What is it about?
Many historic monuments and archaeological sites have been damaged or destroyed by war, natural disasters, and the passage of time. This study explores whether modern artificial intelligence can help reconstruct lost cultural heritage when only limited visual information remains. Building on a reconstruction framework previously developed for the Bamiyan Buddhas, the research tests whether the same approach can be applied to a very different site, the Temple of Bel in Palmyra. The study combines historical images, computer vision, and generative AI techniques to create digital reconstructions and evaluate their accuracy. By examining how well the method generalises across locations, the research helps identify both the opportunities and limitations of AI-assisted heritage reconstruction. The findings aim to support archaeologists, historians, museums, and heritage organisations in preserving and visualising endangered cultural landmarks, while promoting transparent and responsible use of AI in cultural heritage conservation. Ultimately, the work investigates how AI can contribute to safeguarding humanity's shared history for future generations.
Featured Image
Photo by Ashna Jain on Unsplash
Why is it important?
A distinctive feature of this study is that it goes beyond demonstrating an AI-based reconstruction method on a single cultural heritage site. Instead, it tests whether the same reconstruction framework can successfully transfer from the Bamiyan Buddhas, where it was originally developed, to the very different archaeological context of the Temple of Bel in Palmyra. This cross-site evaluation addresses an important scientific question: whether generative AI reconstruction methods are genuinely generalisable or are only effective for the specific monuments on which they were developed. By introducing a bounded and transparent assessment of performance across different heritage settings, the work provides evidence about the reliability, limitations, and practical applicability of AI-assisted heritage reconstruction. The study therefore contributes not only a reconstruction outcome, but also a methodology for evaluating the robustness and transferability of generative AI tools for cultural heritage preservation.
Perspectives
In my view, this research is about testing whether artificial intelligence can become a reliable tool for reconstructing damaged or destroyed cultural heritage sites, not just in one specific case but across different historical contexts. Rather than focusing solely on generating visually convincing reconstructions, the study asks a more important scientific question: can a reconstruction approach developed for one monument also work elsewhere? The paper examines this by applying the HARF methodology beyond the Bamiyan Buddhas to the Temple of Bel in Palmyra, providing a structured assessment of how well the method transfers between sites. I believe the work is important because cultural heritage preservation increasingly faces challenges from conflict, natural disasters, urban development, and environmental change. While generative AI offers powerful reconstruction capabilities, there is a risk that digitally reconstructed heritage may be visually appealing but historically unreliable. By focusing on generalisation, evaluation, and methodological transparency, this research helps move the field beyond demonstrations towards evidence-based practice. Its contribution is not only the reconstruction of a specific monument but also a framework for assessing whether AI-generated heritage reconstructions can be trusted, reproduced, and responsibly used by archaeologists, historians, museums, and conservation organisations. Such evidence is essential if AI is to become a credible tool for preserving humanity's shared cultural legacy.
Prof Tatiana Kalganova
Brunel University
Read the Original
This page is a summary of: Generalising HARF Beyond Bamiyan: A Bounded Cross-Site Evaluation of Generative-AI Reconstrucction at the Temple of Bel, Palmyra, July 2026, MDPI AG,
DOI: 10.20944/preprints202607.2162.v1.
You can read the full text:
Contributors
The following have contributed to this page
