A Comparative Analysis of Synthetically Generated Damage Types on Photographs

The restoration of old, damaged photographs is an important cultural and historical task. The rapid development of artificial intelligence methods, especially deep learning, is leading to significant advances in inpainting and restoration. In this work, we address this challenge by generating synthetic damage on historical photographs, focussing on the creation and comparison of different types of visible deterioration, such as scratches and cracks. Our method applies realistic damage masks to photographs by using alpha blending techniques utilising text-to-image diffusion models. This technique allows real-world damage patterns to be simulated in a controlled and scalable manner, promoting reproducible research and the development of reliable restoration algorithms. Experts assess the generated synthetic datasets to ensure their quality and suitability for the preservation of digital cultural heritage.