2006 IEEE Congress on Evolutionary Computation

Sheraton Vancouver Wall Centre Hotel, Vancouver, BC, Canada

July 16-21, 2006

Evolutionary Image Synthesis Using a Model of Aesthetics

Brian J. Ross, William Ralph, and Hai Zong

Abstract— The automatic synthesis of aesthetically pleas- III describes Ralph’s aesthetic model, which is used as one
ing images is investigated. Genetic programming with multi- of the feature tests by the GP system. System design and
objective fitness evaluation is used to evolve procedural texture experimental details are described in Section IV. Section V
formulae. With multi-objective fitness testing, candidate tex-
tures are evaluated according to multiple criteria. Each criteria presents some example results. Concluding remarks are given
designates a dimension of a multi-dimensional fitness space. in Section VI.
The main feature test uses Ralph’s model of aesthetics. This
aesthetic model is based on empirical analyses of fine art, in II. R ELATED W ORK
which analyzed art work exhibits bell curve distributions of
color gradients. Subjectively speaking, this bell-curve gradient Evolutionary texture and image synthesis is well es-
measurement tends to favor images that have harmonious
tablished [3][4]. Most systems are supervised, in which
and balanced visual characteristics. Another feature test is
color histogram scoring. This test permits some control of the user interactively views and rates texture images from
the color composition, by matching a candidate texture’s color the population, and possibly controls the rate of muta-
composition with the color histogram of a target image. This tion [5][6][7][8][9]. Interactive evolution is less suitable for
target image may be a digital image of another artwork. We evolving preconceived styles of textures, than it is as a
found that the use of the bell curve model often resulted in
creative tool for discovering new textures. Moreover, it can
images that were harmonious and easy-on-the-eyes. Without the
use of the model, generated images were often too chaotic or quickly exhaust the user, who must manually view and
boring. Although our approach does not guarantee aesthetically evaluate every generated texture.
pleasing results, it does increase the likelihood that generated Automatic texture evolution removes the user from the
textures are visually interesting. loop [10][11][12]. Image analysis functions evaluate rudi-
mentary image features such as colour, luminosity, and shape.
I. I NTRODUCTION
These scores are then matched with those of a target texture.
The automatic synthesis of aesthetically pleasing images Another approach is in [13], which evolves procedural tex-
is investigated. This research contributes to the body of work tures for 3D models. Instead of image analysis, this system
using evolutionary computation to generate procedural tex- uses training sets of example texture points.
tures. We use genetic programming to synthesize procedural Automated aesthetic texture evolution has not been widely
texture formulae [1]. Principles of Darwinian evolution are studied. A pioneering work is [14], which uses an artificial
applied to a dynamic population of textures. Those that neural network (ANN) as an image evaluation function
generate images with desired visual features are likely to within a GA. Before evolution, the ANN is trained on sets
reproduce, and have offspring with even more desireable of example images that have been deemed to be aesthetically
features. During the evolution process, the rendered output of appealing. Their system does generate textures according
texture expressions are evaluated by two independent feature to the evaluation by the ANN. It is unclear whether their
analyses. The first analysis uses a novel model of aesthetics ANN has learned any relevant aesthetic principles or useful
by Ralph [2]. This model is derived from an empirical study patterns, especially given the vast quantity of training data
of fine art, in which it was discovered that many examples used.
of art work exhibit a normal distribution of colour gradients. More recently, genetic programming using a fitness func-
The other feature analysis is colour histogram scoring, which tion encoding a model of aesthetics has been used in the
compares an image’s colour distribution histogram to that of NEvAr system [15]. That model posits that images are
a target image. By matching the texture’s colour distribution aesthetically pleasing if they are both visually complex, as
with another image, we are borrowing another artist’s colour well as easy for the brain and visual system to perceive
sensibilities. The combined result of these tests are images and interpret. Two mathematical measurements are used: (i)
that conform to both the bell curve aesthetic and desired image complexity is measured by the JPEG compression
colour composition. ratio; (ii) visual self-similarity is indicated by the fractal
The paper is organized as follows. Related work in evo- compression ratio. Their aesthetic model was applied to the
lutionary texture synthesis is reviewed in Section II. Section TDA test (Test of Drawing Appreciation), which is a stan-
dardized psychological test used to evaluate art appreciation
Brian Ross is with the Department of Computer Science, Brock Univer-
sity, St. Catharines, ON L2S 3A1, Canada (phone: 905-688-5550 ext 4284; [16]. The results were impressive, as the model scored better
fax: 905-688-3255; email: [email protected]). than typical art students.
Hai Zong is a software developer at Teilhard Technologies, St. Catharines, Recently, an aesthetic distance metric was proposed as
ON, Canada (email: [email protected]).
William Ralph is with the Department of Mathematics, Brock University, a means for measuring the information proximity between
St. Catharines, ON L2S 3A1, Canada (email: [email protected]). candidate images and a library of images preclassified to be

0-7803-9487-9/06/$20.00/©2006 IEEE 1087

aesthetically pleasing [17]. Although the metric was success- the diagonal length, scaled for typical pixel densities on CRT
fully tested against user preferences during interactive texture monitors. The overall gradient or stimulus S is then:
evolution, it has not been tested extensively in automatic q
texture evolution. Si,j = |∇ri,j |2 + |∇gi,j |2 + |∇bi,j |2
Measurable aspects of artificial and natural phenomena
often exhibit particular kinds of mathematical distributions for the separate RGB channel gradients. Finally, the response
[18]. For example, the well known 1/f or pink noise dis- R is computed as:
tribution is widely found in diverse areas such as physics
[19], natural images [20], and human cognition [21]. One Ri,j = log(Si,j /S0 )
famous example examining 1/f noise in the arts is [22]. In
where S0 is the threshold of detection, which is taken to be
the music examples analyzed, they show that the differences
2. If Si,j = 0 (no change in colour at a pixel), it is ignored.
in successive pitches in notes (pitch gradients) exhibit 1/f dis-
Step 2: The distribution of the response values for an
tributions. Furthermore, stochastically-generated music based
image is determined. The calculation of the distribution is
on 1/f noise generators is more aesthetically pleasing than
based on the hypothesis that the probability that a viewer
that generated by pure random or white noise generators.
pays attention to a detail of an image is proportional to
the magnitude of the stimulus that resides at that detail.
III. A M ATHEMATICAL M ODEL OF A ESTHETICS
Hence we use a weighted mean and standard deviation, with
A mathematical model characterizing an aspect of fine Ri,j being the weight value for each response. The normal
art has been proposed by Ralph [2]. The following gives distribution of R is then estimated using a weighted normal
a simplified overview of the theory. After analyzing hun- distribution, defined by a mean (µ) and standard deviation
dreds of examples of fine art, it was found that many (σ 2 ):
works consistently exhibit functions over colour gradients 2
Ri,j (Ri,j −µ)2
P P
i,j (Ri,j )
that conform to a normal or bell curve distribution. This µ= P
R σ2 = i,j P
i,j i,j i,j Ri,j
is seen with many artists, such has Cezanne and Seurat,
who create “painterly” images. On the other hand, the visual Once µ and σ 2 are found for an image, the actual distribution
responses to photographs and graphic design work usually of all Ri,j for all pixels in the image is tabulated. Using a
do not have normal distributions. This normal distribution bin width of σ/100, a histogram is calculated, where each
of gradient response is claimed to be difficult to realize, Ri,j updates its corresponding bin using a weight of Ri,j .
since it is essentially characterizing the global distribution Step 3: The closeness of fit between the response actual
of gradient throughout an entire image: local changes affect distribution and the hypothesized bell distribution is calcu-
the distribution found throughout the entire painting. It is lated. This is called the deviation from normality or DFN,
hypothesized that this bell curve distribution has been an and it is calculated as:
implicit aesthetic ideal of many painters throughout history. X pi
The bell curve model posits that a viewer’s response DF N = 1000 pi log( )
qi
to an image is largely determined by his or her psycho-
neurological reaction to visual stimuli. When viewing an where pi is the observed probability in the ith bin of the
artwork, a viewer’s visual system is stimulated by the details histogram, and qi is the expected probability assuming a
of the image. The bell curve model suggests that a viewer is normal distribution with the computed mean and standard
most attracted to changes in an image, for example, the edges deviation above. When qi = 0, that bin is ignored. A DFN
between different colours. The areas with constant colours value of 0 means that a perfect normal distribution exists,
are of less interest. Furthermore, larger changes are more while higher DFN values indicate poorer fits to the normal
noticeable than smaller ones. Since it is known that our ner- distribution.
vous system tends to have a logarithmic reaction to stimuli, There are a number of practical advantages of the bell
this model likewise treats measurements logarithmically. curve model. The DFN score is easily incorporated as a
An image’s bell curve gradient is computed in three steps. fitness score within the genetic algorithm. The bell curve
Step 1: An image’s colour gradient is found. This is done model also permits a means for characterizing general artistic
by computing the following for each pixel (i,j) of an RGB styles. As discussed in [2], different styles of art are often
image (ignoring the extrema row and column of the image characterized by their bell curve fit, as well as the associated
buffer): mean and standard deviation of the distribution. Many paint-
ings with good DFN’s tend to have bell curves with a mean
(ri,j − ri+1,j+1 )2 + (ri+1,j − ri,j+1 )2
|∇ri,j |2 = around 3.0 and a standard deviation of 0.75. Photographs
d2 have poor bell distributions, with means of around 4.2 and
where ri,j is the red value at pixel (i,j). Similar computations standard deviations of 1.2 or more. Graphic designs have
are done for the green and blue channels. The d value is even higher values. These values can be used by the genetic
a scaling factor that is used to scale the result for images programming system as general stylistic target areas for
having different dimensions. We take it to be 0.1% of half evolved images.

1088
 Float (f):
Terminals: x, y, ephem characteristics found in many works of art. The gradient
Basic math: plus(f,f), minus(f,f), diff(f,f), mult(f,f), analysis also computes mean and standard deviation scores.
div(f,f), neg(f) The mean denotes the range of gradient seen in an image. The
Other math: sin(f), cos(f), mod(f,f), log(f), pow(f,f)
Relational: min(f,f), max(f,f), if-then-else(f,f,f) standard deviation represents the changes in gradients seen,
Noise: noise(f,f), turb(f,f), turbflow(f,f,f), ranging from gradual shifts to sharp jumps in colour. We find
cloud(f,f,f,f) it useful to include one or both of these measurements with
Misc: tilerad(f,f), lum(v), chn(v), ichn(f,v)
the DFN when analyzing images.
RGB vector (v): The other feature test, CHISTQ, is a colour analysis
Terminals: colgrad, ephem that compares the colour distribution of a texture image
Noise: marble(f,f,v)
Transform: warprel(f,f,v), warpabs(f,f,v), kaleid(v),
with that of a target image. It uses quadratic histogram
vtile(f,f,v) fitting for colours, which is a test often used in query by
Other: rgb(f,f,f), if-then-else(f,v,v), forv(f) image content systems such as VisualSEEk [25]. The image
is first quantized into 1000 colours. Then a histogram of
TABLE I
quantized colour frequencies is tabulated. The histograms for
T EXTURE LANGUAGE
two images are compared with one another, and an overall
distance between them is calculated. All the histogram entries
in both images are compared exhaustively with one another,
IV. E XPERIMENT to determine how closely the colours distributions match.
This test is fairly relaxed with respect to colour fitting. It is
A. Texture language
also position independent, meaning that colours do not have
The texture language used is similar to one used by Sims to coincide with respect to their placement on an image.
[5]. A summary of the language is given in Table I. The
genetic programming system uses strong typing, and so C. Multi-objective fitness scoring
all language components respect the data type conventions A multi-objective problem is characterized by having two
specified in the table. Two data types are used – float (f), or more fitness criteria [26]. Multi-objective search strategies
and RGB vector (v), which is a tuple of 3 float values that consider each feature test as an independent dimension in the
is interpreted as an RGB colour. search space. This contrasts to approaches that merge scores
The float primitives compute floating point values. The together, perhaps by a weighted sum. Weights are usually
terminals include x and y, which are the current pixel ad hoc, introduce undo bias into the search, and can be
coordinates whose texture colour is being computed. The ter- detrimental to most nontrivial problems.
minal ephem denotes ephemeral random constants [1]. These We interpret the evaluation of textures using the multi-
constants are initialized with a random number generator, and objective approach from [12]. When textures are added to a
retain the initialized value throughout the remainder of their new population, the DFN and CHISTQ tests are performed
existence within the run. Their range is between -1.0 and 1.0. on the generated texture from each formulae, and the fitness
Most of the math and relational operators are straightforward. scores (including mean and standard deviation scores) are
The if-then-else(f,g,h) expression evaluates expression f. If saved. These scores are then used to determine a Pareto
f > 0, then the value of g is returned. Otherwise the value of ranking of all the individuals in the population. Pareto
h is returned. The noise operators use a pink noise or Perlin- ranks are based on the idea of domination. One individual
like noise generator. The basic generator is noise(f,g) [23]. dominates another if it is at least as good in all the scores,
It takes 2 explicit arguments, the frequency f and a random and better in at least one. Using the notation u < v to mean
seed g, and 2 implicit arguments, the current x and y pixel score u is more optimal than v, then u dominates v if:
coordinates. Turb, Turbflow, and Cloud use turbulence and ∀i ∈ (1, ..., k) : ui ≤ vi ∧ ∃i ∈ (1, ..., k) : ui < vi
cloud modeling formulae [23][24]. Tilerad generates a tiling
effect. Individuals having rank 1 are undominated, and are the
The vector functions return RGB vectors as evaluated current best solutions in the population. Those of rank k
results. Vector terminals include ephemeral constants, as > 1 are dominated by all the individuals of ranks < k. All
well as colgrad, which computes an RGB gradient based the individuals in a rank are incomparable with one another.
on the current pixel coordinates. The other vector primitives At the end of a run, all those with rank 1 are considered as
perform a variety of texture and tiling effects, vector con- valid solutions to the problem.
struction and deconstruction, and vector iteration. After the Pareto ranks are determined, the individuals in
each rank are evaluated with respect to their diversity. Tex-
B. Feature evaluation tures in a Pareto rank are considered superior if they are more
The relative worth of a texture image will be measured by diverse with respect to their location in the multi-dimensional
its performance on a suite of feature tests. One feature test is search space, as indicated by the fitness scores in each feature
the bell curve analysis from Section III. The deviation from test. The idea is that a texture with unique characteristics,
normality (DFN) determines how well an image’s gradient likely has correspondingly different feature vector values.
fits to a normal distribution, and hence shares the gradient Likewise, textures that are duplicates or minor variations of

1089
each other probably have very similar or identical feature test 3 random formulae are selected from the population, and
scores. the one with the highest fitness is retained for reproduction.
An indicator of diversity in feature space is the nearest- Crossover applies tournament selection twice, to find two
neighbour distance between feature vectors. Because feature parents.
tests differ widely between each other in terms of valuations, Images rendered during the run are by necessity small,
the raw scores in the vectors are not used to determine given the computational time needed for their generation and
distances. Rather, these distance scores are abstracted into analysis. We often used run-time images with resolutions of
ranks, and the average distance rank is computed instead. approximately 125 by 100, and final rendered output of 1250
This approach will give all feature scores equal weight when by 1000, or 100 times the area. Although smaller run-time
determining distance. The following steps are performed to images greatly speed up a run, a price is paid in feature test
determine diversity for all the textures in each Pareto rank accuracy. A larger resolution image has much more visual
set: detail, which affects the DFN and other scores. Therefore,
1) For each feature test dimension, the nearest-neighbour larger run-time images will result in more accurate results.
distance is determined. If an appropriately-scaled Gaussian blur filter is applied to
2) All the distances are ranked, with the largest distance a large image before feature tests are performed, the results
assigned the lowest rank (1). Each individual ends up obtained are closer to what are seen with smaller resolution
with a vector of distance ranks, with one distance rank versions of the image.
per feature test.
3) The average distance rank is deterined for each dis- V. R ESULTS
tance rank vectors. This is used as the “diversity score”,
where lower averages are preferred. Figure 1 shows some results.1 The search uses four target
Once these steps are computed, the textures within each objectives: DFN=0.0, mean=3.75, std dev=0.75, and colour
Pareto rank set are re-assigned fitness scores, such that more matching with the image in (a) (ie. CHISTQ=1.0). Textures
diverse individuals in the rank have better scores compared that are to be analyzed in the run are rendered at a resolution
to others in the rank set. The desired outcome is a diversity of 126 by 98. The final results are rendered at 1260 by
of generated images. Note that diversity is only applied 980. One result is the image in (b). When processed with
within the Pareto sets, and the Pareto ranks themselves are a 3 by 3 gaussian filter (to approximate the detail in the
maintained. In other words, diversity does not affect the smaller image analyzed in the run), the resulting statistics
overall Pareto ranking between Pareto rank sets. are DFN=6.4, mean=1.8, std dev=0.68, CHISTQ=0.82. Such
results are typical with multi-objective searches, in which
D. Other experimental parameters some scores are strong, while others are weak (in this case,
the mean score). Usually a CHISTQ score above 0.80 means
Parameter Value
Population size 1000
that the image colour yields an acceptable match to the
Generations 50 target. Image (c) shows the gradient response filter applied
Runs/experiment 5 to the solution image. Figure 2 shows the distribution of
Initialization ramped half&half
Initial ramped tree depth 2 to 6
the gradient response for this solution. Figure 3 shows how
Max. tree size (nodes, depth) 200, 20 the DFN stays low as the gaussian filter is increased, which
Crossover rate 0.9 simulates the eye viewing the image at farther distances.
Mutation rate 0.1
Selection scheme tournament (size 3)
Images (d), (e) and (f) in Figure 1 are other results using
the same target image as above. Image (d) has a DFN=7.9,
TABLE II mean=3.7, std dev=0.54, and CHISTQ=0.92. Image (e) has
G ENETIC P ROGRAMMING PARAMETERS ( TYPICAL ) a DFN=15.0, mean=4.6, std dev=0.72, and CHISTQ=0.90.
Image (f) is taken from an earlier generation (25). Less
evolved images such as this one are typically bolder, more
Table II lists other typical parameters used in the exper- primitive, and less refined than latter ancestors, since their
iments. Most are standard within the genetic programming scores are still far from the target values.
literature [1]. Five separate runs using new random number Figure 4 shows some results from an experiment in which
seeds were done per experiment, using a population of 1000 a greyscale image in (a) is used as the colour target.
texture expressions. A maximum of 50 generations were In Figure 5, the Union Jack in (a) is the colour target. That
used. The initial population is generated using the ramped image is detrimental to both the CHISTQ and DFN tests, as
half-and-half tree generation algorithm, which creates a pop- both prefer richer gradients than the 3 colours in the image.
ulation of random trees having an assortment of sizes and One interesting result is in (b). To get a lower DFN score,
shapes. Initial trees have depths between 2 to 6 levels. During a 3D smoke effect was placed between the red and yellow
reproduction, trees can have a maximum of 200 nodes, and bands, as is seen in the detail in (c).
a maximum depth of 20. Crossover is used 90% of the
time, and mutation is selected the remaining 10%. Fitness- 1 Full-sized colour images can be seen at www.cosc.brocku.ca/ ∼
based selection is performed via tournament selection. Here, bross/DF N gallery/.

1090
 (a) Colour target. (b) One result. (c) Colour gradient of (b).

(d) (e) (f)

Fig. 1. Results for Monet sea image palette in (a).

(a) Greyscale target. (b) (c)

(d) Detail from (b). (e) Detail from (c)

Fig. 4. Greyscale colour target experiment

A classic painting by Monet is used as the target in Figure It was evolved in a run that ignored DFN scores, and its
6. Image (b) is one result, and a detail from it is in (c). DFN is over 1400. Images in (c) and (d) show how the
DFN scoring can produce subtlety and balance. Image (e)
Figure 7 shows various results from other experiments. is a detail from an image that had a wood-block print effect.
Images (a) and (b) are examples of poor results having high The image in (f) is from a run in which the target DFN was
DFN scores. The chaotic nature of image (a) gives it a inadvertently set too high. The resulting artistic style is more
DFN=129. On the other hand, image (b) is overly bland.

1091
(a) Colour target. (b) Result. (c) Detail from (b).

Fig. 5. Union Jack experiment

(a) Colour target. (b) Result. (c) Detail from (b).

Fig. 6. Monet experiment

(a) DFN=129 (b) DFN > 1400 (c)

(d) (e) (f)

Fig. 7. Miscellaneous results

1092
 an aesthetic ideal implicitly striven for by many classical
painters. Graphical art and photographs tend to be in the
area marked with the green circle, and did not have strong
DFN scores. We found that our evolved images tended to
reside in the light blue rectangle. When the colour palette
scoring was not performed, the DFN scores were consistently
in the bottom-left of the triangle, in the area of low mean
and standard deviation. The sweet spot was occasionally hit
by our images. As with human artists, the sweet spot is also
technically challenging for genetic programming.

Fig. 2. Gradient plot for Fig.1 (b) VI. C ONCLUSION

This paper has shown that aesthetic texture evolution
benefits from Ralph’s bell curve response model [2]. We
conjecture that the area of image space defined by low DFN
scores is more densely populated with interesting images,
than the areas where the DFN is high. When experiment
parameters were appropriately tuned, we conservatively es-
timate that 10% of our results were visually interesting. On
the other hand, when the bell curve analysis is removed, the
resulting images are chaotic or bland, artificial, mathematical,
and rarely appealing. Thus, the bell curve score is a heuristic
that directs the search into regions of texture space that are
more likely to be balanced, harmonious, painterly, and “easy
on the eyes”. This is important for automatic synthesis, where
Fig. 3. DFN vs blurring for Fig.1 (b)
the user has no control over the direction of search. Never-
theless, the user still makes the ultimate decision whether
graphic design-oriented than the more ambient and painterly the synthesized images are appealing. Art is in the eye of
styles seen with images having low DFN’s (see discussion the beholder.
of Figure 8 below). Ralph’s bell curve response models the degree of visual
non-uniformity that is resident in a large number of classical
paintings. This is not to suggest, however, that all synthetic
images with low DFN’s are necessarily interesting. For
example, we performed runs in which the DFN was the
sole criteria, and the colour score was ignored. Although
the resulting images had excellent DFN scores, they were
inevitably bland and boring, due to the use of a narrow band-
width of RGB space. This was advantageous for obtaining
low DFN scores, but not very interesting to the human eye.
The colour test introduces the need for a palette of colours,
which is the norm in visual art. In fact, the mood of a target
image as intended by the original artist is often injected into
the texture image, visaviz the colour palette.
Interestingly, we found that some images with unusually
poor scores were visually fascinating. This does not suggest
that the fitness criteria is irrelevant. These particular im-
Fig. 8. Gradient response space for art work and evolved images ages have visually pleasing traits inherited from more well-
behaved and refined ancestors. Furthermore, sometimes the
The research in [2] characterizes the observed normal chosen colour target image is “DFN unfriendly”. In such
distributions of art work with the graph in Figure 8. The dark cases, the DFN and colour tests work against each other.
blue pyramid demarcates the area within which the response The resulting “creative tension” often produces the most
curves of paintings are often found. Paintings with low surprising results (image (c) in Figure 5).
DFN scores consistently have means and standard deviations The images in [15] share some stylistic characteristics with
within a “sweet spot” in the pyramid, marked with the many of our results. The aesthetic models of both systems are
red circle. Amateur artists very rarely create pieces that unfavourable to images that are either too static or chaotic.
reside in this locale. This suggests that the sweet spot is Their model’s use of fractal compression tends to favour

1093
images that are self-similar in nature. The NEvAr system [6] J. Graf and W. Banzhaf, “Interactive Evolution of Images,” in Proc.
uses greyscale, while we work with a target colour palette. Intl. Conf. on Evolutionary Programming, 1995, pp. 53–65.
[7] A. Rowbottom, “Evolutionary Art and Form,” in Evolutionary Design
Both the Bell distribution and NEvAr’s aesthetic models by Computers, P. Bentley, Ed. Morgan Kaufmann, 1999, pp. 330–365.
likely reside within 1/f space. The 1/f distribution is applica- [8] M. Lewis, “Aesthetic Evolutionary Design with Data Flow Networks,”
ble to a wide variety of phenomena that reside between total in Proc. Generative Art 2000, 2000.
[9] S. Rooke, “Eons of Genetically Evolved Algorithmic Images,” in
order and total chaos [18]. Its generality makes it too coarse Creative Evolutionary Systems, P. Bentley and D. Corne, Eds. Morgan
for modeling more refined phenomena, such as fine art. Kaufmann, 2002, pp. 330–365.
Research in [27] has applied many of this paper’s ideas, [10] A. Ibrahim, “GenShade: an Evolutionary Approach to Automatic and
Interactive Procedural Texture Generation,” Ph.D. dissertation, Texas
including the Bell curve aesthetic model, toward the evolu- A&M University, December 1998.
tion of non-photorealistic image filters. The results of that [11] A. Wiens and B. Ross, “Gentropy: Evolutionary 2D Texture Gener-
research confirm that the Bell curve model acts as a taming ation,” Computers and Graphics Journal, vol. 26, no. 1, pp. 75–88,
February 2002.
influence to the kinds of filter effects obtained. Furthermore, [12] B. Ross and H. Zhu, “Procedural Texture Evolution Using Multiob-
that research discovered that paint brush strokes are con- jective Optimization,” New Generation Computing, vol. 22, no. 3, pp.
ducive to Bell curve gradient distributions. This partially 271–293, 2004.
[13] A. Hewgill and B. Ross, “Procedural 3D Texture Synthesis Using
explains why we occasionally obtained paint stroke effects Genetic Programming,” Computers and Graphics, vol. 28, no. 4, pp.
on textures, which are not an easy effect to obtain with our 569–584, 2004.
GP texture language and its noise generators. [14] S. Baluja, D. Pomerleau, and T. Jochem, “Towards Automated Artifi-
cial Evolution for Computer-generated Images,” Connection Science,
There are a number of future directions for this research. vol. 6, no. 2/3, pp. 325–354, 1994.
Other measurable aspects of aesthetics could be incorporated [15] P. Machado and A. Cardoso, “All the Truth About NEvAr,” Applied
that address composition and colour selection in more detail. Intelligence, vol. 16, no. 2, pp. 101–118, 2002.
[16] ——, “Computing Aesthetics,” in Proc. XIVth Brazilian Symposium
The evolution of images that have non-normal distributions on AI. Springer-Verlag, 1998, pp. 239–249.
(higher DFN’s) is worth considering. Image (f) in Figure 7 [17] N. Svangard and P. Nordin, “Automated Aesthetic Selection of
shows that interesting graphical designs are possible with Evolutionary Art by Distance Based Classification of Genomes and
Phenomes using the Universal Similarity Metric,” in Applications of
non-bell distributions. New texture languages would result Evolutionary Computing: EvoWorkshops 2004. Springer, 2004, pp.
in new styles of images. The language used here is predomi- 447–456, lNCS 3005.
nated by noise generators, which directly influences the style [18] M. Schroeder, Fractals, Chaos, Power Laws. W.H. Freeman and
Company, 1991.
of abstract art obtained. [19] L. Kiss, Z. Gingl, Z. Marton, J. Kertesz, F. Moss, G. Schmera, and
A. Bulsara, “1/f Noise in Systems Showing Stochastic Resonance,”
ACKNOWLEDGMENT Journal of Statistical Physics, vol. 70, no. 1/2, pp. 451–462, 1993.
Thanks to Andrea Wiens and Han Zhu for their earlier [20] D. Tolhurst, Y. Tadmor, and T. Chao, “Amplitude spectra of natural
images,” Ophthalmic & Physiological Optics, vol. 12, no. 2, pp. 229–
work on Gentropy. B.J. Ross was supported by NSERC 232, 1992.
Operating Grant 138467-1998, and H. Zong by an NSERC [21] J. Pressing, “Sources for 1/f noise effects in human cognition and
USRA award. performance,” Paideusis: Journal for Interdisciplinary and Cross-
Cultural Studies, vol. 2, 1999.
[22] R. Voss and J. Clarke, “1/f noise in music: Music from 1/f noise,” J.
R EFERENCES Acoustical Society of America, vol. 63, no. 1, pp. 258–263, 1978.
[1] J. Koza, Genetic Programming: On the Programming of Computers [23] D. Ebert, F. Musgrave, D. Peachey, K. Perlin, and S. Worley, Texturing
by Means of Natural Selection. MIT Press, 1992. and Modeling: a Procedural Approach. Academic Press, 1994.
[2] W. Ralph, “Painting the Bell Curve: The Occurrence of the Normal [24] H. Elias. (2005) Cloud Cover. [Online]. Available:
Distribution in Fine Art,” In preparation., 2006. http://freespace.virgin.net/hugo.elias/models/m clouds.htm
[3] A. Dorin, “Aesthetic Fitness and Artificial Evolution for the Selection [25] J. Smith and S.-F. Chang, “Visualseek: a fully automated content-
of Imagery from the Mythical Infinite Library,” in Advances in based image query system,” in Proc. ACM Intl. Conf. on Multimedia
Artificial Life – Proc. 6th European Conference on Artificial Life. (ACM-MM), 1996, pp. 87–98.
Springer-Verlag, 2001. [26] C. C. Coello, D. V. Veldhuizen, and G. Lamont, Evolutionary Al-
[4] M. Whitelaw, “Breeding Aesthetic Objects: Art and Artificial Evolu- gorithms for Solving Multi-Objective Problems. Kluwer Academic
tion,” in Creative Evolutionary Systems, P. Bentley and D. Corne, Eds. Publishers, 2002.
Morgan Kaufmann, 2002, pp. 129–145. [27] C. Neufeld, B. Ross, and W. Ralph, “The Evolution of Artistic Filters,”
[5] K. Sims, “Interactive evolution of equations for procedural models,” 2005, submitted.
The Visual Computer, vol. 9, pp. 466–476, 1993.

1094