Shades of History: Reviving Nepal’s Heritage through AI

Abstract

Historical photographs provide vital insights into Nepal’s rich cultural heritage; however, most existing archival collections remain in black and white, limiting visual engagement and cultural comprehension for modern audiences. Despite advancements in AI-driven image colorization, current methods often suffer from inaccuracies in historical and cultural authenticity, highlighting a crucial research gap. This study addresses challenge by employing Conditional Generative Adversarial Networks (cGANs), leveraging a U-Net architecture with a pre-trained ResNet18 backbone. Initially trained using supervised L1 loss and subsequently refined through adversarial training, our method significantly enhances the visual authenticity and accuracy of colorized images. Quantitative assessments yielded a discriminator loss of 0.62 and generator loss of 4.42 for our best model with pretrained backbone. The resulting high-quality colorizations vividly depict historical narratives, greatly enriching the preservation and appreciation of Nepal’s cultural heritage.