NON-PHOTOREALISTIC RENDERING OF ALGORITHMICALLY

GENERATED TREES

Nathan Dudley and Cindy Grimm

Boeing Co, Washington University in St. Louis, USA

Keywords:

Artistic rendering.

Abstract:

This paper presents a novel rendering technique inspired by artistic approaches. Instead of trying to recreate

a trad itional medium, such as charcoal or watercolor, this approach is a mixture of both photo-realism and

abstraction. Artists use a process of abstraction to provide structural information about subjects that do not

have clearly deﬁned shapes, such as groups of leaves in a tree. For example, an artist will use a color wash to

ﬁrst approximate a group of leaves. They then add detail on top of parts of this wash to indicate the presence

of individual leaves. Similarly, we use an abstract shape that approximates the image of leaves clustered at the

end of a branch. To prevent oversimpliﬁcation, we add photo-realistic detail using a blending process. Inter-

frame coherence is achieved both by smoothly interpolating the abstract shapes, and the continuity inherent in

the photo-realistically rendered detail.

1 INTRODUCTION

This paper proposes a novel non-photorealistic

method to render algorithmically generated trees. In-

stead of trying to recreate the traditional medium that

an artist might use, such as watercolor or charcoal,

our approach is inspired by how artists use a combi-

nation of abstraction and detail to convey both shape

and texture in the scene. In particular, we look at how

artists render trees.

Figure 1: Goal of project. Left: a traditional photorealis-

tic rendering. Right: an alternative rendering that blends

abstraction with photorealism.

Long before computers created their ﬁrst com-

puter generated images, humans developed many

methods, some realistic, some not, for depicting com-

plicated shapes such as trees. The ﬁrst step in paint-

ing a tree with real paint is to consider the outline

shape of the tree (Lovett, 2004). Painting every leaf

would result in an overwhelming amount of detail. In-

stead, leaves should be viewed as a mass and painted

in clumps (Boddy-Evans, 2005). To give a sense of

texture, a couple of the leaves should be drawn accu-

rately (Bourdet, 2001). The goal of this paper is to

create images inspired by this approach.

Artists use abstraction to provide structural in-

formation about subjects that do not have obvious

shapes. For example, a wash of color represents a

group of leaves. To convey detail, artists then paint in

representative leaves, but only in places. Similarly,

large branches are depicted in detail while smaller

branches might be left out altogether.

We mimic this approach to render algorithmically

generated trees. Leaves at the end of branches are

clustered into one coherent unit which are then ren-

dered with substitute geometry that suggests the shape

of the original leaves. The resulting image has a

cartoon-like appearance which is then offset by intro-

ducing photo-realistic detail using a blending process.

We render only the larger branches, omitting smaller

ones in the leaf groups.

We provide several artistic options to alter the ap-

pearance of the ﬁnal rendering. Outlining the blobs

provides visual structure to the image and enhances

perspective. The location, and strength, of detail is

important when painting. The rendering application

allows the user to control the amount of detail to use,

197

Dudley N. and Grimm C. (2007).

NON-PHOTOREALISTIC RENDERING OF ALGORITHMICALLY GENERATED TREES.

In Proceedings of the Second International Conference on Computer Graphics Theory and Applications - GM/R, pages 197-203

DOI: 10.5220/0002072201970203

 SciTePress

and where to apply it.

Ensuring inter-frame coherence is a challenge in

non-photorealistic approaches. Because we are us-

ing detail drawn from the photo-realistic rendering

process, this detail is visually stable from frame to

frame. To ensure that the abstract information is sta-

ble, we apply the full rending process every n frames,

then interpolate the abstract data for the intermediate

frames.

Contributions: We present a new, art-based ren-

dering technique that combines abstraction with pho-

torealism. This approach mirrors the artistic process

rather than the re-creating a traditional media. Al-

though in this paper we focus on algorithmically gen-

erated trees, the method could easily be extended to

arbitrary scenes.

Previous work is provided in section 2. A descrip-

tion of L-systems is provided in section 3. Section

4 describes how the groups of leaves are rendered.

Chapter 5 discusses maintaining inter-frame coher-

ence. Results are contained in section 6. Section 7

provides a conclusion and describes future work.

2 PREVIOUS WORK

There is a large (and growing) body of work in non-

photorealistic rendering. We focus on methods that

have been adapted speciﬁcally for rendering trees. In

the following section we discuss the plant generation

software (L-Systems) used to generate the models for

this paper.

In the past, computer graphics rendering fo-

cused primarily on producing photographic im-

ages (Markosian et al., 1997). Non-photorealistic

rendering, on the other hand, may focus on abstrac-

tion and simpliﬁcation, such as the work by Kowal-

ski et al (Kowalski et al., 1999). In this ap-

proach the authors use strokes to render 3D com-

puter graphics scenes in a stylized manner to sug-

gest the complexity of the scene (by drawing a few

leaves at the silhouettes) without representing it ex-

plicitly. Other approaches mimic particular media,

such as stippling (Deussen et al., 2000), charcoal

rendering (Meier, 1996), engraving (Ostromoukhov,

1999), and half toning (Freudenberg et al., 2002).

Deussen (Deussen and Strothotte, 2000), Luft (Luft

and Deussen, 2006) and Fiore (Di Fiore et al., 2003)

focus on using stippling, water colors, and cartoon

rendering, respectively, for trees.

Alvy Smith was the ﬁrst to use Lindenmayer

systems (L-systems) to render representations of

trees (Smith, 1984). The method in this paper also

uses L-systems to create the base 3D models from

which to draw. In particular, we use Prusinkiewicz’s

approach to modeling and visualization of plants us-

ing L-systems (Prusinkiewicz, 2004).

3 USING L-SYSTEMS TO

CREATE TREES

We use L-systems to create the 3D tree mesh data. We

post-process the mesh data to reduce the mesh com-

plexity, ensure that the meshes are closed, and to ex-

tract the branches and leaf groups.

Figure 2: Creating a leaf mesh using L-studio.

3.1 Creating the Mesh

The advantage of using 3-D data (Deussen and

Strothotte, 2000) is that we can use automated tools

to generate the trees. A 3D model of the tree provides

the necessary information about the structure of the

tree. We used L-studio 4.0 (Karwowski and Lane,

2004) to generate the trees used in this paper.

In its basic form, a Lindenmayer system, or L-

system, consists of a starting string of symbols from

an alphabet, and a series of transitions speciﬁed by

a list of search-and-replace rules. During each re-

cursive step, zero or more rule-based transitions are

applied to the string. Each rule consists of an input

substring to be replaced and the string to replace it

with. The ordering of the rules is signiﬁcant. (Fran-

cis, 2002) L-studio also provides several functions to

control the size of the leaves, and the length and size

of the branches based on growth time. The trees used

in this paper have a single leaf shape.

We need to group the leaves based on their level

in the hierarchy (essentially, all leaves produced after

level l are grouped). When L-studio creates the output

ﬁle, it lists the leaves and branches in the order that

they were generated. To determine the group to which

a leaf belongs we use a threshold value based on the

diameter of a branch. The program reads leaves until

the next time that the diameter of a branch crosses

GRAPP 2007 - International Conference on Computer Graphics Theory and Applications

198

the pre-deﬁned threshold value. This triggers a new

group of leaves to be started. This process is repeated

until all of the leaves are read and placed into a unique

grouping.

Leaves and branches constitute the smallest con-

nected grouping in a mesh. We apply one texture to

the branches and a separate one to the leaves. The

meshes belonging to the branches below level l (i.e.,

the trunk and big branches) are separated out and la-

beled as a single group.

3.2 Modifying the Mesh Resolution

The resolution at which L-Studio creates the leaf and

branch meshes is higher than we need, so we simplify

them. Also, the leaf meshes are one sided, so we du-

plicate the leaf mesh, reverse the orientation, and add

thickness.

Textures

No leaf edge

silhouettes

Black leaf edge

silhouettes

Figure 3: Options for drawing a group of leaves. (Left) Tex-

tures used (Center) Leaves drawn without edges hilighted.

(Right) Leaves drawn with edges hilighted. .

The leaf textures (see ﬁgure 3) are a mix of

photographs and hand-painted images. Options for

selecting where to apply texture include leaf only,

branch only, and recursively. The texture can either

be applied to every face or can be stretched out over

every face in a leaf.

The rendering application uses of an id buffer to

determine which pixels a leaf group (or branch) occu-

pies. We assign a unique id to every leaf group and

branch.

4 RENDERING

This section describes the rendering process for cre-

ating a single frame (see ﬁgure 4). The render-

ing process begins by clustering leaves into groups

(blobs). We create approximate, 3D, screen-aligned

geometry for each group (the blob geometry). This

geometry serves two purposes. First, we render the

blob geometry using approximate colors to produce

an abstract rendering of the group. Second, we use the

blob geometry to produce an alpha mask which tells

us where to add in detail. This alpha mask is used to

combine the abstract rendering with a photo-realistic

rendering of the leaves to create a billboard image for

the blob. (Note: By photo-realistic we mean tradi-

tional OpenGL, or ray tracing, Phong, etc., render-

ing.)

Once the billboard images are created, we add

depth and combine them, with the large branches and

trunk, into a single image.

4.1 Lighting

The user speciﬁes two sets of lighting arrangements.

The ﬁrst set of lights, the scene lights, is used to

light the tree in the traditional manner. The second

set of lighting (the blend lighting) is used to deter-

mine where to mix the photorealistic rendering with

the non-photorealistic rendering. This is described in

more detail in section 4.2.2.

4.2 Rendering Blobs

Each group of leaves is processed independently. We

construct several images; an id image which is used

to create the approximate geometry, a rendering of

the approximate geometry, an alpha blend mask, and

the detail rendering. The formula for combining these

images is approximate geometry ∗ alpha blend mask

+ detail ∗ (1 - alpha blend)= ﬁnal.

4.2.1 Blob Approximate Geometry

The goal of grouping the leaves together is to make

a blobby image shape that captures some of the 3D

detail and shading of the leaf group. This process is

done in image space, essentially “shrink-wrapping”

the leaves.

The group of leaves is ﬁrst rendered with all other

branches and leaves turned off. A circle comprised

of typically 20 control points is placed around the

leaves. The user can adjust the number of points used;

twenty provides a balance between creating an inter-

esting shape and abstracting the boundary. The fewer

points used, the smoother the outline will be. The

control points are “shrink wrapped” to the group of

leaves by moving them 20% of the distance between

the current position and the center of the blob. This

process is repeated until the point is on a leaf (see

ﬁg 5b). A B-spline connects the points into a smooth

curve and is also used to draw the blob’s outline.

Next, we create 3D geometry to approximate the

shape of the blob. This geometry takes the form of

NON-PHOTOREALISTIC RENDERING OF ALGORITHMICALLY GENERATED TREES

199

c) Place regional centers, 3D

screen-aligned blob mesh

h) Final blob billboard

d) Blob mesh (viewed

from the side)

b) “Shrink wrap” leaf group

in image plane

a) Render leaf group

f) Blob geometry rendered

with blend lighting (alpha

blend mask)

e) Blob mesh rendered with

average leaf group color,

scene lighting

g) Rendered original 3D

geometry

Figure 4: Creating a single blob. First we render the geometry and shrink wrap it in image space. Second, we construct the

approximate 3D blobgeometry. This geometry is a screen aligned height with the height ﬁeld pointing in the direction of the

viewer. This geometry is used to create two images. The ﬁrst image is the approximate, cartoon-like, rendering. The second is

the alpha-blend mask used to blend between the approximate image and the detailed image. This image becomes a billboard,

trimmed to the shrink wrapped blob outline.

a screen-aligned height ﬁeld, with the height ﬁeld

pointing towards the user. A 10 by 10 unit mesh

square, sufﬁcient to make a smooth-looking blob, is

positioned where the leaves are in 3D space, oriented

to face the viewer. The height is one unit. The mesh is

then scaled so that it covers the leaves in screen space.

To create the height ﬁeld we place three Gaussian

peaks inside of the blob boundary. To ﬁnd the centers

of the peaks, we take the control points of the outline

and divide them into thirds. The average point of one

third of the boundary is then averaged with the center

point of the entire blob to move it slightly inwards.

The color wash used for the blob is an average of

the leaf texture color. The blob geometry is rendered

using the same lighting that is used to render the tree

(see ﬁg 5e).

4.2.2 Adding Detail

The detail texture is created by rendering the leaves.

To emphasize the detail in the leaves, we optionally

let the user highlight the silhouette (Raskar and Co-

hen, 1999) edges of each leaf. This emphasizes the

outline of the leaf and deemphasizes the texture of the

leaf. The silhouette color can vary from black to the

average color of the leaf to white.

To create a combination of photorealism and non-

photorealism, the blob image and the detailed leaves

are blended together. An alpha blend mask is used

to determine how and where to blend. This mask is

created by rendering the blob geometry (colored grey)

with a set of blend lights. Dark regions get the detail

image, light regions the blob one. The user can select

the location of the light and the strength of the light.

At this point there are several saved images: An

image of the leaves with edges (see ﬁg 3), an image

of an outline (see ﬁg 3h), an image of the shape ap-

proximation (see ﬁg 3e), and an image used to calcu-

late the alpha blend mask (see ﬁg 3f). All of these

images are combined into a single image (see ﬁg 3h)

which serves as a billboard image for the group. The

billboard image is clipped to the outline.

To help differentiate neighboring blobs, an outline

GRAPP 2007 - International Conference on Computer Graphics Theory and Applications

200

a) Render trunk and

branches (up to user-

specified level)

b) Blob is rendered as

billboard Billboard depth is minimum depth of corresponding leaf group

Figure 5: Combing branches and blobs into single image.

is drawn around the blob. The color of the outline

is user-controlled blend of the color of the blob and

black (or white). The thickness of the outline is con-

trolled by direction. When the edge is to the left of

a blob image pixel, the outline is drawn wider than if

the edge is in a different direction. A varying thick-

ness provides a greater feeling of depth.

The Painter’s algorithm (Newell et al., 1972) is

used to composite the individual blob images together

(since all of the blob images are screen aligned). The

depth of the blob is determined from the minimum

depth value of any leaf in the blob.

5 ANIMATION

An advantage of computer graphics over traditional

handpainted images is that moving the camera around

the scene is “free”. The biggest challenge is to main-

tain inter-frame image continuity. Rendering each

frame individually causes the boundaries of blobs to

jump, regional centers to bounce around, and blobs to

pop in and out as the depth changes between frames.

To solve these problems, we generate new blobs every

nth frame, then interpolate depth and shape informa-

tion between those frames.

5.1 Pre-rendering

To pre-render, we assume a pre-deﬁned camera path.

A completely new frame is calculated every nth

frame, for a user-deﬁned n. Between the nth frames

the control points for the blob outline and the depth

are interpolated. The regional centers for each blob

remain ﬁxed relative to the changing size of the blob.

The detailed rendering is generated for each frame.

The number of control points used in the bound-

ary is the same between all blobs in all frames. The

boundary between the control points is interpolated

using a cubic B-spline, as is the location of the con-

trol points at each frame.

For each blob in each key frame the minimum

Frame 1 Frame 3 Frame 5

Frame 7 Frame 9 Frame 10

Figure 6: Depth blending over 9 frames.

depth of the leaves in the blob is determined. If one

blob will pass in front of another, we alpha blend be-

tween the two blobs over a period of n frames. This

prevents one blob from “popping” in front of the other

(see ﬁgure 6).

6 RESULTS

In this section we show examples of renderings for a

tree and a bush. We compare adjusting parameters for

the number of control points, light settings for alpha

blob, and color settings for outlines. There is an at-

tached movie that includes rotating and zooming sam-

ples of the results.

The ﬁrst test subject used is a tree. This tree used

a division diameter of 0.042443. The division diam-

eter automatically generates the groupings of leaves

by measuring the diameter of branches. Twenty con-

trol points were used. The edges are black. The alpha

blending controls were set to 70% diffuse, 0% ambi-

ent, and the light position was to the right and up.

The second test subject used is a bush. This bush

used a division diameter of 0.068513. Twenty con-

trol points were used. The edges are black. The alpha

NON-PHOTOREALISTIC RENDERING OF ALGORITHMICALLY GENERATED TREES

201

Figure 7: Bush rendering.

blending controls were set to 70% diffuse, 0% ambi-

ent, and the light position also to the right and up.

The rendering time per frame is approximately

one minute if a purely software-based renderer is

used. This time is proportional to the number of ren-

dered blobs. A tree with approximately 8 blobs takes

30 to 45 seconds to render, while 20 blobs takes 60 to

90 seconds to render.

The number of control points used in the outline

effects the smoothness of the blob shapes (see ﬁg-

ure 8).

The user can also control where detail is placed on

the blob by placing a “blending” light relative to the

view point. The scene is rendered and the brightness

used to control where the detail is placed. Brighter

spots will have less detail then darker ones (see Fig-

ure 9).

The user can also adjust the color of the blob out-

line and the silhouette of the leaves. The images in

ﬁgure 10 illustrate three different settings. Please re-

fer to the attached video. The video shows short clips

of the tree and bush rotating and zooming.

7 DISCUSSION

Trees are an ideal subject for abstraction because they

have both detail and gross structure, and it is fairly

easy to extract this gross structure from the 3D data

in a useful way. Unlike most non-photorealistic tech-

niques, this one replaces the 3D geometry with a sim-

pler 2.5D approximation, rather than applying ﬁlter-

ing. Applying this approach to other geometry, such

as mountains or ﬂowers, would require specifying an

appropriate two or three dimensional geometry sim-

pliﬁcation.

ACKNOWLEDGEMENTS

Thank you to the L-Studio group for the use of their

software. This research was funded in part by NSF

grant CCF-0238062.

REFERENCES

Boddy-Evans, M. (2005). Painting trees: A tips for render-

ing realistic and believable trees. About.com.

Bourdet, S. D. (2001). Painting The Allure of Nature. North

Light Books.

Deussen, O., Hiller, S., van Overveld, C., and Strothotte, T.

(2000). Floating points: A method for computing stip-

ple drawings. Computer Graphics Forum, 19(3):40–

51.

Deussen, O. and Strothotte, T. (2000). Computer-generated

pen-and-ink illustration of trees. In Siggraph 2000,

pages 13–18.

Di Fiore, F., Van Haevre, W., and Van Reeth, F. (2003).

Rendering artistic and believable trees for cartoon an-

imation. In Proceedings of Computer Graphics Inter-

national (CGI 2003), pages 144–151.

Francis, E. M. (2002). L-system.

Freudenberg, B., Masuch, M., and Strothotte, T.

(2002). Real-time halftoning: a primitive for non-

photorealistic shading. In EGRW ’02, pages 227–232.

Karwowski, R. and Lane, B. (2004). L-studio/lpfg: A soft-

ware system for modeling plants.

Kowalski, M. A., Markosian, L., Northrup, J. D., Bourdev,

L., Barzel, R., Holden, L. S., and Hughes, J. (August

1999). Art-based rendering of fur, grass, and trees.

Proceedings of SIGGRAPH 99, pages 433–438.

Lovett, J. (2004). How To Paint Trees.

Luft, T. and Deussen, O. (2006). Real-time watercolor illus-

trations of plants using a blurred depth test. In NPAR,

pages 11—20.

Markosian, L., Kowalski, M. A., Trychin, S. J., Bourdev,

L. D., Goldstein, D., and Hughes, J. F. (1997). Real-

time nonphotorealistic rendering. SIGGRAPH 1997,

31:415–420.

Meier, B. J. (1996). Painterly rendering for anima-

tion. Computer Graphics, 30(Annual Conference

Series):477–484.

Newell, M. E., Newell, R. G., and Sancha, T. L. (1972). A

solution to the hidden surface problem. In ACM’72,

pages 443–450.

Ostromoukhov, V. (1999). Digital facial engraving. In Pro-

ceedings of the 26th annual conference on Computer

graphics and interactive techniques, pages 417–424.

Prusinkiewicz, P. (2004). Art and science of life: Design-

ing and growing virtual plants with l-systems. In ISHS

Acta Horticulturae 630, pages 15–28. Acta Horticul-

turae.

Raskar, R. and Cohen, M. (1999). Image precision silhou-

ette edges. In I3D ’99, pages 135–140.

Smith, A. R. (1984). Plants, fractals, and formal languages.

In SIGGRAPH ’84, pages 1–10.

GRAPP 2007 - International Conference on Computer Graphics Theory and Applications

202

a) 10 points b) 20 points c) 30 points

Figure 8: Number of control points used to create outlines. (a) 10 control points. (b) 20 control points. (c) 30 control points.

b) Blend light lower left

a) Blend light upper right

Figure 9: Blending light. (a) Blending light is placed above and to right of viewer. (b) Blending light if placed below and to

left of viewer.

Leaf edge silhouette: Average color

Blob outlines: Average color + black

Leaf edge silhouette: Black

Blob outlines: Black

Leaf edge silhouette: White

Blob outlines: Average color + black

Figure 10: Changing the silhouette rendering style.

NON-PHOTOREALISTIC RENDERING OF ALGORITHMICALLY GENERATED TREES

203