2nd/10_Wiki/Topics/Architecture/Procedural-Architecture-Systems.md

---
id: wiki-2026-0508-procedural-architecture-systems
title: Procedural Architecture Systems
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [PCG Architecture, Procedural Generation Systems]
duplicate_of: none
source_trust_level: A
confidence_score: 0.85
verification_status: applied
tags: [architecture, pcg, gamedev, generation]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: cpp
  framework: unreal-houdini
---

# Procedural Architecture Systems

## 매 한 줄
> **"매 building / city 매 hand-author 의 X — 매 rules + grammar + noise 매 generate"**. 매 No Man's Sky / Minecraft / Townscaper / Houdini 매 production-grade. 매 2026 LLM-driven layout 매 mainstream (e.g. Inworld + procedural blockout).

## 매 핵심

### 매 layers
- **Macro**: 매 city / region — Voronoi, road network, biome.
- **Meso**: 매 building footprint, lot subdivision (BSP, L-system).
- **Micro**: 매 facade grammar, room layout, props placement.

### 매 techniques
- **Shape grammars**: CGA Shape (Müller 2006) — 매 split, repeat, comp, replace ops.
- **Wave Function Collapse**: tile constraint solver — 매 Townscaper 매 famous.
- **L-systems**: 매 organic / fractal (street trees, ivy).
- **Noise**: Perlin / Simplex / Worley — 매 base terrain.
- **Constraint propagation**: 매 SAT/CSP for room layout.

### 매 응용
1. AAA open-world filler (Ubisoft, GTA city blocks).
2. Roguelike dungeon (Spelunky, Hades).
3. ArchViz mass modeling — Houdini PDG.
4. UGC (Roblox, Dreams) — runtime gen.

## 💻 패턴

### Wave Function Collapse (Python)
```python
import numpy as np, random

def wfc(tiles, adjacency, w, h):
    grid = [[set(tiles) for _ in range(w)] for _ in range(h)]
    while any(len(c) > 1 for row in grid for c in row):
        # pick lowest-entropy cell
        x, y = min(((i, j) for i, row in enumerate(grid)
                    for j, c in enumerate(row) if len(c) > 1),
                   key=lambda p: len(grid[p[0]][p[1]]))
        choice = random.choice(list(grid[x][y]))
        grid[x][y] = {choice}
        propagate(grid, adjacency, x, y)
    return [[next(iter(c)) for c in row] for row in grid]
```

### Shape grammar (CGA-style)
```python
def split_facade(rect, rules):
    # axiom: Facade -> Floor*
    floors = subdivide(rect, axis='y', sizes=[3.0, 3.0, 3.0])
    for f in floors:
        windows = subdivide(f, axis='x', sizes=[1.5]*5)
        for w in windows:
            emit_mesh('window_unit', w)
```

### L-system street network
```python
def lsystem(axiom, rules, depth):
    s = axiom
    for _ in range(depth):
        s = ''.join(rules.get(c, c) for c in s)
    return s

# F=forward, +=turn, [=push
streets = lsystem("F", {"F": "F[+F]F[-F]F"}, 4)
```

### Noise-based terrain (GLSL)
```glsl
float fbm(vec2 p) {
    float v = 0.0, a = 0.5;
    for (int i = 0; i < 6; i++) {
        v += a * snoise(p);
        p *= 2.0; a *= 0.5;
    }
    return v;
}
```

### Houdini PDG-style chunked gen (pseudo)
```python
# Chunk world into tiles, generate independently
for tile in tiles:
    seed = hash((tile.x, tile.z))
    rng = np.random.default_rng(seed)
    buildings = place_buildings(tile, rng)
    export(f"tile_{tile.x}_{tile.z}.usd", buildings)
```

### LLM-augmented blockout (2026)
```python
prompt = f"""Design a {style} village, 12 buildings.
Output JSON: [{{type, x, z, rotation, size}}]"""
plan = json.loads(claude.complete(prompt))
for b in plan:
    spawn_building(b['type'], (b['x'], b['z']), b['rotation'])
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| 매 grid-tile aesthetic | WFC |
| 매 organic city | L-system + agent sim |
| 매 facade detail | Shape grammar (CGA) |
| 매 terrain | Noise + erosion sim |
| 매 narrative-coherent | LLM-driven plan + procedural detail |

**기본값**: 매 hybrid — LLM macro plan + procedural meso/micro.

## 🔗 Graph
- 부모: [[Procedural Generation]]
- Adjacent: [[Perlin Noise]] · [[CSP]]

## 🤖 LLM 활용
**언제**: 매 high-level layout / narrative coherence — LLM 매 plan, procedural 매 detail. 매 prompt-to-blockout (2026 Inworld, Promethean).
**언제 X**: 매 real-time per-frame gen — LLM latency 너무 김. 매 precompute / edge-cache.

## ❌ 안티패턴
- **Pure noise**: 매 visible repetition / no semantic. 매 layered noise + masking 필수.
- **No seed control**: 매 reproducibility 부재 — debugging hell.
- **Monolithic gen**: 매 entire world in one pass — chunked / streamable 필요.

## 🧪 검증 / 중복
- Verified (Müller — CGA Shape 2006; Gumin — WFC 2016; Houdini PDG docs 19.5+).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — full PCG architecture entry |