An investigation into AI-generated art through GANs and ML neural network

Abstract

Art has always been the primary means of human expression, aided by various tools humans were equipped with. Amongst recent tools and advances in artistic technologies there are generative adversarial networks (GANs) and various neural networks, performing the creative task of art generation through the application of different techniques. The scope of this paper is to investigate and compare the use of GANs and machine learning neural techniques to replicate state‐of‐the‐art applications whose code is not publicly available. Furthermore, it aims to evaluate the implementation against the state‐of‐the‐art applications. The VQGAN+CLIP architecture with an additional neural style transfer (NST) layer was tested and compared to other AI-generated art applications. The results showed that while the VQGAN+CLIP images were creative and unique, they lacked structure, had high blurriness, and lower overall quality, which gives a direction for future work. The application called Midjourney outperformed others in most categories.