Evolving Pictures In Image Transition Space

Evolutionary art creates novel images through a processes inspired by natural selection. Images are high dimensional objects, which can present challenges for evolutionary processes. Work to date has handled this problem by evolving compressed or encoded forms of images or by starting with prior images and evolving constrained variations of these. In this work we extend the prior-image concept by evolving interesting images in the transition-space between two bounding images. We define new feature metrics based on proximity to the two bounding images and show how these metrics, combined with other aesthetic features, can be used to drive the creation of new images incorporating features of both starting images. We extend this work further to evolve sets images that are diverse in one and two feature dimensions. Finally, we accelerate this evolutionary process using an autoencoder to capture the transition space and reduce the dimensionality of the search space.