ifc-0274

18.8 A visual-language benchmark

ARTISTIC-T2I–0 uses procedurally generated scenes with a frozen base language of objects, attributes, and elementary relations. The evaluator withholds one relation or composition rule while retaining primitive rendering operations from which it can be realized. Nonce names and generated geometries prevent retrieval of an existing natural-language label.

Two conditions separate different creative claims.

Latent but unnamed.

The frozen generator can already realize the withheld distinction, but the conditioning language cannot reliably express it. The target is productive language acquisition without changing the image model.

Structurally unavailable.

The base generator cannot realize the distinction under any registered prompt. The target is a candidate finite extension through a bounded adapter or control-field extension together with a new language construction; the change is retained only after independent admission.

The new symbol is evaluated on unseen objects, colors, counts, styles, spatial contexts, and grammatical positions. A communication split gives its compact definition and a few admitted examples to a fresh interpreter, which must use it in new requests. A transport split changes seeds and renderers and, where interfaces permit, the underlying diffusion model.

Experiment: ARTISTIC-T2I–0. Epistemic status: proposed integrated benchmark, not a completed visual-language result.
World: procedural visual scenes with a frozen vocabulary and one withheld relation or composition rule.
Joint construction: visual-theory extension, executable grounding or adapter, and an active contrastive-generation policy.
Admission: relation accuracy, counterfactual grounding, disentanglement, unseen composition, fresh-interpreter communication, and cross-seed and cross-renderer transport.
Controls: prompt search, textual inversion, fixed LoRA or adapter, structured attention guidance, the CLIC–OPTIC–RELIC factorial, and semantic, rendering, and query-policy oracles.
Boundary: image novelty or a learned token alone is not an admitted visual language.