An Illusive Image Generation Methodology for Artists in Latent Space: Experiments in VAE-based Generative Deep Model

Abstract The rapid advancements in deep learning-based image generation tools have led to hyper-realistic results, highly favored in various industries. However, for fine arts practice, achieving the intended visual scope is crucial. Variational Autoencoders (VAEs) have gained attention for their ability to produce illusive outcomes with blurry effects. In this research, the latent space of VAEs and their unique visual characteristics are investigated. Through visual analysis, it is found that each VAE model exhibits distinct visual traits, offering diversity while remaining consistent with the original model. This has significant potential to enhance visual art practice. A workflow is proposed that showcases the use of VAEs in fine art, particularly with a single dataset. This workflow guides artists in leveraging VAEs’ distinctive visual characteristics and exploring their application for artistic outcomes. By incorporating VAEs into the creative process, artists can benefit from these models’ unique capabilities.