Hybrid Approaches of Evolutionary and Neural Computing in the Field of Art

Abstract

This Master’s thesis explores the hybridization of Evolutionary- and Neural Computing
for advancing the field of digital art and image generation, employing state-of-the-art
Generative Networks, notably Diffusion Models. The research introduces a versatile framework specifically designed to support various Generative Networks and Fitness
Evaluators. It distinguishes itself from traditional Evolutionary Computing approaches
by effectively decoupling the input arguments from the outputs, placing a Generative Network at the core of the process.
A key aspect of this study is the practical implementation of the framework using
Stable Diffusion models. Unlike conventional methods that focus on promt strings, this research emphasizes optimization within the prompt embedding space, introducing
a novel approach. Evaluation is conducted using multiple fitness criteria, including
the LAION-Aesthetics Predictor V2, CLIPScore, CLIP-IQA, and AI-Detection. The
objective of this study goes beyond simple optimization to include the exploration
of diverse image styles and artistic expressions across both single and multi-objective
optimization scenarios. This exploration is further enriched by employing the Island
Model, a concept borrowed from Evolutionary Computing, to facilitate sophisticated
evolution of image styles.
The results of this research are compelling, demonstrating that optimization with
Evolutionary Algorithms in the prompt embedding space effectively enhances aesthetic quality according to the Predictor V2, achieving both higher and lower aesthetics scores
compared to real images from a LAION sub-dataset. The studies also successfully
showcases style transitions using CLIPScore and optimizes specific image qualities using
CLIP-IQA, handling up to nine criteria simultaneously. Remarkably, the research also
indicates that it is possible to significantly reduce detectability by an AI detection
system, from nearly 100% to almost zero.
This thesis contributes to the field of hybrid Evolutionary and Neural Computing,
particularly within computational artistry. It illustrates how advanced machine learning
techniques can be adapted for artistic creation and image optimization. Laying a solid
foundation for future interdisciplinary research, this work promises exciting new developments in the field of evolutionary generative images and art.

Date of Award	2024
Original language	English (American)
Supervisor	Michael Affenzeller (Supervisor)