Restoring Historical Paintings Using Diffusion Models and GANs
Abstract
Restoring historical paintings is crucial for preserving cultural heritage, but traditional methods often struggle with severe degradation, color fading, and missing details. Recent advancements in deep learning, particularly Generative Adversarial Networks (GANs) and diffusion models have provided innovative solutions for digital restoration. Existing restoration techniques, such as manual retouching and classical image inpainting, require expert knowledge and often fail to reconstruct fine details lost over time. Furthermore, conventional deep-learning methods may introduce artifacts or fail to generalize across diverse artistic styles. We propose a restoration framework combining GANs with diffusion models to address these challenges. GANs, with their adversarial training mechanism, enhance texture and structure recovery, while diffusion models improve noise reduction and fine-grain detail reconstruction. The approach refines missing regions, corrects color degradation, and maintains stylistic consistency through adversarial learning and probabilistic diffusion-based denoising. The proposed method enables museums and researchers to restore paintings more accurately and efficiently, reducing reliance on manual interventions while ensuring historically faithful reconstructions. Experimental results demonstrate that our approach outperforms conventional methods regarding perceptual quality, structural similarity, and artistic fidelity. Integrating GANs and diffusion models offers a powerful tool for cultural heritage preservation.