Liminal Scape is an expressive AI system that has been trained to recognize human emotion and generate abstract images at will. The proposed system receives an image (photographic portrait) and labels it based on the recognized emotional valence. Our system takes this initial photo and paints it red, yellow, or blue depending on the recognized emotions (from the facial expression) using a painterly algorithm which in turn becomes an input for a two modified Deep Convolutional Neural Network (CNN) models known as Deep Dream and Neural Style. These systems along with a final particle system pass generate a range of latent images that convey the initial emotion, unique to the given input (photographic portrait) and the labeled category (R, Y or B). Our system combines emergent and arbitrary behavior and breeding aspects of CNNs with a hybrid ML/particle stroking system to explore art creation within a high complexity space of artificial creativity.
left: overview of the Multi-Region Ensemble CNN (MRE- CNN) framework (adapted from Fan et al., 2018). right: the overview of arbitrary style transfer in real-time with Adaptive Instance Normalization (AdaIN) with a Deep CNN (VGG-19) (adapted from Elgammal et al., 2017; Huang & Belongie, 2017).
Paper: Salimi, M. et al., (2020). “Liminal Scape, an interactive visual installation with expressive AI” In Proceedings of International Symposium on Electronic Art 2020. ISEA.