From Growth Model to Planning Assistant: Terrain-Conditioned Diffusion with Auditable Language Operations and Jurisdictional Rule Packs

Abstract

Municipal planning consists of many distinct duties, and only some of them are automatable. This paper presents a pipeline that covers a defined subset of those duties end to end and states which duties it does not touch. At the core is a conditional denoising diffusion model that generates street-level growth alternatives (roads, continuous built-up density, and proposed amenity locations) for real settlements, trained on observed 1980-2020 change assembled automatically from OpenStreetMap, open elevation tiles, and the multi-epoch GHS-BUILT-S surface. Growth style is provided by a mixture of five experts, one per contemporary growth regime (village, hilly town, planned fringe, informal peri-urban frontier, and megacity edge), dispatched by a hard, auditable router on conditioning statistics; a window in the Swiss Alps and a window on Delhi's edge route to different experts trained on the appropriate morphology. Around the generative core sit the stages a planning office needs to use it: an eleven-metric sustainability scorecard with best-of-N selection, a typed-zoning classifier with per-class abstention, demand forecasting from the same built-up epochs, compliance checking whose thresholds come from jurisdictional rule packs extracted from the planning documents a jurisdiction actually publishes (a zoning ordinance, a comprehensive plan, or a model-code excerpt, read by a deterministic pattern layer and an optional local language model, with every extracted value quoting the sentence it came from), synthesis of public comments into validated plan operations, and deterministic drafting of staff reports in which every sentence traces to a computed quantity. Natural-language control flows through a closed operation schema, and the language model behind it is pluggable, defaulting to a locally served Gemma model so sensitive planning inputs never leave the machine, with a rule parser as the fallback when no model is available. All five experts trained to convergence (final noise-prediction losses 0.024 to 0.042); the zoning classifier reaches leave-one-town-out macro-F1 0.257 across five morphology families; and a controlled reproduction of a previously documented zoning failure on a United States suburb shows the repair came from training data rather than feature calibration, a negative result reported as such. Code, trained weights, a client-side browser deployment, and the full pipeline are public. Duties requiring discretion, negotiation, or legal judgment are identified explicitly and left to people.

Keywords: Diffusion Models, Urban Planning, Generative Urban Design, Land Use, GeoAI, Urban Growth