Worlds from Images: Generative World Models and the Becoming-Operative of Visual Culture

Generative world models promise to turn images into navigable and interactive environments, while video models are promoted as simulators of the physical world. This article situates these claims within the histories of machine vision and computer graphics. Where inverse optics and inverse graphics in machine vision aimed at inferring world structure from images, early computer graphics constructed computational worlds to render images from them. With generative AI, these trajectories are folded together, learning regularities from worlds that have already become images by projecting them forward. In turn, what these systems learn are the specific conventions, genres and affordances of different image ecologies. They thereby complicate the notion of the operative image: it is the metalanguage and excess visuality of images made for the human eye that becomes operative.