ArtGAN: A Self-Supervised GAN with Diffusion Enhancement for Artwork Restoration

Digital restoration of artworks is a demanding and tough task due to deterioration, where it becomes even more difficult when through-damage examples are not available and preserving the style is very important for the cultural heritage. The authors describe ArtGAN, a hybrid system that handles each input between a self-supervised (SimCLR) pretrained Pix2Pix restoration branch for structure-faithful reconstruction and a Stable Diffusion inpainting branch for semantically plausible completion. Then, the outputs are fused with edge-aware alpha blending to minimize seams and color drift. A decision-theoretic controller is the core innovation that decides to which branch and what blending coefficient to apply using estimated damage severity, edge density, and semantic deviation (e.g., perceptual distance), thus relying less on paired labels while maintaining iconography, brushstroke continuity, and palette characteristics among various traditions. ArtGAN shows improvement over strong GAN, LaMa, and diffusion baselines and achieves higher human-preference scores in a conservator-guided evaluation on curated crack, fade, and missing-region benchmarks assembled from public art corpora through the means of PSNR/SSIM and lower LPIPS/FID. Ablations isolate the impact of self-supervised pretraining, the routing policy, and blending coefficient selection, demonstrating consistent gains and robustness under prompt and mask variations. A lightweight, two-column UI supports practitioner workflows (mode selection, prompts, alpha slider, side-by-side preview) with exportable layered outputs to facilitate iterative conservation review. Code, train/validation/test splits, prompts, controller thresholds, and model checkpoints will be released to enable reproducibility and further research on cultureaware restoration.