Philipp Flieger¹, Daniel Kaiser^1,2,3,4; ¹Neural Computation Group, Department of Mathematics and Computer Science, Physics, Geography, Justus Liebig University Giessen, Germany, ²Cluster of Excellence “The Adaptive Mind”, Universities of Giessen, Marburg, and Darmstadt, Germany, ³Center for Mind, Brain and Behavior (CMBB), Universities of Giessen, Marburg, and Darmstadt, Germany, ⁴Center for Applied Computer Science and Data Science (ZAD), Justus Liebig University Giessen, Germany

Looking at a photograph or van Gogh’s rendition of a night sky, we have no problem telling apart style from content. Yet, how the brain represents style and content has not been investigated. Here, we created stylized renditions of 20 natural objects and scenes in seven different artistic styles (3D-render, cubism, expressionism, drawing, pixel art, renaissance, watercolor) using Stable Diffusion XL. In an EEG experiment, participants (N = 25) viewed the stylized images, and the original photographs used to create them, while performing a simple catch-trial detection task. Representational similarity analysis revealed robust neural representations of both art style and content. As expected, we observed higher pairwise decoding accuracies between visually distinct art styles (e.g., 3D-render vs. expressionism), compared to more similar ones. Moreover, neural representations of art style peaked earlier (~100ms post-stimulus) than representations of content (~150–200ms post-stimulus), suggesting that rapidly emerging perceptual codes differentiate art styles. When regressing art-style and content predictors, respectively, onto the neural data while accounting for hierarchical deep neural network (DNN) features and diverse image statistics (e.g., saturation, symmetry) we found that color features were robustly related to style coding while DNN features better explained content coding. Accounting for early- to mid-level DNN features abolished early coding of content while leaving style representations at similar time points intact, and removing the variance explained by high-level features abolished early coding of both art style and content. Conversely, controlling for the variance explained by various simple image statistics left content representations largely unaffected. In summary, our results suggest that neural representations of art style and content emerge with distinct temporal dynamics and, crucially, differ in terms of format.