Core Concepts
Collaborative interactive evolution can generate highly attractive art images by leveraging human feedback to guide the exploration of the latent space of deep generative models.
Abstract
The paper explores a framework for collaborative interactive evolution of art images using Generative Adversarial Networks (GANs). The key highlights are:
The authors employ a specific GAN architecture called Creative Adversarial Networks (CANs) to generate novel art images that mimic the real art distribution but deviate from established styles.
They use evolutionary computing to navigate the latent space of the CAN generator, with two fitness evaluation approaches:
Automatic aesthetic evaluation using a deep learning-based metric (NIMA)
Collaborative interactive evaluation where multiple participants rate the generated images
The evolutionary algorithm includes crossover, mutation, and a local search-based mutation operator to improve image quality. Diversity preservation mechanisms are also introduced to avoid user fatigue.
The results show that the collaborative interactive evolution approach can generate highly attractive art images, as perceived by human participants, while the automatic aesthetic evolution did not achieve the same level of success.
The local search-based mutation did not lead to significant improvements beyond random level, highlighting the challenges in automatically assessing the aesthetic quality of art.
The collaborative approach reveals the subjectivity in perceiving art, with diverse ratings from participants, and also demonstrates the importance of human guidance in the evolution of creative art.
Quotes
"Collaborative interactive evolution can generate highly attractive art images when the evolution is guided by collaborative human feedback."
"The collaborative approach reveals the subjectivity in perceiving art, with diverse ratings from participants, and also demonstrates the importance of human guidance in the evolution of creative art."