AI Art Neural Constellation: Revealing the Collective and Contrastive State of AI-Generated and Human Art

Discovering the creative potentials of a random signal to various artisticexpressions in aesthetic and conceptual richness is a ground for the recentsuccess of generative machine learning as a way of art creation. To understandthe new artistic medium better, we conduct a comprehensive analysis to positionAI-generated art within the context of human art heritage. Our comparativeanalysis is based on an extensive dataset, dubbed “ArtConstellation,”consisting of annotations about art principles, likability, and emotions for6,000 WikiArt and 3,200 AI-generated artworks. After training variousstate-of-the-art generative models, art samples are produced and compared withWikiArt data on the last hidden layer of a deep-CNN trained for styleclassification. We actively examined the various art principles to interpretthe neural representations and used them to drive the comparative knowledgeabout human and AI-generated art. A key finding in the semantic analysis isthat AI-generated artworks are visually related to the principle concepts formodern period art made in 1800-2000. In addition, through Out-Of-Distribution(OOD) and In-Distribution (ID) detection in CLIP space, we find thatAI-generated artworks are ID to human art when they depict landscapes andgeometric abstract figures, while detected as OOD when the machine art consistsof deformed and twisted figures. We observe that machine-generated art isuniquely characterized by incomplete and reduced figuration. Lastly, weconducted a human survey about emotional experience. Color composition andfamiliar subjects are the key factors of likability and emotions in artappreciation. We propose our whole methodologies and collected dataset as ouranalytical framework to contrast human and AI-generated art, which we refer toas “ArtNeuralConstellation”. Code is available at:https://github.com/faixan-khan/ArtNeuralConstellation