ReCoDi - Refinable Colorization of Historic Imagery via Conditional Diffusion Models

Film archives and museums all over the world contain large amounts of historic videos and imagery, only available as either monochrome or even black-and-white documents. While digitization and in some cases digital restoration preserves these cultural assets from decay, marketing of these historic documents is difficult due to limited public acceptance caused by their archaic look. Colorization techniques make the depictions of historic events visually appealing such that they are interesting to a larger audience. However, colorizing historic imagery remains a highly labor- and cost-intensive process, which prevents its more widespread use to this day. ReCoDi addresses the most costly step in the colorization process of historic films, namely the colorization of single keyframes respecting authentic historic looks. Such keyframes are the basis for any further color propagation tools. This project aims to combine state-of-the-art deep learning-based image generation approaches, novel conditioning strategies for these generative models, and efficient user control into a unified optimization framework.