One million comic books panel

Comic books make use of white space — or gutters — to propel the story forward, relying on readers’ intuitive ability to fill in the gaps between panels. To see whether computers could learn to make the same inferences, a group of computer scientists built a giant corpus of public-domain comics and tried training a series of neural networks on it. (Spoiler: Humans are much better at this.) The underlying dataset contains 1.2 million panels from nearly 200,000 scanned pages of nearly 4,000 books in the Digital Comic Museum, all published during the 1938–1954 “Golden Age” of American comics. It also contains 2.5 million chunks of text extracted from the comics’ speech balloons, thought bubbles, and narration boxes. [h/t Robin Sloan]

Sign In


Reset Password

Please enter your username or email address, you will receive a link to create a new password via email.