The digitization of large scale cultural heritage collections has opened up new perspectives for the study of global culture. Recent works in digital humanities have started to rely on automated tools to uncover viral networks of reprinted textual works. The Numapresse project aims to extend this approach to visual content, using a pilot corpus of 100,000 illustrations extracted from magazines and dailies from 1850 to 1914. The identification of reprinted images is based on a « twisted » deep learning model. Instead of using the classification output, we focus on an intermediary step: the data generated by the third layer of the neural network which encodes hybrid information between the original visual rendering and the abstract categories used for classification. This experimental tool has not only allowed us to retrieve hundreds of reprints but also to derive structural information on the « flow » of reprint from one periodic to another. For instance, it turned out that a large share of reuses occurred in the Voleur illustré (the illustrated thief), which acted as a nineteenth-century aggregator of visual content.