Generative art and creative coding: algorithms as aesthetic medium

Intuition Beginner

Generative art is a way of making art in which the artist builds a process — a set of rules, an algorithm, or a system — and lets that process produce the actual image, sound, or animation. The artist does not paint every line by hand. Instead the artist writes the instructions, and a computer (or a person following the instructions) carries them out. The result can surprise even its creator.

This idea is older than computers. In the 1960s, artists like Vera Molnar, Manfred Mohr, Frieder Nake, and A. Michael Noll used mainframe computers and mechanical pen plotters to draw geometric patterns. They wrote programs that turned mathematical choices into ink on paper. Around the same time, the American artist Sol LeWitt made wall drawings defined only by written sentences, with other people doing the drawing.

A simple example shows the idea. Imagine a rule: place one thousand dots on a page, where each dot's position is chosen by a random number, and connect any two dots that sit closer than a fixed distance. You wrote one short rule, but you get a complex web of lines you could never have planned by hand. That gap between a short rule and a rich result is the heart of generative art.

Some generative artists borrow patterns from nature. Conway's Game of Life turns a few survival rules into crawling, growing shapes. Ken Perlin invented a "noise" function in 1985 that makes textures looking like clouds, marble, and terrain. Aristid Lindenmayer's L-systems use rewriting rules to grow plant-like branches. Alan Turing described in 1952 how chemicals could spontaneously form stripes and spots.

A key question follows. If the machine produces the pixels, who is the artist? Most people credit the designer of the rules. The algorithm is a tool, like a brush, but a brush cannot make its own choices. This is what makes generative art philosophically interesting: the creative act moves from the finished surface to the design of the system itself.

Visual Beginner

Lineage	Period	Idea	Representative figures
Early computer / plotter art	1960s	Algorithms drawn by pen plotters	Vera Molnar, Manfred Mohr, Frieder Nake, A. Michael Noll
Conceptual-generative precursors	1960s-2007	Instruction sets executed by others	Sol LeWitt (wall drawings)
Cellular automata art	1970-present	Local rules make global patterns	Conway (Game of Life), McCabe-Parkes
Procedural natural texture	1985-present	Smooth noise for organic surfaces	Ken Perlin
L-systems	1968-present	Grammar rewriting grows plants	Aristid Lindenmayer
Reaction-diffusion	1952-present	Chemical patterns form stripes/spots	Alan Turing
Evolutionary art	1980s-present	Selection breeds images	Karl Sims, William Latham
Creative coding platforms	2001-present	Code-as-medium for artists	Casey Reas, Ben Fry (Processing, p5.js)
Contemporary generative practice	2018-present	Algorithm as long-form aesthetic	Tyler Hobbs (Autumn, Fidenza)

Worked example Beginner

Let us compute one value of a two-dimensional Perlin noise texture at a point. This is the building block behind cloud, marble, and terrain textures.

Pick a query point with coordinates $x=3.4$ and $y=2.7$. The grid cell holding this point has corners at $(3,2)$, $(4,2)$, $(3,3)$, and $(4,3)$. At every grid corner the system stores a fixed random direction, called a gradient, such as "point up-right" or "point down".

For each corner, form a vector from the corner to the query point, then measure how much that vector lines up with the corner's gradient. A corner whose gradient points roughly toward the query point gives a positive value; a corner whose gradient points away gives a negative value. This produces four signed numbers, one per corner.

Next, smooth the position inside the cell. The fractional offsets are $u=0.4$ and $v=0.7$. A shaping function bends these so that the blend is flat near the cell edges, avoiding visible seams. The bend is $u^{\prime }=u^3(6u^2-15u+10)$, giving $u^{\prime }\approx 0.43$, and similarly $v^{\prime }\approx 0.85$.

Finally, blend. Combine the two lower-corner values using $u^{\prime }$, combine the two upper-corner values using $u^{\prime }$, then combine those two results using $v^{\prime }$. The single output number might be $0.12$. Color every pixel by its own output and the whole canvas fills with smooth, cloud-like texture instead of a blocky checkerboard.

Check your understanding Beginner

Formal definition Intermediate+

A generative art system is a function $G:\mathcal{S}\times \mathcal{P}\to \mathcal{O}$, where $\mathcal{S}$ is a space of seeds (sources of variation), $\mathcal{P}$ is a space of artist-controlled parameters, and $\mathcal{O}$ is an output space (an image, animation, or sound field). The artist fixes $G$ and selects parameters $p\in \mathcal{P}$; the seed $s\in \mathcal{S}$ picks one concrete realization. The pair $(G,p)$ is the work; each $s$ yields one instance. This makes precise the division of authorship between intention (encoded in $G$ and $p$) and autonomy (encoded in $s$ and in $G$'s internal logic).

Value noise on a lattice ${\mathbb{Z}}^2$ assigns to each lattice point $(i,j)$ a fixed pseudo-random scalar $V_{i,j}\in [-1,1]$. The value at a point $(x,y)$ is a bilinear blend of the four corner scalars, weighted by smooth falloff functions of the fractional offsets $u=x-\lfloor x\rfloor$, $v=y-\lfloor y\rfloor$. Because it interpolates scalar heights, value noise leaves faint square grid artifacts visible in its isolines.

Gradient (Perlin) noise instead assigns to each lattice point $(i,j)$ a fixed pseudo-random unit gradient ${\mathbf{g}}_{i,j}\in S^1$. The contribution of a corner is the dot product ${\mathbf{g}}_{i,j}\cdot (\mathbf{x}-(i,j))$ between that gradient and the vector from the corner to the query point, and the four contributions are blended by the quintic fade $f(t)=t^3(6t^2-15t+10)$ applied to $u$ and $v$. Because the value is zero to first order at each lattice point, the lattice structure is far less visible.

An L-system is a rewriting grammar $L=(\Sigma ,\omega ,P)$, where $\Sigma$ is an alphabet, $\omega \in {\Sigma }^{\ast }$ is the axiom, and $P\subseteq \Sigma \times {\Sigma }^{\ast }$ is a set of production rules. Beginning from $\omega$, all applicable rules are applied in parallel each generation. Interpreting letters as turtle-graphics commands yields self-similar branching forms that model plant growth.

A two-state cellular automaton is a tuple $(\Lambda ,Q,N,f)$ with lattice $\Lambda$ (often ${\mathbb{Z}}^2$), finite state set $Q$, a neighborhood $N$, and a local update $f:Q^N\to Q$ applied synchronously to every cell. Conway's Game of Life uses $Q=\{0,1\}$, the Moore neighborhood of eight cells, and rules in which a cell lives in the next generation if it has two or three live neighbors.

Reaction-diffusion, after Turing, models two chemicals with concentrations $u(x,y,t)$ and $v(x,y,t)$ evolving as $\partial u/\partial t=D_u{\nabla }^{2}u-u{v}^2+f(1-u)$ and $\partial v/\partial t=D_v{\nabla }^{2}v+u{v}^2-(f+k)v$, producing stable spot and stripe patterns from uniform initial conditions.

Key result Intermediate+

Result (gradient noise is band-limited and approximately isotropic). Let $n:{\mathbb{R}}^2\to \mathbb{R}$ be Perlin gradient noise built from unit gradients and the quintic fade $f(t)=t^3(6t^2-15t+10)$. Then (i) $n$ is of class $C^2$ everywhere, (ii) the power spectrum $|\hat{n}(\mathbf{\xi }){|}^2$ is concentrated at spatial frequencies $|\mathbf{\xi }|\lesssim 2\pi$, and (iii) the expected autocorrelation is invariant in expectation under rotation of the query direction, so the noise has no preferred axis. By contrast, value noise based on bilinear interpolation of lattice scalars is only $C^1$ and exhibits visible square lattice artifacts along rows and columns.

Sketch. The quintic fade $f$ satisfies $f(0)=0$, $f(1)=1$, and $f^{\prime }(0)=f^{\prime }(1)=f^{\prime \prime }(0)=f^{\prime \prime }(1)=0$. Because the second derivative of the falloff vanishes at the lattice points, the piecewise polynomial pieces join with matching value, slope, and curvature, giving global $C^2$ regularity; this is exactly the regularity gain over the cubic Hermite fade $3t^2-2t^3$, which is only $C^1$. Within a cell, each corner contribution ${\mathbf{g}}_{i,j}\cdot (\mathbf{x}-(i,j))$ is linear in $\mathbf{x}$, and the product of two one-dimensional quintic weights is a polynomial of degree at most ten; its Fourier transform is therefore a Schwartz-class object concentrated at low frequency. Averaging over the random gradients, $\mathbb{E}[\mathbf{g}{\mathbf{g}}^{\top }]=\frac{1}{2}I$ on the unit circle, so no lattice axis is preferred and the expected spectral density is rotationally symmetric. Value noise, lacking the gradient cancellation (corner contributions need not vanish at lattice points), retains a square-periodic component whose Fourier signature sits at the lattice harmonics, visible as faint grid lines.

Bridge. This is exactly why a quintic fade and random gradients, and not a smoother blend of random heights, became the substrate of three decades of texture synthesis: the central insight is that killing the second derivative at the lattice corners makes the grid invisible to the eye. The result generalises to three and four dimensions (volume and animated noise) and is dual to the reaction-diffusion patterns of Turing, where diffusion smooths high frequencies by physical rather than interpolation means. Putting these together, the bridge is that all naturalistic generative texture is a story of controlled low-pass filtering, and the pattern recurs whenever an artist needs chaos that still looks organic; it builds toward the procedural worlds of contemporary creative coding and appears again in the reaction-diffusion and L-system derivations below.

Exercises Intermediate+

Advanced results Master

The deeper theory of generative texture is spectral. A noise function is a random field, and its visual character is governed by its power spectrum $|\hat{n}(\mathbf{\xi }){|}^2$. Pure white noise has a flat spectrum and looks like television static — every frequency equally present, no spatial coherence. Natural phenomena have $1/f^{\beta }$-type spectra: power falls off with frequency, so large smooth features dominate and fine detail contributes progressively less. The Perlin kernel, being a compactly supported blend of low-degree polynomials, produces a field whose spectrum rolls off smoothly above the lattice frequency, approximating the pink-ish character of clouds and terrain. Fractal Brownian motion (fBm) generalises this by summing octaves $n(2^k\mathbf{x})/2^k$ for $k=0,1,2,\dots$, stacking self-similar copies at increasing frequency and decreasing amplitude to build turbulence, mountains, and marble veins ^{[perlin1985 Perlin 1985]}.

Reaction-diffusion, formulated by Turing in 1952 ^{[turing1952 Turing 1952]}, offers a second route to natural pattern through dynamics rather than interpolation. Two chemicals that diffuse at different rates and react autocatalytically can drive an initially uniform state unstable: a perturbation at a favored wavelength grows while others decay, a linear instability now called a Turing bifurcation. The fastest-growing wavenumber is set by the ratio of diffusion constants, so the system selects spot or stripe spacing from the physics rather than from an artist's choice. Gray, Murray, and Koch extended this into a quantitative theory of mammalian coat markings, fish scales, and finger ridges, giving generative artists a biologically grounded pattern generator.

Cellular automata compress the same logic into discrete space. Conway's Game of Life, a two-state automaton on a square lattice with the Moore neighborhood, is computationally universal — it can simulate any Turing machine — yet emerges from four lines of survival rules. McCabe and Parkes's generative works (exhibited from 2006 onward) ^{[mccabe2006 McCabe and Parkes 2006]} use continuous-valued automata and reaction-like update rules to produce paintings that evolve frame by frame, treating the canvas as a small ecosystem. The aesthetic point is that striking global order is recoverable from aggressively local rules.

Lindenmayer's 1968 grammars ^{[lindenmayer1968 Lindenmayer 1968]} model development by parallel rewriting, distinguishing them from Chomsky grammars that rewrite one symbol at a time. This parallelism captures the simultaneous division of cells in a growing organism. Adding stochastic rule selection, parametric interpretation, and age-dependent productions yields trees, algae, and entire procedural ecosystems; Prusinkiewicz later unified this with the biological theory of phyllotaxis to reproduce exact plant architectures from measured parameters.

Evolutionary art, pioneered by Karl Sims and William Latham in the late 1980s and early 1990s, externalises authorship further: the artist defines a genotype (a symbolic expression tree), a mutation operator, and a fitness criterion, then breeds images by selection. Sims's 1991 work let a human act as the fitness function, choosing which mutants survived, turning the artist into a gardener of forms. This raises a sharp version of the authorship question — the artist wrote neither the final image nor, in a strong sense, the rule that produced it, only the loop that searches the space.

The arrival of generative AI reframes this lineage rather than breaking it. A diffusion model is still a generative system $G(s,p)$: $p$ is now a text prompt plus conditioning, $s$ is the initial noise tensor, and $G$ is the learned denoising flow. The continuity is real — Tyler Hobbs's Autumn (2018) and the rule-based tradition are the conceptual ancestors of prompt-driven work — but the locus of aesthetic labor shifts from writing a compact law to curating a high-dimensional latent space. The algorithm as aesthetic statement survives; the algorithm simply becomes very large and learned rather than small and written.

Synthesis. The central insight unifying every technique in this unit is that organic-looking imagery is controlled low-pass structure seeded by randomness: the foundational reason Perlin noise, Turing patterns, and fBm all look natural is that each suppresses high-frequency artifacts while preserving a smooth, self-similar low-frequency body. This is exactly what the fade function enforces by hand and what diffusion enforces by physics, so the pattern recurs across interpolation, reaction, and learned generators; it generalises from two dimensions to volumes and animations, and it is dual to the spectral view in which an artist composes by shaping power rather than pixels. Putting these together, the bridge is that generative art is not randomness made visible but randomness made smooth, and that single idea — shape the spectrum, then vary the seed — builds toward every contemporary procedural world and appears again in the formal proof of why the quintic fade is the minimal polynomial that does the job.

Full proof set Master

Proposition (uniqueness of the quintic Perlin fade). Among real polynomials $f$ of degree at most five satisfying the endpoint Hermite conditions

$$f(0)=0,f(1)=1,f^{\prime }(0)=f^{\prime }(1)=f^{\prime \prime }(0)=f^{\prime \prime }(1)=0,$$

there is exactly one, namely $f(t)=6t^5-15t^4+10t^3=t^3(6t^2-15t+10)$. This polynomial is the lowest-degree blend that makes one-dimensional Perlin noise $C^2$ at every lattice boundary, and no degree-four polynomial can achieve this.

Proof. Write the general quintic as $f(t)=a_0+a_{1}t+a_2{t}^2+a_3{t}^3+a_4{t}^4+a_5{t}^5$ with derivatives $f^{\prime }(t)=a_1+2a_{2}t+3a_3{t}^2+4a_4{t}^3+5a_5{t}^4$ and $f^{\prime \prime }(t)=2a_2+6a_{3}t+12a_4{t}^2+20a_5{t}^3$.

Apply the six Hermite conditions. From $f(0)=0$ we get $a_0=0$. From $f^{\prime }(0)=0$ we get $a_1=0$. From $f^{\prime \prime }(0)=0$ we get $2a_2=0$, hence $a_2=0$. The polynomial therefore has the form $f(t)=a_3{t}^3+a_4{t}^4+a_5{t}^5$. The three remaining conditions are

$$f(1)=a_3+a_4+a_5=1,f^{\prime }(1)=3a_3+4a_4+5a_5=0,f^{\prime \prime }(1)=6a_3+12a_4+20a_5=0.$$

This is a linear system $M\mathbf{a}=\mathbf{b}$ with

$$M=\left(\begin{array}{ccc}1& 1& 1\\ 3& 4& 5\\ 6& 12& 20\end{array}\right),\mathbf{b}=\left(\begin{array}{c}1\\ 0\\ 0\end{array}\right).$$

Its determinant is $\det M=1(80-60)-1(60-30)+1(36-24)=20-30+12=2\ne 0$, so the solution is unique. Solving, subtract three times the first row from the second and six times the first from the third to obtain $a_4+2a_5=-3$ and $6a_4+14a_5=-6$; dividing the latter by two gives $3a_4+7a_5=-3$, and combining with $a_4+2a_5=-3$ yields $a_5=6$, then $a_4=-15$, then $a_3=10$. Hence $f(t)=10t^3-15t^4+6t^5=t^3(6t^2-15t+10)$, the claimed kernel.

To see no quartic can suffice, note that the six Hermite conditions impose six independent linear constraints, but a quartic has only five coefficients; the system is overdetermined. Concretely, setting $a_5=0$ in the reduced system forces $a_4=-3$, $a_3=0$, giving $f(t)=-3t^4$, which violates $f(1)=1$. The cubic Hermite kernel $3t^2-2t^3$ satisfies the four $C^0,C^1$ conditions but not the two $C^2$ conditions (its second derivative is $\pm 6$ at the endpoints, as shown in the exercises), so $C^2$ regularity genuinely requires degree five.

Because the second derivative vanishes at every lattice point, the piecewise-polynomial pieces of one-dimensional Perlin noise join with matching value, slope, and curvature, giving global $C^2$ regularity; in two dimensions the same argument applies to the tensor-product kernel $f(u)f(v)$, ensuring no crease is visible along any grid line. $\square$

Connections Master

• Digital-media foundations 34.08.01. This unit is the depth companion to the digital-media-art chapter: where the foundational unit surveys the whole field from video art to NFTs, this one specializes to the generative subfield and supplies the mathematical machinery (the fade kernel, gradient noise, L-systems) that the survey only names. The formal model $G:\mathcal{S}\times \mathcal{P}\to \mathcal{O}$ used here sharpens the survey's prose definition of generative art into something a programmer can implement.

• Art history and conceptualism 34.04.01. The lineage from Sol LeWitt's instruction-based wall drawings to algorithmic code is a direct line within twentieth-century art history: LeWitt's claim that "the idea becomes a machine that makes the art" is the conceptual bridge between post-minimal conceptual practice and the generative tradition of Molnar, Reas, and Hobbs. Understanding where generative art sits in the canon requires the modern-contemporary framing that unit supplies.

• Visual-art elements and composition 34.03.01. A generative artist still composes with line, shape, color, balance, and rhythm; the algorithm merely automates their placement. The design vocabulary for evaluating a Tyler Hobbs or Casey Reas piece — figure/ground, hierarchical proportion, color harmony — is exactly the vocabulary of the visual-art-elements unit, now applied to work whose surface no human hand touched.

Historical & philosophical context Master

The pioneers of computer art worked against serious institutional resistance. In the early 1960s, A. Michael Noll at Bell Telephone Laboratories produced some of the first computer-generated images — including a grid of vertical lines perturbed by a Gaussian random function (1962) and a computer-generated version of Piet Mondrian's Composition with Lines (1965) that most viewers, in a small study Noll ran, preferred to the original. Frieder Nake and Georg Nees in Germany and Vera Molnar in France were producing plotter drawings in the same years, treating the plotter pen as a drawing instrument and the program as a score. Manfred Mohr's 1971 solo show Une Esthétique Programmée at the Musée d'Art Moderne de la Ville de Paris was the first museum exhibition devoted entirely to algorithmic work ^{[mohr1971 Mohr 1971]}. These artists were often scientists or engineers, and their work was exhibited at engineering venues before it was accepted by art galleries.

Sol LeWitt's wall drawings, beginning in 1968, gave the generative idea an art-world legitimacy that pure computer art initially lacked ^{[lewitt1971 LeWitt 1967]}. LeWitt's insistence that the concept, not the execution, constituted the work reframed authorship in a way that made algorithmic art intelligible within the mainstream contemporary tradition. When Casey Reas and Ben Fry released Processing in 2001 ^{[reas2007 Reas and Fry 2007]}, they lowered the technical barrier so that artists without engineering backgrounds could write generative code; p5.js later carried this into the browser. Reas's own practice argues that the algorithm itself is the aesthetic statement, a position Tyler Hobbs extends in long-form generative works like Autumn (2018) and Fidenza (2021).

Philosophically, the central question is where creativity resides when the artist does not touch the surface. Three positions recur. The instrumentalist view treats the algorithm as a sophisticated tool and credits the designer entirely. The system-as-author view, strengthened when algorithms produce outputs that surprise their creators, grants the system some co-authorial role. The curatorial view, ascendant in the diffusion-model era, relocates creativity to selection: the artist's act is to navigate and choose from a vast space of possibilities. These are not merely academic distinctions — they decided the 2023 US Copyright Office rulings on whether AI-assisted images can be registered, and they decide who owns a generative edition ^{[Manovich 2001]}.

Bibliography Master

@article{turing1952,
  author  = {Turing, A. M.},
  title   = {The Chemical Basis of Morphogenesis},
  journal = {Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences},
  volume  = {237},
  number  = {641},
  pages   = {37--72},
  year    = {1952}
}

@article{lindenmayer1968,
  author  = {Lindenmayer, Aristid},
  title   = {Mathematical Models for Cellular Interactions in Development {I--II}},
  journal = {Journal of Theoretical Biology},
  volume  = {18},
  number  = {3},
  pages   = {280--315},
  year    = {1968}
}

@article{perlin1985,
  author  = {Perlin, Ken},
  title   = {An Image Synthesizer},
  journal = {ACM SIGGRAPH Computer Graphics},
  volume  = {19},
  number  = {3},
  pages   = {287--296},
  year    = {1985}
}

@book{reas2007,
  author    = {Reas, Casey and Fry, Ben},
  title     = {Processing: A Programming Handbook for Visual Designers and Artists},
  publisher = {MIT Press},
  year      = {2007}
}

@article{lewitt1967,
  author  = {LeWitt, Sol},
  title   = {Paragraphs on Conceptual Art},
  journal = {Artforum},
  volume  = {5},
  number  = {10},
  pages   = {79--83},
  year    = {1967}
}

@article{noll1962,
  author  = {Noll, A. Michael},
  title   = {Computers and the Visual Arts},
  journal = {Bell Telephone Laboratories Technical Memorandum},
  year    = {1962}
}

@book{rush2005,
  author    = {Rush, Michael},
  title     = {New Media in Art},
  edition   = {2},
  publisher = {Thames and Hudson},
  year      = {2005}
}

@book{manovich2001,
  author    = {Manovich, Lev},
  title     = {The Language of New Media},
  publisher = {MIT Press},
  year      = {2001}
}