What is it about?
Old textiles often survive only as small, faded, broken pieces. The Viking Age Oseberg ship burial in Norway (834 CE) left hundreds of such fragments, and fitting them back together by hand is slow, careful work. Computer programs that solve jigsaw puzzles from images already exist, but they were trained on everyday photos, not on woven fabric. We built TEXMET, a collection of 18,644 images of historical textiles from the Metropolitan Museum of Art's open collection, and used it to test three modern puzzle-solving AI models. We cut each textile image into square pieces, shuffled them, and measured how often each model put every piece back in the right place.
Featured Image
Why is it important?
This is the first image collection made specifically for teaching puzzle-solving AI about textiles. Our tests show that models trained only on everyday photos perform poorly on fabric, but an extra round of training on TEXMET changes that. The best model, DiffAssemble, solved 85.89% of textile puzzles after this training. On images of 45 heavily damaged Oseberg fragments, its success rate rose from 26.67% to 64.44%. The lesson is simple: AI tools for cultural heritage need training data that looks like the real material. The same approach could help museums and archaeologists working with other broken objects such as pottery, wall paintings, and manuscripts.
Perspectives
This paper grew out of my master's thesis. What struck me most was seeing a model that struggled with the damaged Oseberg fragments become far more reliable once it had learned from museum textiles. We are not yet reassembling real, irregular fragments, but I hope TEXMET gives other researchers a shared starting point for getting there.
Muhammad Hamza Zafar
Norwegian University of Science and Technology
Read the Original
This page is a summary of: TEXMET: A Deep Learning Benchmark for Archaeological Textile Reconstruction, Journal on Computing and Cultural Heritage, September 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3846195.
You can read the full text:
Contributors
The following have contributed to this page
