Dr. Stefanie Schneider

Stefanie Schneider is an Assistant Professor at the Institute of Art History at LMU Munich. With an academic background in Statistics, Computer Science, and Sociology, she earned her PhD in Art History in 2024. Since 2016, she has been active in the field of Digital Humanities, exploring the intersection of traditional hermeneutic approaches and contemporary quantitative methods in art-historical enquiry. Her first book, Art and the Artificial Eye: Algorithmic Views of the Human Posture (2025), investigated the utilization of computational methodologies to construct a virtual space of associations, references, and similarities for the historical embedding of the human figure, and its posture, in visual art. Her work has received several awards, including an honorary mention in the 2023 European Union Prize for Citizen Science, and has been featured in leading academic venues such as IEEE, ACM, and ADHO. She is a member of the Association for Digital Humanities in the German-speaking Countries (DHd) and the Working Group for Digital Art History.

Project at CAS:

At CAS, Stefanie Schneider is exploring the challenges that Vision-Language Models (VLMs) encounter when interpreting non-real-world imagery. Since the mid-2010s, VLMs – especially those based on transformer architectures – have made remarkable progress, enabling systems to “attend” to multiple aspects of input data simultaneously. However, these advances often falter when applied to non-real-world images: visual materials shaped by artistic, cultural, or conceptual intervention (such as paintings, film stills, or archaeological drawings). Such types of images are underrepresented in training datasets, resulting in shallow or idiosyncratic representations that fail to capture the inherently hierarchical meanings of these images. Consequently, VLM-generated representations often detach from the images’ original contexts, raising concerns about their interpretive coherence. This project investigates how VLMs reconfigure non‑real‑world images. It pursues two goals: first, to highlight the representational shortcomings of VLMs by identifying where and why they fail to encode the semantics of non‑real‑world imagery; and second, to interpret VLMs as ‘mediators of meaning’ by exploring how their outputs navigate – and sometimes distort – the space between aesthetic intention and computational abstraction.