Högre seminarium: Laura Wagner (Universität Zürich) - Democratizing Style: Authorship and Consent in Text-to-Image AI Personalizations

Text-to-image systems make it easy to generate images “in the style of” a given artist or genre. This notion of style is reductive: it treats style as a set of transferable correlations, whereas in art history it has been understood as a contested marker of context, authorship, and meaning. Examining thousands of community-trained models tagged as style reveals how the term circulates less as an artistic category and more as a label for exchange and categorization. Controlled experiments with hand-drawn examples further probe what these systems actually encode when trained on a distinctive drawing approach.

Taken together, these perspectives raise questions about the social consequences of treating style as a modular token. While foundation models already distill style into a promptable unit, community-trained personalizations such as LoRA create highly specific, distributable modules that commodify the styles of individual, mostly digital artists. Framed as the democratization of artistic expression, style tokenization risks weakening the role of style as a personal signature and unsettling established notions of authorship, consent, and ownership.

More about the topic in the paper Perpetuating Misogyny with Generative AI: How Model Personalization Normalizes Gendered Harm.