Ue Procedural Generation

ue-procedural-generation

You are an expert in Unreal Engine's procedural generation systems, including the PCG framework, ProceduralMeshComponent, instanced static meshes, noise functions, and spline-based generation.

Context Check

Before advising, read .agents/ue-project-context.md to determine:

• Whether the PCG plugin is enabled (plugins list)
• Target generation type: world layout, terrain, dungeon, vegetation, runtime mesh
• Performance constraints (mobile, console, Nanite enabled)
• Multiplayer requirements (server authority vs. deterministic seeding)

Information Gathering

Ask for clarification on:

1. Generation type: world population (PCG), runtime mesh (ProceduralMeshComponent), instanced geometry (ISM/HISM), or spline-driven?
2. Timing: editor-time baked result or runtime dynamic generation?
3. Instance count: hundreds (ISM) or tens of thousands (HISM)?
4. Collision: does generated geometry need physics collision?
5. Determinism: same seed must produce same result across sessions or network clients?

1. PCG Framework (UE 5.2+)

Node-based rule-driven world generation. Operates on point clouds with transform, density, color, seed, and metadata attributes.

Plugin Setup

// Build.cs
PublicDependencyModuleNames.Add("PCG");

// .uproject Plugins array
{ "Name": "PCG", "Enabled": true }

Core Classes

UPCGComponent Key API (from PCGComponent.h)

// Assign graph (NetMulticast)
void SetGraph(UPCGGraphInterface* InGraph);

// Trigger generation (NetMulticast, Reliable) — use for multiplayer
void Generate(bool bForce);

// Local non-replicated generation
void GenerateLocal(bool bForce);

// Cleanup
void Cleanup(bool bRemoveComponents);
void CleanupLocal(bool bRemoveComponents);

// Notify to re-evaluate after Blueprint property change
void NotifyPropertiesChangedFromBlueprint();

// Read generated output
const FPCGDataCollection& GetGeneratedGraphOutput() const;

Generation triggers (EPCGComponentGenerationTrigger):

• GenerateOnLoad — one-shot on BeginPlay
• GenerateOnDemand — explicit Generate() call only
• GenerateAtRuntime — budget-scheduled by UPCGSubsystem

UPCGGraph Node API (from PCGGraph.h)

// Add node by settings class
UPCGNode* AddNodeOfType(TSubclassOf<UPCGSettings> InSettingsClass, UPCGSettings*& DefaultNodeSettings);

// Connect two nodes
UPCGNode* AddEdge(UPCGNode* From, const FName& FromPinLabel, UPCGNode* To, const FName& ToPinLabel);

// Graph parameters (typed template)
template<typename T>
TValueOrError<T, EPropertyBagResult> GetGraphParameter(const FName PropertyName) const;

template<typename T>
EPropertyBagResult SetGraphParameter(const FName PropertyName, const T& Value);

Custom Blueprint PCG Node

Derive from UPCGBlueprintBaseElement:

UCLASS(BlueprintType, Blueprintable)
class UMyPCGNode : public UPCGBlueprintBaseElement
{
    GENERATED_BODY()
public:
    UFUNCTION(BlueprintNativeEvent, BlueprintCallable, Category = "PCG|Execution")
    void Execute(const FPCGDataCollection& Input, FPCGDataCollection& Output);
};

// In Execute:
FRandomStream Stream = GetRandomStreamWithContext(GetContextHandle()); // deterministic seed

for (const FPCGTaggedData& In : Input.GetInputsByPin(PCGPinConstants::DefaultInputLabel))
{
    const UPCGPointData* InPts = Cast<UPCGPointData>(In.Data);
    if (!InPts) continue;

    UPCGPointData* OutPts = NewObject<UPCGPointData>();
    for (const FPCGPoint& Pt : InPts->GetPoints())
    {
        FPCGPoint NewPt = Pt;
        NewPt.Density = Stream.FRandRange(0.5f, 1.0f);
        OutPts->GetMutablePoints().Add(NewPt);
    }
    Output.TaggedData.Emplace_GetRef().Data = OutPts;
}

Key UPCGBlueprintBaseElement properties:

• bIsCacheable = false — when node spawns actors or components
• bRequiresGameThread = true — for actor spawn, component add
• CustomInputPins / CustomOutputPins — extra typed pins

PCG Determinism

PCG graphs are deterministic by default — the same seed produces identical output. Each node receives a seeded random stream via GetRandomStreamWithContext(). To vary output across instances, set the PCG component's Seed property. For multiplayer, ensure all clients use the same seed (replicate via GameState or pass as spawn parameter).

// Set PCG seed at runtime for deterministic variation
UPCGComponent* PCG = FindComponentByClass<UPCGComponent>();
PCG->Seed = MyDeterministicSeedValue;
PCG->Generate(); // Regenerate with new seed

PCG Data Types

Point data is the most common — most PCG nodes consume and produce point collections. Also available: UPCGTextureData, UPCGPrimitiveData, UPCGDynamicMeshData.

See references/pcg-node-reference.md for all node types, settings fields, and pin labels.

2. ProceduralMeshComponent

// Build.cs
PublicDependencyModuleNames.Add("ProceduralMeshComponent");

Core API

// Create section: vertices, triangles (CCW = front), normals, UVs, colors, tangents
void CreateMeshSection(int32 SectionIndex,
    const TArray<FVector>& Vertices, const TArray<int32>& Triangles,
    const TArray<FVector>& Normals,  const TArray<FVector2D>& UV0,
    const TArray<FColor>& VertexColors, const TArray<FProcMeshTangent>& Tangents,
    bool bCreateCollision);

// Updates vertex positions (incl. collision if enabled). Cannot change topology.
void UpdateMeshSection(int32 SectionIndex,
    const TArray<FVector>& Vertices, const TArray<FVector>& Normals,
    const TArray<FVector2D>& UV0,    const TArray<FColor>& VertexColors,
    const TArray<FProcMeshTangent>& Tangents);

void ClearMeshSection(int32 SectionIndex);
void ClearAllMeshSections();
void SetMeshSectionVisible(int32 SectionIndex, bool bNewVisibility);
void SetMaterial(int32 ElementIndex, UMaterialInterface* Material);

Terrain Grid Example

void ATerrainActor::Build(int32 Grid, float Cell)
{
    TArray<FVector> Verts; TArray<int32> Tris; TArray<FVector> Norms;
    TArray<FVector2D> UVs; TArray<FColor> Colors; TArray<FProcMeshTangent> Tangs;

    for (int32 Y = 0; Y <= Grid; Y++)
    for (int32 X = 0; X <= Grid; X++)
    {
        float Z = SampleOctaveNoise(X * Cell, Y * Cell, 4, 0.5f, 2.f, 80.f);
        Verts.Add(FVector(X * Cell, Y * Cell, Z));
        Norms.Add(FVector::UpVector);
        UVs.Add(FVector2D((float)X / Grid, (float)Y / Grid));
    }
    for (int32 Y = 0; Y < Grid; Y++)
    for (int32 X = 0; X < Grid; X++)
    {
        int32 BL = Y*(Grid+1)+X, BR=BL+1, TL=BL+(Grid+1), TR=TL+1;
        Tris.Add(BL); Tris.Add(TL); Tris.Add(TR);
        Tris.Add(BL); Tris.Add(TR); Tris.Add(BR);
    }
    ProceduralMesh->CreateMeshSection(0, Verts, Tris, Norms,
                                       UVs, Colors, Tangs, /*bCreateCollision=*/true);
}

Performance Notes

• One draw call per CreateMeshSection. Keep vertex count < 65K per section.
• UpdateMeshSection updates vertex positions and collision (if enabled) but cannot change topology — call CreateMeshSection for new triangles.
• ProceduralMesh does not support Nanite.
• Compute vertex data on background thread; call CreateMeshSection on game thread only.

Async Mesh Generation

Generate vertices on a background thread, then apply on the game thread:

// Background task — compute vertices
class FMeshGenTask : public FNonAbandonableTask
{
public:
    TArray<FVector> Vertices;
    TArray<int32> Triangles;
    void DoWork() { /* Marching cubes, noise sampling, etc. */ }
    FORCEINLINE TStatId GetStatId() const { RETURN_QUICK_DECLARE_CYCLE_STAT(FMeshGenTask, STATGROUP_ThreadPoolAsyncTasks); }
};

// Launch and poll
// Use FAsyncTask (not FAutoDeleteAsyncTask) when polling IsDone() is needed.
// FAutoDeleteAsyncTask deletes itself on completion — calling IsDone() afterward is a use-after-free.
auto* Task = new FAsyncTask<FMeshGenTask>();
Task->StartBackgroundTask();
// Poll safely: if (Task->IsDone()) { /* use Task->GetTask().Vertices */ delete Task; }

Collision on Procedural Meshes

Set UProceduralMeshComponent::bUseComplexAsSimpleCollision = true to use the rendered triangles directly for collision. This is accurate but expensive — only use for static geometry. For dynamic or high-poly meshes, generate simplified convex hulls instead.

3. Instanced Static Meshes (ISM / HISM)

Key ISM API (from InstancedStaticMeshComponent.h)

virtual int32 AddInstance(const FTransform& T, bool bWorldSpace = false);
virtual TArray<int32> AddInstances(const TArray<FTransform>& Ts,
    bool bShouldReturnIndices, bool bWorldSpace = false, bool bUpdateNavigation = true);
virtual bool UpdateInstanceTransform(int32 Idx, const FTransform& NewT,
    bool bWorldSpace = false, bool bMarkRenderStateDirty = false, bool bTeleport = false);
virtual bool BatchUpdateInstancesTransforms(int32 StartIdx, const TArray<FTransform>& NewTs,
    bool bWorldSpace = false, bool bMarkRenderStateDirty = false, bool bTeleport = false);
bool GetInstanceTransform(int32 Idx, FTransform& OutT, bool bWorldSpace = false) const;
virtual bool RemoveInstance(int32 InstanceIndex);  // O(n) for ISM; triggers async BVH rebuild for HISM
virtual void PreAllocateInstancesMemory(int32 AddedCount);
int32 GetNumInstances() const;

// Per-instance custom float data (read in materials via PerInstanceCustomData)
virtual void SetNumCustomDataFloats(int32 N);
virtual bool SetCustomDataValue(int32 Idx, int32 DataIdx, float Value,
    bool bMarkRenderStateDirty = false);
virtual bool SetCustomData(int32 Idx, TArrayView<const float> Floats,
    bool bMarkRenderStateDirty = false);

Culling properties: InstanceStartCullDistance, InstanceEndCullDistance, InstanceLODDistanceScale, bUseGpuLodSelection.

Vegetation Scatter (HISM + Terrain Trace)

HISM->SetStaticMesh(TreeMesh);
HISM->SetNumCustomDataFloats(1);
HISM->PreAllocateInstancesMemory(Count);
FRandomStream Rand(Seed);
TArray<FTransform> Transforms; Transforms.Reserve(Count);

for (int32 i = 0; i < Count; i++)
{
    FVector Loc(Rand.FRandRange(Min.X, Max.X), Rand.FRandRange(Min.Y, Max.Y), 0);
    FHitResult Hit;
    if (GetWorld()->LineTraceSingleByChannel(Hit,
        Loc + FVector(0,0,5000), Loc - FVector(0,0,5000), ECC_WorldStatic))
        Loc.Z = Hit.Location.Z;

    Transforms.Add(FTransform(
        FRotator(0, Rand.FRandRange(0,360), 0), Loc,
        FVector(Rand.FRandRange(0.8f, 1.3f))));
}

TArray<int32> Indices = HISM->AddInstances(Transforms, true, true);
for (int32 i = 0; i < Indices.Num(); i++)
    HISM->SetCustomDataValue(Indices[i], 0, Rand.FRand(), false);
HISM->MarkRenderStateDirty();

Foliage System

The editor's Foliage paint mode uses AInstancedFoliageActor which internally wraps UHierarchicalInstancedStaticMeshComponent. For procedural foliage at scale, use UProceduralFoliageComponent with UProceduralFoliageSpawner — it distributes foliage via simulation (species competition, shade tolerance) rather than manual painting.

Per-instance collision: Enable bUseDefaultCollision on the ISM component. Each instance inherits the static mesh's collision. For custom per-instance collision shapes, use separate actors — ISM does not support unique collision per instance.

Platform limits: HISM GPU buffer caps vary by platform (~1M on desktop, ~100K on mobile). Monitor with stat Foliage. Split large populations across multiple HISM components.

4. Noise and Math

// Built-in Perlin (all output in [-1, 1])
float N1 = FMath::PerlinNoise1D(X * Freq);
float N2 = FMath::PerlinNoise2D(FVector2D(X, Y) * Freq);
float N3 = FMath::PerlinNoise3D(FVector(X, Y, Z) * Freq);

// Octave noise
float OctaveNoise(float X, float Y, int32 Oct, float Persist, float Lacu, float Scale)
{
    float V=0, A=1, F=1.f/Scale, Max=0;
    for (int32 i=0; i<Oct; i++) {
        V += FMath::PerlinNoise2D(FVector2D(X,Y)*F) * A;
        Max += A; A *= Persist; F *= Lacu;
    }
    return V / Max;
}

// Seeded deterministic random
FRandomStream Stream(Seed);
float R = Stream.FRandRange(Min, Max);
int32 I = Stream.RandRange(MinI, MaxI);
FVector Dir = Stream.VRand();

Height/density maps: Sample UTexture2D pixel data via FTexturePlatformData to drive terrain height or placement density. Lock with BulkData.Lock(LOCK_READ_ONLY), read, then unlock.

Poisson disc sampling (minimum-separation scatter for natural placement) — full Bridson algorithm implementation in references/procedural-mesh-patterns.md.

5. Spline Components

USplineComponent API (from SplineComponent.h)

// Build spline (always batch with bUpdateSpline=false, call UpdateSpline() once after)
void AddSplinePoint(const FVector& Pos, ESplineCoordinateSpace::Type Space, bool bUpdate=true);
void SetSplinePoints(const TArray<FVector>& Pts, ESplineCoordinateSpace::Type Space, bool bUpdate=true);
void ClearSplinePoints(bool bUpdate=true);
virtual void UpdateSpline();  // Rebuild reparameterization table

// Query by arc-length distance
FVector    GetLocationAtDistanceAlongSpline(float Dist, ESplineCoordinateSpace::Type Space) const;
FVector    GetDirectionAtDistanceAlongSpline(float Dist, ESplineCoordinateSpace::Type Space) const;
FVector    GetRightVectorAtDistanceAlongSpline(float Dist, ESplineCoordinateSpace::Type Space) const;
FRotator   GetRotationAtDistanceAlongSpline(float Dist, ESplineCoordinateSpace::Type Space) const;
FTransform GetTransformAtDistanceAlongSpline(float Dist, ESplineCoordinateSpace::Type Space, bool bUseScale=false) const;
float      GetSplineLength() const;

// Point editing
int32  GetNumberOfSplinePoints() const;
void   SetSplinePointType(int32 Idx, ESplinePointType::Type Type, bool bUpdate=true);
void   SetClosedLoop(bool bClosed, bool bUpdate=true);
void   SetTangentsAtSplinePoint(int32 Idx, const FVector& Arrive, const FVector& Leave,
           ESplineCoordinateSpace::Type Space, bool bUpdate=true);

Point types: Linear, Curve, Constant, CurveClamped, CurveCustomTangent.

FindInputKeyClosestToWorldLocation(WorldLocation) — returns the spline key nearest to a world position (useful for snapping actors to splines).

Runtime modification: Call AddSplinePoint(), RemoveSplinePoint(), or SetLocationAtSplinePoint() then UpdateSpline() to rebuild. Batch modifications before calling UpdateSpline() — each call recalculates the entire spline.

Spline Placement Example

// Place instances evenly along spline
float Len = Spline->GetSplineLength();
for (float D = 0.f; D <= Len; D += Spacing)
{
    FTransform T = Spline->GetTransformAtDistanceAlongSpline(D, ESplineCoordinateSpace::World);
    HISM->AddInstance(T, /*bWorldSpace=*/true);
}

USplineMeshComponent (Mesh Deformation)

#include "Components/SplineMeshComponent.h"
USplineMeshComponent* SM = NewObject<USplineMeshComponent>(this);
SM->SetStaticMesh(PipeMesh);
SM->RegisterComponent();

FVector SP, ST, EP, ET;
Spline->GetLocationAndTangentAtSplinePoint(Seg,   SP, ST, ESplineCoordinateSpace::Local);
Spline->GetLocationAndTangentAtSplinePoint(Seg+1, EP, ET, ESplineCoordinateSpace::Local);
SM->SetStartAndEnd(SP, ST, EP, ET, /*bUpdateMesh=*/true);
SM->SetForwardAxis(ESplineMeshAxis::X);

6. Runtime Mesh Generation Patterns

See references/procedural-mesh-patterns.md for full implementations:

• Marching Cubes — isosurface from 3D density scalar field
• Dungeon BSP — BSP partition into rooms, L-corridor carving, tile-to-mesh
• L-System — string rewriting + turtle interpreter to HISM branches
• Wave Function Collapse — constraint-propagation tile grid layout
• Async mesh generation — background thread vertex computation, game thread CreateMeshSection
• Spline road extrusion — cross-section profile swept along USplineComponent

// Marching Cubes result → ProceduralMesh
ProceduralMesh->CreateMeshSection(0, MarchVerts, MarchTris, MarchNormals,
    MarchUVs, {}, {}, /*bCreateCollision=*/true);

Common Mistakes and Anti-Patterns

PCG

• Calling GenerateLocal() in Tick — generation is not free; use GenerateOnDemand and regenerate only on data change.
• Using GenerateLocal() in multiplayer — it is NOT replicated; use Generate(bForce) (NetMulticast).
• Heavy custom nodes with bIsCacheable = true — only cache if output depends solely on inputs + seed.
• Graphs with bIsEditorOnly = true fail to cook into packaged builds.

ProceduralMeshComponent

• Passing bCreateCollision=false to CreateMeshSection — characters fall through the mesh.
• Calling UpdateMeshSection expecting topology to change — vertex count must match; use CreateMeshSection for new triangles.
• Using ProceduralMesh for Nanite-scale terrain — not supported; use Landscape or PCG + ISM.
• Wrong triangle winding (CW instead of CCW) — polygons are invisible due to back-face culling.

ISM / HISM

• Using ISM above ~500 instances — switch to HISM for BVH culling.
• Setting bMarkRenderStateDirty=true on every UpdateInstanceTransform in a loop — only set true on the last call.
• Skipping PreAllocateInstancesMemory before bulk add — repeated realloc degrades performance.

Splines

• Calling AddSplinePoint(bUpdateSpline=true) in a loop — rebuilds reparameterization table every call; use false and call UpdateSpline() once.
• Using spline input key (not distance) for even spacing — key is NOT proportional to arc length.

Multiplayer

• Procedural content must be deterministic (same seed) or server-authoritative.
• GenerateLocal() does not replicate; Generate(bool) is NetMulticast, Reliable.

Related Skills

• ue-actor-component-architecture — component lifecycle, registration, replication
• ue-physics-collision — collision profiles, complex vs. simple on generated geometry
• ue-cpp-foundations — NewObject, TSubclassOf, TArray, memory management

Reference Files

• references/pcg-node-reference.md — all PCG node types, pin labels, settings fields, determinism checklist
• references/procedural-mesh-patterns.md — quad grid, marching cubes, dungeon BSP, L-system, WFC, spline road