Creative art generation with a style-based architecture for generative adversarial networks
Combining the state of the art Generative Adversarial Network (GAN) architecture of StyleGAN2 with the artistic style ambiguity loss of Creative Adversarial Network (CAN) has promises for improving the creative art eld because styles are more controlled and higher resolutions can be reached than before. Even though my implementation does not achieve the expected results but su ers from common GAN problems, contributions are made towards the development and evaluation of such improved creative networks, and its societal relevance is discussed. A questionnaire for human evaluation is developed as I argue that objective GAN metrics give ambiguous results in the creative eld. Furthermore, this area gives contributions towards Arti cial General Intelligence, but also additional developments towards identifying deep fakes are required. Overall, my research together with follow up implementation improvements help understanding the possibilities and complexities of this intriguing eld of creative computer-generated art.
Faculteit der Sociale Wetenschappen