← Bonabeau Apps Karl Sims: Genetic Images ↗

Interactive evolution of images

A browser replication of Karl Sims, “Artificial Evolution for Computer Graphics” (1991) and “Interactive Evolution of Equations for Procedural Models” (1993). Each picture is a Lisp expression evaluated at every pixel. You do the selecting.

Generation 1.

Select and reproduce

Click one image to breed mutations from it, or two images to mate them.

Expected mutations per offspring: 1.5

Genotype

Select an image to read its expression.

Library

Saved genotypes appear here. They can be selected as parents like any other image.

Load expressions

Paste one Lisp expression per line, in the format shown under Genotype.

What is replicated, and what differs from the papers

Genotypes are Lisp expression trees built from the 1993 texture function set: + - * / mod round min max abs expt log and or xor sin cos atan if dissolve hsv-to-rgb vector transform-vector bw-noise color-noise warped-bw-noise warped-color-noise blur band-pass grad-mag grad-dir bump ifs warped-ifs kaleidoscope warp-abs warp-rel warp-by-grad.
Mutation uses the seven types in section 2.1 (1993): new random subexpression, scalar adjustment, vector adjustment, function change, wrapping in a new function, an argument replacing its function, and copying another node. The per-node rate is scaled inversely with expression size, and shrinking is weighted slightly above growing.
Mating uses both methods from section 4.3 (1991): matched traversal that copies either parent’s version at each difference, and crossing over of random nodes.
Offspring with a high estimated compute cost are discarded before display, as in the papers. This page also discards images that are nearly a single flat color, which the papers do not do.
The cross dissolve follows section 2.3 (1993): identical nodes are shared and differing branches are interpolated.
The papers do not specify the internals of the functions, so the noise, warp, gradient, bump, and IFS implementations here are this page’s own. Images will not match the published figures.
Values are signed: X and Y run from −1 to 1, and −1 displays as black, +1 as white. Functions that need a single constant (blur radius, kaleidoscope segments, IFS coefficients) use the mean of their argument.
Evaluation runs per pixel in JavaScript on one thread in place of a CM-2. Volume textures, the Time variable, input images, 3D shapes, and dynamical systems are not included.