Data model and graph design

PCG data shapes

Shape	Use
Volume	bounded 3D domain and boolean/spatial sampling
Surface	landscape/2D domain projected into the world
Line	Spline and Landscape Spline data
Point Data	explicit candidates with transform, bounds, density, steepness, color, seed
Polygon 2D	closed planar areas and polygon union/intersection/difference
Composite Data	lazy spatial union/intersection/difference before concretizing
Attribute Set	nonspatial typed metadata/parameters

Keep spatial operations lazy as long as useful; convert/sample to explicit points only when a downstream point operation or spawner requires it.

Point contract

For every point stream, document:

• coordinate space and projection surface;
• bounds represented by each point;
• density meaning and expected range [0,1];
• seed source and any Mutate Seed boundary;
• required metadata attributes, types, units, and defaults;
• whether transform scale is semantic size or spawn transform;
• which node owns rejection versus presentation.

Point seeds are derived in part from position for world-position consistency. Operations that duplicate points can also duplicate random behavior; use Mutate Seed where independent random branches are intentional.

Graph layering

Use a readable staged graph:

Input and bounds
  -> Sampling/projection
  -> Environmental attributes
  -> Hard filters and exclusion
  -> Density/weight shaping
  -> Selection attributes
  -> Transform variation
  -> Spawn/output

Name reroutes and comment the contract at each stage. Delete temporary attributes before expensive copy/spawn operations when they are no longer needed.

Attribute rules

• Prefer semantic names such as SurfaceSlopeDegrees, BiomeId, Mesh, ClearanceCm.
• Do not silently reuse one attribute with multiple units or meanings.
• Use graph parameters for tool-level inputs; use point attributes for per-point decisions.
• Use Match And Set Attributes for data-driven selection tables and weighted choices.
• Use typed complex attributes in UE 5.8 only when downstream nodes/cook targets support them.

Determinism test

Generate, record point count/transforms/selected assets, cleanup, and regenerate with unchanged inputs. Results must match wherever deterministic output is required. Then change only the exposed seed and confirm intended variation changes without violating hard constraints.