MultiMatte is a promptable image background removal model. Tell it what you want to keep, and MultiMatte will remove the rest.

How does language mediate visual meaning in AI image generation? This article theorizes AI image generation as mediated through three successive linguistic nets that together constitute a new epistemic order of the visual. Building on Vilém Flusser’s media philosophy, we assume that the translation of written language into images is forever incomplete: each act of translation opens a gap in meaning, with visual richness escaping linguistic capture, while simultaneously adding another layer of meaning embedded in language itself. The first net, dataset annotation, decodes images into linguistic categories determining which visual expression becomes computationally traceable. The second net, user prompting, is where visual intentions are encoded into technical idioms that circulate through communities of practice. The third net, system prompts, is where platform-embedded instructions invisibly modify user requests according to corporate and legal imperatives. Synthesizing recent research, we show how these nets interact: each successive translation reorients the space of visual possibilities, producing cumulative constraints on visual culture. The result of this configuration is a new epistemic order of AI image-making that determines which images can be generated and increasingly shape what users learn to imagine, what we call promptable imagination: the trained disposition to conceive visual ideas in terms that can successfully traverse all three nets.

Recent advances in visual generative models have enabled high-quality image and video generation, but evaluating these models often demands sampling hundreds or thousands of images or videos, which is computationally expensive. Existing evaluation methods also rely on rigid pipelines that overlook …