Digital Inpainting of Damaged Frescoes Using a Fine-Tuned Diffusion Model
Loading...
Date
2025
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
The Eurographics Association
Abstract
Frescoes are a vital part of cultural heritage, but they are increasingly deteriorating due to environmental and human factors. Traditional restoration methods are costly, labor-intensive, and may risk compromising the original artwork-particularly when the damaged content is uncertain. These challenges highlight the need for innovative approaches that can complement conventional techniques. This paper explores the application of a deep learning-based method for the virtual restoration of frescoes. The focus is on image inpainting, a process that fills in missing fragments by leveraging information from the undamaged parts of the image, while preserving consistency in texture, color, and artistic style. When enhanced by deep learning models, this approach enables the generation of realistic reconstructions, even for areas where the original appearance is unknown. The study evaluates the effectiveness of fine-tuned models in restoring both minor and major damage, such as small cracks and missing sections, using different sets of hyperparameters. Model performance was assessed using a combination of objective quality metrics and subjective evaluations. Additionally, an intuitive web-based tool was developed to make the restoration process more accessible and user-friendly.
Description
CCS Concepts: Computing methodologies → Artificial intelligence; Computer vision; Reconstruction
@inproceedings{10.2312:dh.20253072,
booktitle = {Digital Heritage},
editor = {Campana, Stefano and Ferdani, Daniele and Graf, Holger and Guidi, Gabriele and Hegarty, Zackary and Pescarin, Sofia and Remondino, Fabio},
title = {{Digital Inpainting of Damaged Frescoes Using a Fine-Tuned Diffusion Model}},
author = {Smolka, Milena and Smolka, Bogdan},
year = {2025},
publisher = {The Eurographics Association},
ISBN = {978-3-03868-277-6},
DOI = {10.2312/dh.20253072}
}