Teaching AI to recognize ancient Roman painting styles from small broken fragments

What is it about?

Can a computer recognize an ancient painting style from just a small broken piece? We examine the unique case of the four classic Roman wall-painting styles, and show that modern AI can identify which style a realistic fragment belongs to, far more accurately than earlier methods.

Featured Image

Photo by Ahmet Yüksek ✪ on Unsplash

Photo by Ahmet Yüksek ✪ on Unsplash

Why is it important?

AI models that recognize artistic style have been studied extensively, but almost always on complete, well-framed images of artworks. We ask a harder question: how much can these models tell from just a small piece? To find out, we built a new dataset of realistic fragments by cutting images of Pompeian wall paintings using the exact shapes of real fragments excavated at Pompeii. We then compared a wide range of modern AI models, from huge foundation models to small models that can run on a phone. Our case study is the four Roman wall-painting styles that scholars have used to study and date Pompeii for over a century, a classification that is subtle even for experts when only part of a painting is visible. The best models identify a fragment’s style far more accurately than earlier methods, showing that a small piece holds more stylistic information than one might expect. This can help experts sort and catalogue fragment collections, and the approach could extend to mosaics, manuscripts and textiles.