Discourse-driven Comic Generation

Narrative generation enables a range of opportunities for understanding the creative act of storytelling. Prior approaches have mostly converged on a pipeline model, wherein story structure is generated as a precursor to discourse structure, mapping individual story events to discourse elements. This model, however, unnecessarily limits narrative possibilities, which most prior work avoids by implicitly assuming a textual output medium. We investigate a new generation approach that treats discourse as primary, using comic generation as a testbed. Our approach is based on leading discourse theories for comics by McCloud (panel transitions) and Cohn (narrative grammar). Rather than rearranging preexisting panels, we generate panel contents based on notions of relatedness supported by cognitive theories of visual language. We present a proof-of-concept generator with a wide range of abstract comic output, a computational realization of McCloud’s and Cohn’s comics theories, and a modular algorithm that affords the evaluation of visual discourse theories.