STERR-GAN: Spatiotemporal Rerendering for Facial Video Restoration

Old facial films are valuable historical resources that vividly depict significant figures of the past. However, these films suffer from low quality and visual distortion due to outdated camera technology and poor preservation. Facial video restoration offers a solution, incorporating sub-tasks such as noise removal, colorization, and facial enhancement. A naive approach that restores each frame independently often results in inconsistency. In this paper, we introduce Spatio-temporal Re-rendering for Facial Video Restoration (STERR-GAN), which features a Temporal Recurrent Network to leverage adjacent frame information for maintaining temporal coherence, and a specialized GAN module to enhance facial details. Our experimental results demonstrate that STERR-GAN significantly outperforms existing methods in both quantitative and qualitative analyses. Additionally, to address the insufficiency of current datasets, we propose the Facial Video dataset for Antique Restoration (Face-VAR) to support further research in facial video restoration.