Experimentation with Artbreeder

This experiment explores how human decisions may influence outcomes when using such a system to produce images. Using the image generation system Artbreeder (Simon 2018), one may only see one set of results at a time, meaning that one only has a partial idea of what might be produced, but this lacks breadth showing all possible outcomes that one might choose among. To get around this issue and to develop an idea of the diversity of images produced by the Artbreeder system, I gathered together all possible outcomes descending from an initial starting image over several generations. In so doing, I was able to visualise how an individual user’s choices might change the final outcome from repeatedly “breeding” images to produce new “offspring” images. The result of this process is something akin to a family-tree of images from a starting image to its many potential outcomes.

What I found was disappointing is that the Artbreeder system demonstrated a lack of diversity among the different lineages of images. While there was some variation among the different sets of images, the overall aesthetics of what was produced tended toward certain similar sets of traits. From several attempts at experimenting with this system, I found that it was possible to produce interesting outcomes by using Artbreeder as intended. While this does not offer the user a comprehensive overview, it does enable a more explorative approach than painstakingly considering all possible outcomes associated with a given starting combination.

The Artbreeder system has subsequently changed to enable users greater control on the genome used for producing images. One may combine, for example: african grey parrot; turnstile; church; plastic bag; dam; and chaos, to produce the image above. This affords users greater control to over the process than was possible previously, combining various kinds of images together.

Leave a Reply