The Compositional Rhetoric of Multimodal Generative AI

In this article, I address the enunciative rhetoric of AI-generated images, focusing, in particular, on two questions. The first concerns the composition rules learned by generative models in representing the human face through prompts asking for the negation, subtraction, addition, or substitution of some fundamental parts of the face such as the nose. These tests have the objective of studying the (supposed) path the machine realizes during the process of image instauration and of describing the image’s compositional rhetoric in relation to the difficulty of these models in figuring uncommon figures such as noseless faces. Indeed, a noseless face is something unexpected and rarely found in the datasets which have trained these kinds of models. The second part of the article uses another corpus as a demonstration of the model’s work with prototypic, unique, or unusual topics. It explores how GenAIs, solicited through prompts mentioning recognized artistic painterly styles (e.g., Van Gogh’s and Klee’s styles) bring forth stereotypical forms and textures. In this context, by starting from the familiar forms of historical styles in the artistic tradition, it becomes possible to estimate both the level of stereotyping and the level of creativity of these models.