18 Generative Visual Languages

ARTISTIC: Acquiring Representational Theories through
Infinitesimal Semantic Transport in Image Creation

Text-to-image diffusion makes the tension at the center of this book visible. A model generates images never observed during training by following many small denoising steps. The result may be novel and aesthetically valuable, yet the language controlling that process remains largely inherited: a fixed tokenizer, text encoder, conditioning interface, and repertoire of prompt constructions. Generative novelty is therefore not automatically representational creativity.

This chapter introduces ARTISTIC: Acquiring Representational Theories through Infinitesimal Semantic Transport in Image Creation. ARTISTIC specializes the AGENTIC contract to visual languages. It couples the image–noise geometry of diffusion to a second direction in which the conditioning theory itself can be diagnosed, extended, and independently admitted.

The evidence reported here realizes only bounded parts of that architecture. It progresses from controlled lexical and part-structure calibration to estimated declarations, uncertainty-aware acquisition, and theory-guided repair of generated photographs. Together these experiments supply evidence for an operational substrate for visual-language construction; they do not yet establish an abstract double involution algebroid or unrestricted invention of visual objects, relations, or artistic techniques. Chapter 19 takes up that stronger target explicitly.

The artifact dossier makes the boundary especially visible here. A repaired image is an operational output; a visual-language extension is a structural delta; masks, prompts, relations, and protected content form its transport record; held-out observers and abstention supply admission evidence; and the versioned declaration makes the result replayable. Most natural-image results below admit repairs relative to a supplied or estimated declaration. They do not thereby admit a nontrivial extension of the visual language.

Accordingly, “ground truth” in this chapter is not one privileged target image. In controlled studies it is a registered scene graph, mask, relation, or formation rule. In natural-image studies it is an independently supplied annotation or structural observer, with uncertainty retained when the image does not determine a unique judgment. A repair can satisfy that reference and preserve protected content without being the only acceptable image or the most aesthetically valuable one.

¹1. Chapter roadmap. The chapter states the ARTISTIC architecture, calibrates lexical and structural declarations, and then tests visual repair and concept induction. It closes by separating demonstrated assurance mechanisms from the remaining measurement and mathematical obligations. ↩