions, or geom- at runtime as much non-periodic texture (or distributions, or geom-
etry) as needed. We leverage previous texture synthesis work and etry) as needed. We apply these results to modeling and rendering
adapt it to fill Wang Tiles. We demonstrate how to fill individual problems.
tiles with Poisson distributions that maintain their statistical proper- Wang Tiles only define matching edge constraints. This is insuf-
ties when combined. These are used to generate a large arrangement ficient to maintain coherent features that cross more than one edge
of plants or other objects on a terrain. We show how such envi- (i.e., corners). To accommodate such features, we extend the notion
ronments can be rendered efficiently by pre-lighting the individual of Wang Tiles to include corner constraints in the spirit of [Neyret
Wang Tiles containing the geometry. and Cani 1999]. The increase in degrees of freedom also allows us
We also extend the definition of Wang Tiles to include a coding to modulate the texture by mixing two source textures.
of the tile corners to allow discrete objects to overlap more than one
edge. The larger set of tiles provides increased degrees of freedom.
1.1 Relation to Previous Work
CR Categories: I.3.3 [Computer Graphics]: Picture/Image Gen-
eration - Viewing Algorithms— [I.3.6]: Computer Graphics— There are a variety of methods where small samples are used to gen-
Methodology and Techniques erate more complex patterns. Much of the research involved in these
methods strive to avoid visual artifacts arising from repeatedly using
Keywords: non-periodic tiling, Poisson distributions, texture syn- the same sample data. Related work can be divided into two groups:
thesis, Wang Tiles 2D- and 3D-texturing methods, and geometry creation for com-
plex outdoor scenes and related level-of-detail methods. We also
touch on the relation of Wang Tiles to the broader literature on tiling.
1 Introduction
Texture Synthesis: Stam was the first to consider the use of
Modeling and rendering scenes that capture the complexity of the square Wang Tiles for texture synthesis [Stam 1997]. Based on a
real world is a difficult and time consuming task. A now well-known deterministic algorithm that uses a set of 16 tiles with colored edges
[Grünbaum and Shephard 1987] he created large non-repetitive tex-
tures, water surfaces and caustics by defining texture samples on
the tiles following the border constraints introduced by the shared
edges. In his paper, he describes the construction of a limited set of
patterns. A general algorithm for filling the tiles and quickly assem-
bling a tiling is not given. We extend his work in these directions.
Recent work in texture synthesis provides a means to avoid repe-
tition in two dimensions by using a small texture tile as an example
that is then used to create a larger amount of non-repetitive texture
that has the same visual properties.
Wei and Levoy [Wei and Levoy 2000] use a statistical model that
selects a pixel of the given texture by analyzing the local neighbor-
 (a) (b) (c) (d)
Figure 1: a) Eight Wang Tiles that can stochastically tile the plane; b) A small portion of the plane with a valid tiling. c) 12 Wang Tiles
constructed from two horizontal and three vertical colors; d) 18 Wang Tiles constructed from three horizontal and vertical colors.

hood of the pixel in the so far synthesized texture. Efros and Free- 2 Wang Tiles
man [Efros and Freeman 2001] randomly select square regions from
the sample image and fit (or quilt those parts to the already created A Wang Tile set consists of square tiles with color-coded edges. The
large texture. One of the challenges in this work is to maintain the squares cannot be rotated. A valid tiling of the infinite plane con-
visual appearance of large coherent regions. The work cited above sists of any number of copies from the set laid down such that all
achieves these goals at a computational and storage cost. We lever- contiguous edges have matching colors. This tiling is named after
age these results to fill a small set of Wang Tiles, such that the re- Hao Wang who conjectured in 1961 that any set of tiles that can
sultant tiling maintains the ability to include large coherent features. produce a valid tiling of the plane must also be able to produce a pe-
The tiling itself mitigates the computational and memory require- riodic tiling of the plane [Grünbaum and Shephard 1987]. This was
ments when creating large textures. later refuted in a series of papers that eventually led to a set of only
13 tiles discovered by Culik that could be shown to be strictly ape-
Several algorithms have been proposed to extend the methods riodic1 [Berger 1966; Culik II 1996; Kari 1996]. Theoreticians have
to arbitrary surfaces [Turk 2001; Wei and Levoy 2000; Tong et al. taken an interest in aperiodic sets, using them to simulate the proper-
2002], but these create large textures to cover the entire geometry. ties of proposed DNA computers, and have shown they can simulate
The work presented here is most closely related to the non-periodic any Turing machine. They have even been featured in a leading role
texture mapping method based on triangles presented by Neyret et in a science fiction story, “Wang’s Carpets” by Greg Egan. We will
al [Neyret and Cani 1999]. Using this kind of mapping, they are not draw on these more theoretical aspects of Wang Tiles, but rather
able to texture a variety of objects. Because the triangles are rotated on their simplicity; specifically, that they are square and that a small
to achieve a mapping, triangles are not well suited for textures set can lead to a non-periodic tiling of the plane.
with directional content such as strokes or oriented patterns. Wang
Tiles avoid these problems since they are square and cannot be
rotated, but the tilings are restricted to surfaces that map to a plane.
The triangle rotation also requires the edges to have an orientation
effectively doubling the number of colors. Beyond these differences,
we focus in this paper primarily on new methods to construct the
tile interiors themselves.

Modeling of complex ecosystems: A complex data genera-

tion problem arises when modeling outdoor scenes. Deussen et al.
(a) (b)
[Deussen et al. 1998] simulate such scenes by covering the plane
with tens of thousands of synthetic plants. Approximate instancing
helps them to reduce the geometric complexity but the storage of
the instancing information is sometimes a challenge, as stated in
[Deussen et al. 2002]. Wang Tiles can store plants or plant instances
on a small set of tiles while constructing large distributions through
tiling the plane. Lighting calculations can also be leveraged by
performing them only once per instance. For example, Layered
Depth Images [Gortler et al. 1996; Levoy and Hanrahan 1996;
Shade et al. 1998] showed how a point sampled representation
of such geometry can be pre-lighted and rendered from changing (c) (d)
viewpoints at interactive speeds. We show how the same methods
can be applied to Wang Tiles. Figure 2: Assemblies of tiles comparing stochastic tilings with
strictly aperiodic sets. Each small pseudo-colored square pixel rep-
resents one whole tile: (a) 32 × 32 tiles stochastically assembled (b)
Tiling: There is also a vast literature on tiling methods [Grünbaum 32 × 32 strictly aperiodic set [Grünbaum and Shephard 1987] (c)
and Shephard 1987; Glassner 1998], however, the square nature of 256 × 256 stochastic (d) 256 × 256 aperiodic.
Wang Tiles avoids most of the issues related to assembling a tiling,
thus this rich literature is mostly outside the scope of this work. Ex- We introduce a new stochastic process to laying down individual
panses of texture from a single square tile can be generated easily tiles. This leads to the ability to generate large tilings at runtime
by simply mirroring the tile, a process easily supported by texture that are guaranteed to be non-periodic.
mapping hardware. Wang Tiles can be thought of as a way to over-
come the repetitive nature of such a simple scheme at only a modest 1 We use the term non-periodic to indicate a tiling which is not periodic
cost in memory, while maintaining the interactive speeds of hard- (although the set itself could lead to a periodic tiling), as opposed to the term
ware supported texture mapping. aperiodic that refers to a set of tiles that can never produce a periodic tiling.
 3 Tile Design
A stochastic tiling algorithm: We work with sets that can be
trivially shown to be able to tile the plane and also trivially shown Wang Tiles provide the framework for constructing large expanses
not to be aperiodic. In fact, a single tile, such as tile b in Figure 1(a) of texture from a small amount of input. We will first demonstrate
can tile the plane on its own. When tiling the plane, each choice of and discuss an interactive tile design system. We follow this with
tile typically has two constraints. If one is placing tiles from West to an automated system that takes discrete texture primitives, such as
East and from North to South, then each tile placed must have N and images of leaves, as input. We introduce an optimization method
W edges that match the E and S edges already placed. If there are that assures that discrete primitives crossing tile boundaries are kept
K colors then there are K 2 combinations of colors for two adjacent intact while minimizing artifacts within the tile interiors.
edges of the tiles (e.g., the south (S) and east (E) edges)2 . So long as
there is at least one tile in the set with each north (N) and west (W) 3.1 Interactive Tile Design
combination, then the following simple procedure produces a valid
tiling of any portion of the plane: The simplest way to design a set of Wang Tiles is to create them
interactively. A simple tool allows us to place geometric primitives
1. select any tile for the NW corner; within the set. The primitives can be imported from any drawing
tool. If a primitive is moved such that it overlaps an edge, it will
2. tile the top row left to right by choosing tiles for which the W automatically appear in all tiles with the same and opposing edges.
edge of the new tile matches the E edge of the previous tile; Thus, primitives can reside partially in one tile and partially in tiles
with the opposite edge of the same color. A perspective projection
3. select a first tile for the next row such that the N edge matches of a portion of the tiled plane provides immediate feedback of the
the S edge from above; overall appearance and large scale patterns that might occur. Fig-
ure 3 shows two manually created sets from line patterns and the re-
4. continue this row by randomly selecting tiles for which the N sulting tilings with several hundred tiles. Although the texture was
and W edges match the S and E edges from above and the left created from only eight Wang Tiles, almost no repetition artifacts
respectively (since our set contains all NW combinations, this can be seen.
is always possible); and

5. go to step 3 for as many rows as desired.

If the set contains at least two tiles for each NW combination (i.e.,
2 × K 2 ), then there are always at least two choices at each step. If
this choice is made with uniform probability, the plane will always
be tiled non-periodically because each step reduces to an indepen-
dent random process (i.e., a coin flip). One such set of 8 tiles with
2 colors is shown in Figure 1(a). Note that there are, in practical
terms, only two colors since the red and green used for the NS edges Figure 3: Two interactively designed eight tile sets to the left of
could be replaced with the blue and yellow of the EW edges with no resulting tilings.
change in the set.
In Figure 1(a) tiles a through d contain all NW combinations as
do tiles e through h. The SE edge combinations are also each rep-
resented twice thus tilings can proceed from south to north as well. 3.2 Automatic Tile Design for Texture Synthesis
Many other sets are possible with all required properties. A valid To create a continuous texture from a sample, textures must be found
tiling of a small portion of the plane may look like Figure 1(b). for each tile that fit together across the boundaries with matching
This simple algorithm results in tilings that have a random ap- colors. In our variant of the algorithm we generate such textures.
pearance as can be seen in Figure 2(a)-(d). In Figure 2(a) a tiling A tile is created by combining diamond shaped (squares rotated by
constructed from the tile set of Figure 1(a) is shown. Each tile was 45 degrees) sample portions of the source image, one for each edge
given its own color to make them distinguishable from one another. color of horizontal and vertical edges. Therefore, for a set of eight
In Figure 2(b) a deterministic aperiodic tiling [Grünbaum and Shep- tiles we need four sample images, and for the set of 18 tiles six
hard 1987] of 16 tiles can be seen to still contain some repetitive sample images.
appearance. Figures 2(c) and (d) show larger parts of stochastic and
deterministic tiling.
Often larger sets of tiles are needed to fully avoid artifacts. Us-
ing three horizontal colors results in 12 tiles (Figure 1(c)) and using
three horizontal and three vertical colors results in 18 tiles (Figure
1(d)). The main advantage of a larger set is that it can substan-
tially reduce repetition artifacts as in Figure 6(d). Another advan-
tage is that using such a set we are able to introduce larger patterns
to the tiling of the plane. We will further address these issues in Sec-
tion 3.5 by introducing inhomogeneity into the tiling patterns. But (a) (b)
first we continue by describing our methods for creating valid tile
sets for two dimensional applications such as textured surfaces or Figure 4: a) Four subimages are combined to form each Wang Tile;
non-repetitive drawings. b) construction of an eight tile set.

2 The orientable triangles in Neyrets work [Neyret and Cani 1999] would The set of Wang Tiles can also be generated automatically. Our
require (2K)2 /3, K is effectively doubled by the orientation, but reduced method is derived from work by Efros and Freeman [Efros and Free-
by a factor of 3 since each triangle has 3 edge pair possibilities. man 2001]. In their method a large texture is created from small
 (a) (b) (c) (d)

Figure 6: a) Source image; b) texture generated using image quilting; c) 18 Wang Tiles automatically generated based on set in Figure 1(d);
d) resulting part of infinite texture (8×6 tiles) with some tile instances highlighted.

Figure 7: Three textures using 18 Wang Tiles.

monds that correspond to the edge colors of the tile. In this case
we have to find four cutting paths to combine the four samples to
form the tile (cf. Figure 4(a)). This is done using the optimiza-
tion algorithm given in [Efros and Freeman 2001]. Next, the four
combined samples that together have a diamond shape are cut along
their diagonals. This forms the final tile slightly smaller than the
sample diamonds due to the samples’ overlap. Because the same
image sample is used for all tiles that share the edge color the tiles
will always fit together. If the north edge of a tile is colored red, the
lower diagonal part of the “red” diamond is used for synthesizing
the upper part of tile, and vice versa for a red south edge. Figure 4
illustrates this process and its use on a set of eight tiles. Figure 5
shows a pseudo-colored result of a tiling with these eight tiles.
Some visible artifacts may remain (see Figure 8) for two reasons:
(1) because so few sample patches of the original texture are used,
Figure 5: A 32×32 pseudocolored tiling where each diamond shape and (2) due to artifacts introduces by the Efros’ quilting algorithm.
color represents the label of the sample chosen, or equivalently, one The first reason can be overcome, albeit at some cost, by increasing
of four edge colors. This shows the non-periodicity achieved with the number of tiles. Often one can effectively reduce the artifacts,
only four subimages. without increasing the number of tiles, by minimizing the quilting
errors. To do this, we iterate over sets of samples. For each set, its
overall quilting error is the sum of the pixel color errors that occur
along all cutting paths (e.g., 32 paths for the eight tile set). If the
square sample images obtained randomly from the source image. error is too high, four new sample images are selected, the paths are
The sample is fitted to the already created image by optimizing a calculated and the overall error is computed. This is repeated until
cutting path in an overlapping region. If the color differences of the a small overall error is obtained or a preset number of iterations is
pixels along the cutting path are below a given threshold, the sample completed.
image is added to the texture. In our trials, a set of eight tiles requires about 3 seconds of opti-
Each tile is now constructed by combining the four sample dia- mization to produce good fitting tiles. An 18 tile set takes 15 sec-
 Poisson disc distributions: To represent distributions, the tiles
each contain a set of point positions. The goal is find a set of point
distributions on each tile, that when combined will result in an over-
all distribution with desired properties. Purely random distributions
are trivial to create by simply placing random positions on the tiles.
When combined, the overall distribution remains random since there
are no dependencies between points.
However, in many cases purely random distributions are not the
best choice. In a crowd each person might position themselves to
leave a certain distance from any other person. Biological and en-
vironmental forces also do not lead to random placements of plants.
In both cases the distribution can often be described by a so called
Poisson disc distribution that avoids crowding. In such a distribu-
tion, a small disc around each object defines a region in which no
other object of the distribution is found.
Figure 8: An example of a synthesized texture where the diamond Using Wang Tiles, the challenge is to find positions on each tile
shaped artifacts are apparent. that generate a Poisson disc distribution across all possible tilings.
This presents a challenge since points placed near the border of a
tile have influence on the placement of points in any other tile with
onds on a standard PC with 2 GHz. Once the set of samples is found, a matching opposite edge.
the generation of large textures is very fast. This is a major advan- Often, dart throwing is used to generate Poisson disc distribu-
tage of tiling or pattern based methods over other texture synthesis tions: a series of random positions is created one at a time (the
techniques where the creation of texture is dependent on the overall darts). Each position is added to the already generated point set if
amount of texture required [Wei and Levoy 2000; Efros and Free- it does not overlap with any of the discs of points already inserted.
man 2001]. This works fine (given infinite time) for an infinite plane but is
Figure 6 shows some results. A small image showing some not sufficient in the case of Wang Tiles. If a dart falls near a tile
daisies in a meadow is selected (a). In (b) a texture created by the boundary, one has to check all the discs of all points that possibly
original image quilting algorithms is shown. The tiles of an eighteen lie on the opposite tile with the same edge. In essence, for a point
tile set are given in (c). (d) shows the result if 8×6 tiles are com- near a border, part of its disc is repeated in all possible neighboring
bined. Figure 7 shows some other examples of images generated by tiles. Thus, the total area in all tiles that cannot accept a new dart
eighteen Wang Tiles. due to this point is larger than one disc. This causes less points to
be inserted near the border (see Figure 9) distorting the distribution
in the overall tilings.
3.3 Automatic Tile Design for Object Distributions

Figure 9: Dart throwing problem for Poisson distributions near

edges. The example above illustrates the problem with two tiles with
blue east edges and two tiles with corresponding blue west edges.
A dart landing at a has a disc (and thus an influence) that extends (a) (b)
beyond the border of tile 1 in which it lands. This extension into
tiles 3 and 4 appears as the grayed out areas a0 . No new darts can
land in these regions. A new dart can land at b. Its disc lies fully in-
side tile 2. This new dart at b will now always have an empty space
to its right of almost 2 disc radii since the dart at a has precluded
new darts landing nearby in tiles 3 and 4. This illustrates the prob-
lem with pure dart throwing for defining a Poisson distribution with
Wang Tiles.

The creation of pseudo-random distributions is another area

where Wang Tiles can be used efficiently. For complex ecosystems
like those presented in [Deussen et al. 1998], tens of thousands of (c)
objects have to be distributed. Instancing can minimize the amount
of geometry needed, but even the storage of plant instances can Figure 10: Eight Wang Tiles with joint Poisson distributions, a por-
consume lots of memory for large distributions. Alternatively, a tion of the plane resulting from these tiles; b) Fourier transform
few objects can be positioned on a small set of Wang Tiles and showing some weak grid structure, from [Hiller et al. 2001]; c) com-
then the full population can be generated by combining the tiles parison of a distribution generated from one tile with 160 plants (up-
non-periodically. per row) and another one created from 8 tiles with 20 plants each
(lower row).
 (a) (c)

(b) (d) (e)

Figure 11: The corner problem: a) A Wang tile with all combinations of marked corners. b) Tiling from the expanded tile set that shows various
combinations of marked/unmarked tile corners. Constructing an inhomogeneous tiling: c) two source images; d) tiling that contains density
information. e) a tiled plane (12x12) constructed from a tile set with inhomogeneities.

Lloyd’s method: An alternative for creating Poisson disc distri- paragraph, the solution also offers the possibility to generate tilings
butions is to use Lloyd’s method to optimize an initial set of point that contain inhomogeneities.
positions [Deussen et al. 2000; Hausner 2001; McCool and Fiume A solution is found by coding the corners of a tile as an addi-
1992]. Lloyd’s method has the advantage over dart throwing in that tional bit of information, essentially coloring the corners as well as
points can interact with each other. the edges3 . For a single tile, 24 = 16 possibilities of corner codings
A starting Poisson distribution point set is created independently can occur (cf. Figure 11(a)). A set of tiles with two colored edges in
in each tile. Then, one tile is selected and surrounded by 8 tiles from this case is expanded to 16×8=128 tiles. The number can be reduced
the set. A Voronoi diagram is constructed across the nine tile set, to 64 tiles and still maintain the stochastic selection since there are
and the points within the selected center tile are relaxed to satisfy only three corners that must match as each tile is inserted. In the ex-
the Poisson disc constraints across the tile boundaries. The local panded set for each of these two possibilities all corner combinations
tilings and relaxation steps are repeated several times for all tiles in the tile have to be stored. In Figure 11(b) a tiling with marked and
until the distribution stabilizes. This results in consistent appearing unmarked corners in arbitrary combinations is shown.
point distributions, such as one shown in Figure 10 for a set of eight
Wang Tiles. Details and a numerical evaluation of this optimization
process can be found in Hiller et al. [Hiller et al. 2001] and as an 3.5 Introducing Inhomogeneity
application to sampling in Deussen [Deussen et al. 2000]. Figures So far the generated textures and distributions are quite homoge-
12 and 13(e) show this process applied to positioning sunflowers in neous. The selection process for textures and object positions en-
a simulated sunflower field. ables us to avoid visible patterns, but the generated textures appear
uniform. To improve the appearance it would be helpful to allow the
3.4 The Corner Problem tiling process to locally change the texture content.
The corner marked tile sets can help here. To create an inhomoge-
Edge coloring and the texture synthesis process ensures that the tiles neous tiling, the texture synthesis process takes two source images
always fit together. But a problem arises if an object or a visible instead of one. The tiles are generated by choosing one source im-
artifact of the texture is placed across a corner of a tile. In this case age for creating all those parts of a tile with marked corners and the
the vertical edge color constraint enforces all tiles with same col- other one for the rest (cf. Figures 4(b) and 11(c)). Of course, the
ored opposite vertical edges to include the remainder of this object two images must potentially fit together via the quilting algorithm to
to ensure the fitting condition. As the same object is also on the achieve natural textures.
horizontal edge, all tiles that contain the corresponding horizontal This enables us to create tilings with two different density levels
edges have to be adapted similarly. In total, in a two colored set of as shown in Figure 11(c). Here, the marked corners are shown as re-
eight tiles six tiles have to be adapted, in a three colored set (18 tiles) gions of different density. In Figure 11(e) a result can be seen using
twelve tiles. This results in visible repetitive patterns.
We are able to create a tile set that avoids such corner problems, 3 It should be noted that the triangular tiles in [Neyret and Cani 1999]

albeit at the cost of adding more tiles to the set. As shown in the next already take into account the corners as part of their coding as well.
two images of daisies, one with a high daisy density and the other
one with low density. The texture quilting optimization process in
this case lasts longer (about 1 minute on a 2 GHz PC). But, once this
is run, generating large amounts of textures at runtime is relatively
free both in texture memory and time.

4 Three Dimensional Applications

(a) (b)

(c) (d)

Figure 12: A Wang tiled field filled with sunflowers

The manual effort required to model natural environments and the

computational cost required to render such scenes can both be very
high. Systems that attempt to render realistic looking natural envi-
ronments have solved either one problem, or the other, but not both.
The systems presented in Weber and Penn [Weber and Penn 1995],
Mech and Prusinkiewicz [Měch and Prusinkiewicz 1996], Deussen
et al. [Deussen et al. 1998], and Prusinkiewicz et al. [Prusinkiewicz (e)
et al. 2001] are capable of creating realistic plants and environments.
However, the models created by these systems are of such high detail Figure 13: LOD selection. (a) A typical portion of the sunflower
that they can not be rendered in real-time. Deussen et al. [Deussen field. (b) This view straddles two of the 8 wedges that divide the
et al. 2002] achieve frame rates from 4-8 fps for complex vegeta- space of directions around a tile. (c) Levels of detail false-colored.
tion, but they need lots of memory to store their plant models. This Red (not seen below us) is highest detail, orange is lowest detail.
makes their approach not so well suited for larger areas. At the other (d) Layout of directions and resolution levels from viewpoint; e) A
end of the spectrum, geometric level-of-detail algorithms [Hoppe Frame from interactive LDI renderer with 20 flowers per tile. Sky is
1996; Perbert and Cani 2001] have increased the effective geometric environment mapped, mountains are geometry. Frames rendered at
complexity that can be rendered in real-time, but provide no direct between 2.7 and 4.0 fps on a P4 1.7 GHz PC.
support for realistic shading.
We address both problems with the use of Wang Tiles as con-
tainers for 3D geometry. In our example application of a sunflower
field, we use tiles to group a set of flowers into a single lighting 3D textures are each represented as Layered Depth Images (LDIs)
environment. We use eleven variations of a model of a sunflower that convert complex geometry into a sparse set of pre-shaded sam-
and position them stochastically in each of 8 Wang Tiles based on ples [Shade et al. 1998].
the Poisson disc distribution discussed earlier. Sunflowers that inter- Real time rendering is enhanced by extending LDIs with a novel
sect a tile boundary are repeated in all tiles with opposite matching view-dependent multi-resolution formulation of the representation.
colors. A tiling draped over a height field can be rendered with a Since tiles further from the viewpoint require fewer samples, we cre-
standard ray tracer as seen in Figure 12. ate 7 levels of detail of each tile, each with successively fewer sam-
ple points. Tiles viewed from a particular direction need include
only those points of geometry seen from that direction. To address
4.1 Layered Depth Image Tiles this we create 8 variations of each tile specialized for viewing from
A lower quality, but real-time rendering can be achieved by Wang each of 8 compass directions.
Tiles in which the geometry within each tile is pre-shaded. This is The complete system consists of an offline and an online compo-
possible since Wang Tiles are not allowed to be reoriented when nent. Offline, we create and pre-shade the hierarchical set of tiles
constructing the tilings. We discuss one possible represention in composed of instances of the eleven sunflower models. At runtime,
which the geometry is sampled into a hierarchy of 3D textures. The we compute a tiling of the plane that is draped over a terrain and
render the visible tiles in back to front order. Each tile is selected D EUSSEN , O., C OLDITZ , C., S TAMMINGER , M., AND D RETTAKIS , G.
from the appropriate hierarchy level and from the appropriate view- 2002. Efficient rendering of complex ecosystems using points and lines.
ing direction based on its distance and orientation to the camera. In IEEE Visualization 2002, IEEE, 219–226.
Each sample point within each tile is splatted to the screen in back E FROS , A., AND F REEMAN , W. 2001. Image quilting for texture synthesis.
to front order as described in the original LDI paper [Shade et al. In Proceedings of SIGGRAPH 2001, 341–346.
1998] (see Figure 13).
G LASSNER , A. 1998. Aperiodic tiling, part 1. IEEE Computer Graphics
Because we focus here on the application and construction of the and Applications 18, 4 (May).
Wang Tiles, we leave the details of how the original geometry is
sampled to create the hierarchical LDIs to a technical report [Shade G ORTLER , S. J., G RZESZCZUK , R., S ZELISKI , R., AND C OHEN , M. F.
et al. 2002]. 1996. The lumigraph. In Proceedings of SIGGRAPH 1996, 43–54.
We are able to render the sunflower field in real-time on a 2GHz G R ÜNBAUM , B., AND S HEPHARD , G. C. 1987. Tilings and Patterns. W.
PC with an Nvidia GeForce4 graphics card. Across the data sets, our H. Freeman and Company. ISSN 0716711931.
software renderer was able to maintain a rendering rate of between H AUSNER , A. 2001. Simulating decorative mosaics. In SIGGRAPH 2001
4.5 and 5.7 million depth pixels per second. This corresponds to Conference Proceedings, 573–578.
about 4 fps for the sunflower field. Given the point-rendering capa-
H ILLER , S., D EUSSEN , O., AND K ELLER , A. 2001. Tiled blue noise
bilities of modern graphics hardware, namely point-sprites in mod-
samples. In Proceedings of Vision Modelling Visualization 2001, Infix-
ern graphics APIs, we believe point-sampled tiles can reach 30 fps.
Verlag, 265–272.
H OPPE , H. 1996. Progressive meshes. In Proceedings of SIGGRAPH 1996,
5 Discussion and Future Work 99–108.
K ARI , J. 1996. An small aperiodic set of wang tiles. Discrete Mathematics
We have shown the versatility and usefulness of Wang Tiles for im- 160, 259–264.
age and texture synthesis, and for generating large point distribu- L EVOY, M., AND H ANRAHAN , P. 1996. Light field rendering. In Proceed-
tions. Wang Tiles were interesting in their own right as a theoretical ings of SIGGRAPH 1996, 31–42.
entity. They should also now prove their usefulness in a variety of
M C C OOL , M., AND F IUME , E. 1992. Hierarchical poisson disk sampling
practical graphics applications.
distributions. In Graphics Interface ’92, 94–105.
Computer games can benefit from the extremely compact repre-
sentation for texture using tiles. We also look forward to seeing how M ĚCH , R., AND P RUSINKIEWICZ , P. 1996. Visual models of plants in-
well a hardware point based renderer works with the LDI tiles. teracting with their environment. In Proceedings of SIGGRAPH 1996,
397–410.
Based on a reviewer comment, we are also investigating the possi-
bility of generating the Poisson distributions with an algorithm sim- N EYRET, F., AND C ANI , M.-P. 1999. Pattern-based texturing revisited. In
ilar to the way we texture the tiles in section 3.2. Proceedings of SIGGRAPH 1999, 235–242.
In addition, techniques that use volumetric textures [Neyret N EYRET, F. January-March 1998. Modeling, animating, and rendering com-
January-March 1998] or slices [Perbert and Cani 2001] can poten- plex scenes using volumetric textures. IEEE Transactions on Visualiza-
tially save a lot of memory by tiling non-periodically with only small tion and Computer Graphics 4, 1, 55–70.
samples of their data represented in the tiles. P ERBERT, F., AND C ANI , M. 2001. Animating prairies in real-time. In
Another three-dimensional application is to create tilings of vol- 2001 ACM Symposion on Interacive 3D Graphics, 103–110.
umes with Wang-like cubes. In this case, each tile is a small cube
with sides colored. Doing so, we would need a set of 32 tiles to P RUSINKIEWICZ , P., M ÜNDERMANN , L., K ARWOWSKI , R., AND L ANE ,
B. 2001. The use of positional information in the modelling of plants. In
tile the space non-periodically. Applications for such tilings can be
Proceedings of SIGGRAPH 2001, 289–300.
particle systems that represent clouds or stars in galaxies. We are
certain there are many other applications of these elegant entities we S HADE , J., G ORTLER , S., H E , L., AND S ZELISKI , R. 1998. Layered depth
have not thought of. images. In Proceedings of SIGGRAPH 1998, 231–242.
S HADE , J., C OHEN , M. F., AND M ITCHELL , D. 2002. Tiling layered depth
images. ftp://ftp.cs.washington.edu/tr/2002/12/UW-CSE-02-12-07.pdf.
Acknowledgements
S TAM , J. 1997. Aperiodic texture mapping. Tech. rep., R046. Euro-
The authors would like to acknowledge the many useful discussions pean Research Consortium for Informatics and Mathematics (ERCIM).
with Don Mitchell during preparation of an early version of this http://www.ercim.org/publication/technical reports/046-abstract.html.
work. We also wish to thank the reviewers (both in this and past T ONG , X., Z HANG , J., L IU , L., WANG , X., G UO , B., AND S HUM , H.-Y.
years) for many helpful comments. 2002. Synthesis of bidirectional texture functions on arbitrary surfaces.
ACM Transactions on Graphics 21, 3 (July), 665–672. ISSN 0730-0301
(Proceedings of ACM SIGGRAPH 2002).
References T URK , G. 2001. Texture synthesis on surfaces. In Proceedings of ACM
SIGGRAPH 2001, ACM Press / ACM SIGGRAPH, Computer Graphics
B ERGER , R. 1966. The undecidability of the domino problem. Memoirs
Proceedings, Annual Conference Series, 347–354. ISBN 1-58113-292-1.
American Mathematical Society, 66, 1–72.
WANG , H. 1961. Proving theorems by pattern recognition II. Bell Systems
C ULIK II, K. 1996. An aperiodic set of 13 wang tiles. Discrete Mathematics Technical Journal 40, 1–42.
160, 245–251.
WANG , H. 1965. Games, logic, and computers. Scientific American
D EUSSEN , O., H ANRAHAN , P., P HARR , M., L INTERMANN , B., M ĚCH , (November), 98–106.
R., AND P RUSINKIEWICZ , P. 1998. Realistic modeling and rendering of
plant ecosystems. In SIGGRAPH 98 Conference Proceedings, 275–286. W EBER , J., AND P ENN , J. 1995. Creation and rendering of realistic trees.
In Proceedings of SIGGRAPH 1995, 119–128.
D EUSSEN , O., H ILLER , S., VAN OVERVELD , K., AND S TROTHOTTE , T.
W EI , L.-Y., AND L EVOY, M. 2000. Fast texture synthesis using tree-
2000. Floating points: A method for computing stipple drawings. Com-
structured vector quantization. In Proceedings of SIGGRAPH 2000, 479–
puter Graphics Forum, Eurographics 2000 Conference Proceedings 19,
488.
4, 40–51.