mapwright 0.19.0

Domain-neutral procedural fantasy map & world generation: Voronoi terrain, hydraulic erosion, biomes, rivers, Markov place-names, and shaded-relief SVG.

mapwright

⚠️ Early development (v0.x, alpha). The API is still moving and may change without notice between versions. Usable today, but pin a version (e.g. mapwright==0.10.0) if you depend on it.

Domain-neutral procedural fantasy map & world generation — Voronoi terrain with hydraulic erosion, climate-driven biomes, rivers, Markov place-names, and shaded-relief SVG rendering. Pure Python, numpy-only, fully seed-deterministic.

mapwright produces neutral data (cells, biomes, rivers, polygons) and a self-contained SVG renderer. It has no opinion about your application's models — map its output onto your own tiles/entities however you like.

Gallery

AtlasRenderer — the same neutral terrain, skinned with a hand-drawn art pack. The art here is original, generated through mapwright's companion image service and stamped where the physics put it (mountains on the ranges, forests by climate, sea serpents offshore). mapwright itself ships no art — the pack is the skin:

Render themes — the same continent (same cells, rivers, roads, settlements), re-skinned by swapping a Theme (palette + biome vocabulary). No regeneration:

theme="neon"

theme="dune"

theme="blueprint"

The same theme= drives the town and dungeon renderers too — one skin across all three:

Settlement, theme="neon"

Dungeon, theme="blueprint"

Below: deterministic shaded-relief renders of each built-in preset (or a dungeon), produced by examples/gallery.py:

continent	archipelago	islands
highlands	desert	arctic
pangaea	tropical	DungeonGenerator
SettlementGenerator	Settlement (port)	Settlement (citadel)
RegionalRoadGenerator	RegionGenerator	template="isthmus"
template="atoll"	land_age=0 (young)	land_age=1 (old)
elevation_hint=…	A continent whose macro shape is art-directed by a caller-supplied elevation_hint (a coarse painted land/elevation mask). The host (or an LLM) draws the shape; mapwright fills in organic coasts, erosion, rivers and climate — so maps need never look circular.

The land_age pair above are the same continent at land_age=0 (young, jagged, snow-capped peaks) vs land_age=1 (old, worn down to rounded hills) — a mapwright-original "geological age" knob.

Regenerate them with python examples/gallery.py (SVGs always; PNGs when cairosvg is installed).

Install

pip install mapwright
# hand-drawn / themed atlas rendering (adds Pillow):
pip install "mapwright[atlas]"
# latest from git:
pip install git+https://github.com/sligara7/mapwright.git
# or, for local development:
pip install -e ".[dev]"

Quickstart

from mapwright import SeededRNG, RegionalTerrainGenerator, RegionalSVGRenderer, Marker

# Same seed -> same world, every time.
terrain = RegionalTerrainGenerator(SeededRNG(7)).generate(width=60, height=40)

markers = [Marker(name="Eldmoor", x=30, y=18, kind="settlement_city")]
svg = RegionalSVGRenderer().render(terrain, markers)
open("world.svg", "w").write(svg)

Shape the world with WorldMapConfig — or describe it and let an LLM fill the config:

from mapwright import WorldMapConfig, RegionalTerrainGenerator, SeededRNG

desert = WorldMapConfig.preset("desert")          # ready-made worlds...
custom = WorldMapConfig(continents=7, sea_level=0.55, temperature=-0.8)  # ...or tune
world  = RegionalTerrainGenerator(SeededRNG(1)).generate(60, 40, config=desert)

# Every field is a bounded scalar with a clear meaning, so it doubles as a schema
# a host app (or an LLM) can populate. from_dict clamps junk to valid ranges:
WorldMapConfig.from_dict({"temperature": 5, "continents": -3})  # -> safe, clamped

Presets: continent, pangaea, archipelago, islands, highlands, desert, arctic, tropical.

Terrain defaults to a tectonic-plate simulation (organic coasts + mountain ranges). For a controllable continent archetype, pass a template (Azgaar-style composed heightmap ops) — config still drives sea level, climate, and rivers on top of it:

from mapwright import RegionalTerrainGenerator, SeededRNG, WorldMapConfig, TERRAIN_TEMPLATES

print(list(TERRAIN_TEMPLATES))   # archipelago, volcano, peninsula, isthmus, atoll, continents
world = RegionalTerrainGenerator(SeededRNG(5)).generate(
    80, 58, WorldMapConfig(sea_level=0.55), template="archipelago")

Save and reload worlds (and dungeons) — JSON round-trips losslessly, so a reloaded world renders byte-identically:

from mapwright import RegionalTerrainGenerator, SeededRNG, TerrainResult

terrain = RegionalTerrainGenerator(SeededRNG(7)).generate(60, 40)
open("world.json", "w").write(terrain.to_json())          # ...later...
same = TerrainResult.from_json(open("world.json").read())  # bit-identical

to_dict/from_dict (and to_json/from_json) are available on TerrainResult, Dungeon, and Marker. Numpy rasters and full-precision floats are preserved.

Procedural place-names in several culture styles:

from mapwright import SeededRNG, NameGenerator

namer = NameGenerator(SeededRNG(7))
namer.settlement("nordic")    # -> 'Eirmundheim'
namer.settlement("elvish")    # -> 'Faelynnwood'
namer.region("dwarvish")      # -> 'The Korvald Reach'

Generate a dungeon and render it:

from mapwright import SeededRNG, DungeonGenerator, DungeonSVGRenderer

dungeon = DungeonGenerator(SeededRNG(3)).generate(48, 32)
svg = DungeonSVGRenderer().render(dungeon, labels=True)  # number the rooms
open("dungeon.svg", "w").write(svg)
print(dungeon.ascii())  # or eyeball it as text

Generate a town — an organic footprint split into named wards, each subdivided into building lots, threaded with streets, and optionally walled (try the port and citadel presets):

from mapwright import SeededRNG, SettlementGenerator, SettlementConfig, SettlementSVGRenderer

town    = SettlementGenerator(SeededRNG(7)).generate(90, 90)
port    = SettlementGenerator(SeededRNG(5)).generate(90, 90, SettlementConfig.preset("port"))
citadel = SettlementGenerator(SeededRNG(3)).generate(90, 90, SettlementConfig.preset("citadel"))
open("town.svg", "w").write(SettlementSVGRenderer().render(town))

Settlement presets: hamlet, village, town, city, port, citadel.

What's inside

Component	What it does
SeededRNG	One seed drives everything; .derive(label) yields independent, reproducible sub-streams (unifies stdlib + numpy).
NameGenerator	Order-k character Markov names over hand-authored culture namebases; reproducible across processes.
RegionalTerrainGenerator	Voronoi cells (Lloyd-relaxed) → tectonic-plate heightmap (organic coasts + mountain ranges at plate collisions; percentile sea level) → Planchon–Darboux depression fill → flux + hydraulic/creep erosion → rivers + inland lakes → latitude/elevation climate with rain-shadow → Whittaker biomes. Accepts a template= archetype or an elevation_hint= (caller-drawn macro shape).
compute_cell_polygons	Reconstructs convex Voronoi polygons (half-plane clipping) for vector rendering.
RegionalSVGRenderer	Shaded-relief (hillshade) SVG: biome polygons, coastline, rivers, roads, labelled markers. Takes a theme=.
Theme / THEMES	A render palette + biome vocabulary; re-skins the same terrain — and its towns and dungeons — via one theme= (parchment / neon / dune / blueprint, or your own). The "Dominant Medium" layer.
environment_affordances / summarize_cells	Neutral ecology helpers: biome + climate → affordance tags (scarce_water, predator, …); reduce a set of cells to a CellSummary (dominant biome, mean climate, hydrology, affordances). A host decides what tags mean mechanically.
AtlasRenderer / ArtPack	Hand-drawn / themed PNG: stamps symbols from an external art pack (mountains, forests, hills, settlements, sea decorations) onto the terrain. mapwright ships no art — a pack is a skin. Needs pip install "mapwright[atlas]".
RegionalRoadGenerator	Connects settlement sites with trade routes — an MST whose edges are A*-routed over the terrain (avoids sea, climbs/crosses rivers at a cost).
RegionGenerator	Partitions land into named factions/territories: spread capitals seed a flood fill over the land graph (sea divides them); each Region is Markov-named.
DungeonGenerator	BSP-partitioned rooms + minimum-spanning-tree corridors → rooms, corridor cells, and a walkable grid (with Dungeon.ascii()).
DungeonSVGRenderer	Renders a Dungeon to SVG: walls, carved floor, room outlines, optional tile grid and per-room labels. Takes a theme=.
SettlementGenerator	Self-contained town layout: an organic footprint divided into named Voronoi wards (market, docks, …), each subdivided into building lots, a street network (MST over ward adjacency + main roads from gates to the market), an optional defensive wall (towers + gate gaps, opened at the harbour when coastal), and optional coastline.
SettlementSVGRenderer	Renders a Settlement to SVG: sea, footprint, kind-coloured wards, building lots, streets, wall with towers/gatehouses, labels. Takes a theme=.

Everything is neutral: RegionalTerrainGenerator returns a TerrainResult of TerrainCells (each with a Biome), and you decide how a Biome maps to your world.

Atlas rendering & art packs

RegionalSVGRenderer draws a clean shaded-relief map. For a hand-drawn (or neon, or scrap-metal, or any) look, AtlasRenderer stamps little symbol images — mountains, trees, hills, towns, sea monsters, a compass — placed exactly where the physics put them.

mapwright bundles no art. The renderer is the engine; the art is a separate art pack you point it at, so the same world can wear any style without re-generating anything:

from mapwright import SeededRNG, RegionalTerrainGenerator, ArtPack, AtlasRenderer, Marker

terrain = RegionalTerrainGenerator(SeededRNG(7)).generate(80, 56)
markers = [Marker("Eldmoor", 40, 28, kind="settlement_castle")]

pack = ArtPack.from_directory("path/to/my-art-pack")   # needs mapwright[atlas]
png = AtlasRenderer(pack, scale=12, seed=7).render(terrain, markers, land_age=0.3)
open("atlas.png", "wb").write(png)

An art pack is just a directory of transparent PNG symbols plus an optional manifest.json that maps mapwright's neutral concepts onto art slots:

{
  "name": "my-pack",
  "colors": {"parchment": "#ecdfbf", "water": "#b5cad1",
             "coast": "#463c2c", "label": "#2b2218"},
  "slots": {
    "mountain.young": {"files": ["mountains/sharp/*.png"], "width": 2.0, "anchor": "bottom"},
    "mountain.old":   {"files": ["mountains/eroded/*.png"]},
    "hill":           {"files": ["hills/*.png"]},
    "tree.pine":      {"files": ["trees/pine/*.png"]},
    "tree.deciduous": {"files": ["trees/leafy/*.png"]},
    "city.castle":    {"files": ["cities/castle*.png"]},
    "decoration.compass": {"files": ["compass/*.png"], "anchor": "center"}
  }
}

Slots the renderer asks for: terrain relief — mountain.young / mountain.mid / mountain.old (chosen by land_age), hill, tree.pine / tree.deciduous / tree.cactus (by climate), dune; settlements — city.castle / city.large / city.town / city.village (by marker kind); decorations — decoration.creature / decoration.ship / decoration.compass. A missing fine slot falls back to a coarser sibling (mountain.mid → any mountain.*), so partial packs still render. With no manifest.json, ArtPack.from_directory() auto-discovers slots from a conventional folder layout. Because packs are pure data, a host like an image-generation service can produce them on demand in any style — the generation stays the same; the pack is the skin.

Render themes

The vector RegionalSVGRenderer takes a Theme — a palette plus an optional biome vocabulary — so the same neutral terrain re-skins into wildly different worlds without regenerating anything. The neutral Biome enum never changes; a theme just decides how each biome looks and is named:

from mapwright import RegionalSVGRenderer, SettlementSVGRenderer, DungeonSVGRenderer, THEMES

svg  = RegionalSVGRenderer(theme="neon").render(terrain, markers, roads=roads)
town = SettlementSVGRenderer(theme="neon").render(settlement)   # same theme skins the town
dgn  = DungeonSVGRenderer(theme="blueprint").render(dungeon)    # …and the dungeon
# built-ins: "parchment" (default), "neon" (Tron/digital-grid), "dune" (sand), "blueprint"
THEMES["neon"].biome_label(Biome.OCEAN)   # -> "Void"  (the vocabulary layer)

All three renderers take the same theme=, so one theme skins the world map, its towns, and its dungeons together (a Theme carries nested SettlementPalette + DungeonPalette, importable from mapwright.themes for custom packs). A Theme is plain hex-string data (JSON-friendly), so a host — or the same image service that makes art packs — can author new ones. This is the "Dominant Medium" idea from mapwright's longer-term vision: a sand planet, a digital grid, and an irradiated waste are the same map wearing different skins. Pair a theme with a matching ArtPack for a full restyle of both the vector and hand-drawn renders.

Determinism

Every generator draws from a SeededRNG. The same seed (and parameters) reproduces an identical world — terrain, names, rivers, and SVG — across runs and across processes (the Markov chains are built in sorted order, so output never depends on PYTHONHASHSEED).

Performance

Pure Python + numpy, single-threaded. Typical map/town sizes generate in well under a second; examples/benchmark.py prints a table for your machine. Rough figures (numbers are machine-dependent):

Generator	Size	Time
Terrain	64×44 (≈470 cells)	~150 ms
Terrain	120×90 (1500 cells, capped)	~1.8 s
Dungeon	80×60 (≈50 rooms)	~9 ms
Settlement	pop 9000 (50 wards, ~1100 lots)	~65 ms
Roads / regions	on a 120×90 map	a few ms

Two things worth knowing:

• Terrain cell count is capped at 1500 (cell_area clamp in generate), which bounds the hydrology/climate/graph work — but the initial Voronoi rasterisation is per-pixel, so total time still grows roughly linearly with width × height on large maps. Raise cell_area (fewer, coarser cells) to trade detail for speed, e.g. generate(w, h, cell_area=12).
• Dungeon corridor connection is a dense MST (~O(rooms³)), so dungeons with hundreds of rooms get slow — keep them modest or raise DungeonConfig.min_leaf for fewer, larger rooms.

API stability & contract

The public API is exactly the names exported in mapwright.__all__ — that's the contract. It's pinned by tests/test_api_contract.py (public surface, key signatures), so an accidental breaking change fails CI.

For the world parameters specifically, WorldMapConfig.json_schema() returns a JSON Schema (draft 2020-12) — the machine-readable contract a host app or LLM can validate/generate against, then feed through WorldMapConfig.from_dict() (which clamps to valid ranges). Schema and runtime clamping are generated from the same field spec, so they can't drift.

Versioning follows SemVer. While at 0.x the API may still change between minor versions; every change is recorded in CHANGELOG.md. Pin a tag or commit if you depend on it.

Development

python -m venv .venv && . .venv/bin/activate
pip install -e ".[dev]"
pytest

Credits & license

MIT licensed (see LICENSE). Algorithms were implemented clean-room from the publicly described techniques of Azgaar's Fantasy-Map-Generator (MIT) and Martin O'Leary / Ryan L. Guy's FantasyMapGenerator (Zlib); see NOTICE for details. The bundled name lists are original.