When Comics Chat: a deep prior for text-balloon reading order

The success of machine learning over the last 20 years, has been nothing less but a celebration of the power of big data. Comics are big data; they constitute one of the largest and most diverse visual–textual narrative corpora available today. Yet, either framed as a visual aesthetic, as ghiblification, or ossified as an outdated form of world-building, comics interaction with ML has been lukewarm. The resulting introversion of comics, can however serve as a research direction: What happens when comics can only speak with itself? Instead of training a machine learning model to speak like comics, something already demonstrated in projects like the Neural Yorker, this work carves an orthogonal path. Using a deep learning approach it takes a vast collection of comic text-balloons and learns a prior of their reading order. Using this prior it then predicts any plausible rearrange discussions that can arise from our database of comics using text-balloons that didn't appear together in the training set. Formatted into the most common contemporary discursive interface, comic balloons are then rendered into an endless chat, a conceptual device simulating the experience of comics talking to itself.