In this paper, we attempt to make ink talk in historical documents by extracting visual information materialized by the trace left on the support, thus addressing two tasks: anomaly localization in hand-pressed ornaments in the framework of Rey’s publishing and source separation in Queen Marie-Antoinette’s correspondence where the letters are hidden with loops. Anomalies are defined as changes in shape with respect to the original skeleton. A logarithmic image processing is applied for both anomaly visualization and image generation by mixing letters with loops to train a source separation model, respectively. Such an image generation is tackled from spectral modeling in halftoning.