Revealing Palimpsests with Latent Diffusion Models: A Generative Approach to Image Inpainting and Handwriting Reconstruction

One of the significant challenges in ancient manuscript analysis is accurately reconstructing missing text. Palimpsests, a unique type of historical document, present distinct difficulties due to the overlap between overwritten and underlying text. In this work, we address this challenge by applying an effective image inpainting technique based on a generative model. Our method leverages a Latent Diffusion Model (LDM) backbone with key modifications to the conditioning mechanism, enabling the model to effectively utilize contextual information from the neighboring regions of the mask. To enhance the generation process, we provide an initial approximation of the masked region's pixels as a starting condition. Additionally, we incorporate intermediate representations within cold diffusion and employ a combined perceptual loss function. These advancements result in more refined and visually realistic handwriting reconstructions. Finally, we demonstrate the efficiency of our model through quantitative and qualitative evaluations on both synthetic and real palimpsest multispectral imaging (MSI) examples.