My First Scientific Publication

My paper Demographic Influences on Contemporary Art with Unsupervised Style Embeddings, written with Noa Garcia and Yuta Nakashima, was accepted at the VISART workshop at the European Conference on Computer Vision 2020.

The question

Does where an artist comes from, or who taught them, leave a measurable trace in the visual style of their paintings? The idea had been discussed in art history for decades, but never tested at scale with computational methods.

The dataset

We built contempArt: 14,398 images of paintings and drawings by 441 early-career artists from 15 German art schools, collected between 2018 and 2020. The schools range from HGB Leipzig (67 artists, 2,464 images) to Weissensee in Berlin (8 artists, 144 images), covering a broad spectrum of programs and teaching philosophies across Germany.

For each artist we collected demographic metadata (school, gender, nationality, professor class) and, where available, Instagram data: follower counts, engagement metrics, and a directed social graph of 456,056 edges across 247,087 accounts.

The full dataset is available on Zenodo and documented in the contempart repository.

The method

We extracted unsupervised style embeddings using a VGG-19 convolutional neural network. Two types of features: FC7 activations (4,096 dimensions, capturing high-level content) and Gram texture descriptors (4,096 dimensions, capturing brushwork, color relationships, and surface patterns). We also computed node2vec embeddings from the Instagram follower graph to represent social proximity between artists.

With these representations in hand, we tested whether demographic variables (nationality, gender, birth decade, art school) could predict structure in the learned style space.

The finding

Short answer: barely. Nationality and birth decade showed small but consistent associations with style clusters. Gender and school showed weaker effects. The social network showed no meaningful correlation with style similarity. We concluded that artistic style, at least as captured by VGG-19 features, appeared largely independent of the demographic and social variables we measured.

The full analysis pipeline is reproducible end-to-end, with pre-computed embeddings available on Zenodo.

Osaka

This project would not have happened without my time at Osaka University in early 2020. I owe a great deal to Noa Garcia and Yuta Nakashima, who hosted me at the Intelligence and Sensing Lab and the Institute for Datability Science.

More information on this project is available on the project website.

2026 update: content vs. appearance

Six years later, I revisited the dataset with modern embeddings following the methodology of Kim et al. (2025). Instead of VGG-19, I used two models that separate what a painting depicts from how it looks:

• C-vectors (CLIP ViT-L/14, 768 dimensions): semantic content, what is painted
• A-vectors (Stable Diffusion 2.0 VAE, 16,384 dimensions): visual appearance, colors, composition, texture

The results reframe the original paper’s conclusion. For appearance, the null result holds: A-vectors, like VGG, show no significant link to school or professor class. But for content, the picture is different.

Professor class explains 10.5% of content variance (p=0.0001). School explains just 0.4% once you control for professor. In other words, the school effect is almost entirely a professor effect: students paint similar subjects because the same person teaches them, not because of any broader institutional culture.

Artists who follow each other on Instagram also produce more semantically similar work (r=0.111, p=0.009), but only in content, not appearance. And an unexpected asymmetry: gender predicts appearance slightly (2.0% variance) but not content.

The original paper asked the right question but had the wrong lens. VGG-19 mixed content and style together. Once you pull them apart, the demographic signal shows up clearly in what artists choose to paint, while how they paint remains their own.

The full re-analysis is in the contempart-clip repository.