Collaborative Interactive Evolution of Art in the Latent Space of Deep Generative Models

TL;DR

The paper tackles the challenge of controlling GAN-generated art by navigating the latent space with evolutionary computing, using Creative Adversarial Networks (CANs) as the generator. A hybrid framework combines collaborative interactive evolution with an intelligent local-search mutation and automatic NIMA-based aesthetics to assess fitness, alongside diversity-preserving mechanisms. Empirical results show that collaborative human-guided evolution yields significantly more attractive images than random baselines, while automatic evolution alone does not consistently outperform randomness; local search provides mixed gains in human perception. The work demonstrates the value of human-in-the-loop evaluation for art synthesis and presents a scalable approach for co-creative image generation via a latent-genotype-to-phenotype mapping in GANs, with potential for broader exploration in AI-assisted artistry.

Abstract

Generative Adversarial Networks (GANs) have shown great success in generating high quality images and are thus used as one of the main approaches to generate art images. However, usually the image generation process involves sampling from the latent space of the learned art representations, allowing little control over the output. In this work, we first employ GANs that are trained to produce creative images using an architecture known as Creative Adversarial Networks (CANs), then, we employ an evolutionary approach to navigate within the latent space of the models to discover images. We use automatic aesthetic and collaborative interactive human evaluation metrics to assess the generated images. In the human interactive evaluation case, we propose a collaborative evaluation based on the assessments of several participants. Furthermore, we also experiment with an intelligent mutation operator that aims to improve the quality of the images through local search based on an aesthetic measure. We evaluate the effectiveness of this approach by comparing the results produced by the automatic and collaborative interactive evolution. The results show that the proposed approach can generate highly attractive art images when the evolution is guided by collaborative human feedback.

Figures (13)

• Figure 1: Block diagram of the evolutionary algorithm.
• Figure 2: Illustration of the preserve diversity metric. While the examples (a) and (b) would be judged as too similar, the examples (c) and (d) exhibit sufficient differences.
• Figure 3: (a) Fitness improvement of 20 random images over 100 generations of local search according to the automatic evaluation metric. (b) Proportion of participants preferring the local search results over the original image, averaged over all 20 images.
• Figure 4: Images (b) and (d) result from images (a) and (c) using the local search. (b) was preferred in 77% over (a), while (d) was preferred only in 10% over (c).
• Figure 5: Mean fitness through generations of the automatic evolution, with the shaded area representing the standard error. The results are averaged over five runs.