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

TL;DR

The paper presents ArtNeuralConstellation, a framework to contrast AI-generated art with human art by combining Wölfflin principles, 15 general art concepts, OOD analysis in CLIP space, time mapping, and emotion/likability studies. Leveraging eight generative models (GAN-based and diffusion) trained on WikiArt and evaluated against 6,000 human artworks, the authors map artworks into neural representations and semantic spaces, uncovering a consistent bias of AI art toward modern (1850–2000) aesthetics and a tendency toward incomplete figuration in GAN outputs. The study shows AI art is generally ID with human art for landscapes and geometric abstracts but OOD for deformed figures, indicating distinct visual regimes in machine-generated work. Emotion and likability results reveal AI art can be highly engaging, with DDPM-based outputs often most likable and a diverse emotional palette, suggesting AI as a viable contemporary artistic medium with unique stylistic signatures. The work provides a scalable, data-driven protocol and an open data/code resource to guide future comparisons of human and AI-generated art across stylistic, temporal, and affective dimensions.

Abstract

Discovering the creative potentials of a random signal to various artistic expressions in aesthetic and conceptual richness is a ground for the recent success of generative machine learning as a way of art creation. To understand the new artistic medium better, we conduct a comprehensive analysis to position AI-generated art within the context of human art heritage. Our comparative analysis is based on an extensive dataset, dubbed ``ArtConstellation,'' consisting of annotations about art principles, likability, and emotions for 6,000 WikiArt and 3,200 AI-generated artworks. After training various state-of-the-art generative models, art samples are produced and compared with WikiArt data on the last hidden layer of a deep-CNN trained for style classification. We actively examined the various art principles to interpret the neural representations and used them to drive the comparative knowledge about human and AI-generated art. A key finding in the semantic analysis is that AI-generated artworks are visually related to the principle concepts for modern period art made in 1800-2000. In addition, through Out-Of-Distribution (OOD) and In-Distribution (ID) detection in CLIP space, we find that AI-generated artworks are ID to human art when they depict landscapes and geometric abstract figures, while detected as OOD when the machine art consists of deformed and twisted figures. We observe that machine-generated art is uniquely characterized by incomplete and reduced figuration. Lastly, we conducted a human survey about emotional experience. Color composition and familiar subjects are the key factors of likability and emotions in art appreciation. We propose our whole methodologies and collected dataset as our analytical framework to contrast human and AI-generated art, which we refer to as ``ArtNeuralConstellation''. Code is available at: https://github.com/faixan-khan/ArtNeuralConstellation

Figures (17)

• Figure 1: This figure shows how our study reveals the collective state of AI-generated art within the context of human art heritage based on individual analysis from the five aspects: Wölfflin's concepts, general art principles, Out-Of-Distribution analysis (OOD), time, and emotion & likability. The example bar plots and the answers in individual analysis present the actual results in our experiments for the AI portrait from the generative model CWAN2.
• Figure 2: The distribution example of the first principle: linearly vs. painterly. The bins represent the values from the left value to the right value. As we can see in this figure, AI-generated art is less uniform than human art for Wölfflin's principles. Machine artworks are highly populated on the right side concepts more in general. More distribution figures for other principles are shown in Appendix \ref{['expaned_woff_supp']}.
• Figure 3: For each visual concept, the densities for human (blue lines) and generated art (red lines) are estimated to contrast their centers. Dotted densities are also added to show how human artworks visually changed from classical to modern times and to see where the generated art falls on the historical transitions. The plots' ranges are magnified around centers, and the names of the concepts are colored red when the two population means are significantly different. We found AI-generated art centers are different from the human centers for most of the visual concepts except for "geometric", "open", and "linear"; in general, generated artworks are visually biased toward modern periods artworks.
• Figure 4: For the various CLIP spaces by different PCA dimensional reductions, OOD states are compared along with t-SNE representations (perplexity 30): human (greenish dots) and generated arts (orange dots). We found that for the top principal axes from 1 to 17 (50% data variance), OOD is relatively few, but the number of OOD is increased as more additional marginal dimensions are considered. This phenomenon is observed in both OOD counts and t-SNE plots.
• Figure 5: The most ten OOD (red lines) and eight ID samples (black lines) in 95% PCA are presented along with their t-SNE locations. All ID samples show landscape or abstract and geometric patterns, but OOD samples show irregular and twisted figures.