Most failures of generated city architecture are not rendering failures. The image is crisp; the light is plausible; something deeper is wrong. A greystone wears a turret it would never have worn. A market hall cupola grows a clock face the record never authorized. A desert house arrives with a pitched roof borrowed from a wetter climate. GenAnywhere, the content-shaping persona used across this series, treats these as schema violations, and it catches them with a validation model built on one idea: a typology is a constrained vocabulary of parts, and a believable district is a graph that stays inside its vocabularies.
Vocabularies, not styles
A style is a mood; a vocabulary is a list. The Beaux-arts apartment hotel typology, in the GenAnywhere library, admits a limestone base with an arched entry, brick and terra-cotta upper stories, ornate cornices, a steep mansard roof with dormers, and a carriage gate behind an iron fence. It admits those parts with proportion ranges and adjacency rules: the mansard crowns, the carriage gate sits on the side garden, the entry arch stays blank of lettering. Generation then samples from the list rather than improvising, which is why two runs of the same typology produce cousins rather than strangers.
The payoff is diagnostic. When a generated building reads wrong, the validator can name the offending part and the rule it broke, instead of leaving a reviewer to squint at a pretty image and distrust it.
Negative constraints: the parts a typology refuses
The admitted list defines what a typology can say; the refused list defines its accent. Chicago greystones refuse turrets. Charleston single houses refuse full-frontal facades. Prairie school houses refuse vertical emphasis of any kind. Mid-century desert houses refuse pitched roofs and shutters. These negative constraints are stored with the same weight as the positive ones, because pastiche accumulates one borrowed part at a time: a turret here, a shutter there, and the district sounds like nowhere.
Negative constraints also carry the editorial rules that keep packages publishable. Facade entities carry a lettering boolean set false across the run, which is why cast-iron storefronts render as blank dark glazing, cupolas carry no clock faces, and armory gates carry no insignia. The validator rejects any variant whose surfaces acquire markings, long before a human sees them.
Cross-entity invariants at district scale
Per-building vocabularies are necessary but not sufficient; districts need invariants that span entities. Street-wall continuity holds along each block face within a setback tolerance, with courtyard courts and fenced squares as declared exceptions. Scale coherence limits the step between neighboring rooflines so a triple-decker never towers over a bungalow row without a declared transition. Seasonal and light coherence binds every image in a run to one season and one sun direction, because a district photographed in two weathers is two districts. Vista termination requires each long view to close on a declared landmark: a cupola, a castellated tower, a mountain rim.
These invariants run at validation time, on the graph, in milliseconds. A violation returns the two entities in conflict and the rule name, which turns an aesthetic argument into a field edit. That is the whole economic case for the model: milliseconds of validation instead of hours of re-rolling.
Validators, variant selection, and the rejection log
A run emits several variants per block; the validator scores each against vocabularies and invariants and rejects the rest with reasons. GenAnywhere keeps every rejection in a log, because the log is the team's training data: recurring rejections point at typology definitions that need tightening, and rare ones point at prompts that overreached. Selection among surviving variants is then an editorial act — viewpoint, light, and which facade leads the frame — performed on a candidate set that is already believable.
Rendering the validated graph
Only validated graphs reach the renderers. The preview renderer outlines the graph for editorial review on every save. The asset renderer frames one cover and two inline views per article from distinct viewpoints, so the bungalow belt, the garden square, and the trinity alley each lead exactly one image. The document renderer flattens the same graph into each platform's native format, from plain markdown essays to HTML templates with label footers. Because every renderer reads one validated graph, an edit to a typology's proportion range propagates to previews, images, and documents in a single build — and the district keeps its accent everywhere it appears.



Top comments (0)