Combination of certainty and uncertainty: Using FusionGAN to create abstract paintings

In the study of generative art, it is relatively easy at present to achieve a high degree of certainty or uncertainty. However, the combination of certainty and uncertainty has always been an area of difficulty in generative art. In this paper, we present a novel FusionGAN system to automate the generation of abstract paintings. These generated abstract paintings combine the factors of certainty and uncertainty. First, we collect an APdataset consisting of three parts: abstract paintings drawn by artists, sketches, and abstract paintings generated by other neural network methods. We then train the proposed FusionGAN system on the collected dataset to learn the expression of abstract paintings. Corresponding to the two-step operation of the combination of certainty and uncertainty in the artist's creation, the proposed FusionGAN system is also divided into two steps for the generation of abstract paintings. More specifically, the first step is the basic structure establishment, which corresponds to the fundamental certainty element in the painting creation. The second step is the realization of details, which integrates the uncertain details based on the basic structure. The experimental results achieved by our system in abstract painting generation enrich the diversity of artistic creation and have been recognized by art institutions, with some results displayed on their websites.