Crafting Generative Art through Genetic Improvement: Managing Creative Outputs in Diverse Fitness Landscapes

TL;DR

This paper tackles open-ended generative art creation via genetic improvement under diverse fitness landscapes, aiming to disentangle the contributions of individual fitness functions and align outputs with human-art preferences. The authors employ Grammatical Evolution and Genetic Improvement to evolve parameterized sequences of drawing techniques, guided by Lexicase selection and a CNN-based art classifier acting as a proxy for artistic quality. Their experiments reveal that, with few fitness functions, specific techniques tend to sweep the population, while adding objectives can induce niche diversification; the art classifier helps filter out noisy or blurry results but remains imperfect as a standalone measure. The findings highlight practical pathways for controlling creative outputs in evolutionary art and point to future work in alternative selection mechanisms and more robust perceptual evaluators.

Abstract

Generative art is a rules-driven approach to creating artistic outputs in various mediums. For example, a fluid simulation can govern the flow of colored pixels across a digital display or a rectangle placement algorithm can yield a Mondrian-style painting. Previously, we investigated how genetic improvement, a sub-field of genetic programming, can automatically create and optimize generative art drawing programs. One challenge of applying genetic improvement to generative art is defining fitness functions and their interaction in a many-objective evolutionary algorithm such as Lexicase selection. Here, we assess the impact of each fitness function in terms of the their individual effects on generated images, characteristics of generated programs, and impact of bloat on this specific domain. Furthermore, we have added an additional fitness function that uses a classifier for mimicking a human's assessment as to whether an output is considered as "art." This classifier is trained on a dataset of input images resembling the glitch art aesthetic that we aim to create. Our experimental results show that with few fitness functions, individual generative techniques sweep across populations. Moreover, we found that compositions tended to be driven by one technique with our current fitness functions. Lastly, we show that our classifier is best suited for filtering out noisy images, ideally leading towards more outputs relevant to user preference.

Figures (6)

• Figure 1.1: Sample generative art outputs.
• Figure 1.2: CNN visualizations.
• Figure 1.3: Total time required for each individual technique across 100 randomly-instantiated invocations.
• Figure 1.4: Sample collages of final outputs from $EC_{ut,ac}$ and $EC_{pc,gc,ut,cd}$ from one replicate. Each collage comprises the outputs from the final evolved population to demonstrate the difference between the different objectives used for selection.
• Figure 1.5: Heatmap of normalized and averaged fitness values for final population per experimental configuration (three minute evaluation timeout, flow-field removed). Experimental configuration numbers correlate to a truth table of activated fitness objectives.