LIVERPS & Scene Composition

Nick Porcino application memory
USDOpenUSDLIVERPS

OpenUSD LIVERPS Composition Algorithm

Overview

The LIVERPS algorithm defines OpenUSD’s composition strength ordering, determining how scene description layers combine to produce the final composed stage. The acronym represents seven composition operators applied in decreasing strength order: Local, Inherits, VariantSets, Relocates (a modifier), References, Payloads, Specializations.

Layer Loading and PrimSpec Parsing

Layer Resolution Phase

  1. Root Layer Identification: The composition engine begins with the root layer specified when opening a stage
  2. Dependency Discovery: Each layer is parsed to identify composition arcs (references, payloads, inherits, etc.)
  3. Layer Loading: Referenced layers are loaded on-demand, creating a dependency graph
  4. PrimSpec Extraction: For each loaded layer, primspecs are parsed and indexed by prim path

PrimSpec Structure

Each primspec contains:

  • Metadata: Composition arcs, active state, instanceable flag
  • Properties: Attributes and relationships with their values and metadata
  • Child Prim Names: References to nested primspecs

LIVERPS Introduction

The LIVERPS mnemonic orders the USD compositional arcs in order of strength from strongest to weakest (with the exception of relocates as will be shown later). The LIVERPS arcs may be grouped into four functional categories that work together to build composed stages. In the following explanations, the term “opinion” is often used. An opinion is just a piece of authored data about a prim or property, whether it comes from a layer, a reference, or an override.

1. Local Foundation

  • L (Local): directly authored opinions in the current layer
  • E (Relocates): Enable non-destructive namespace remapping of composition source content via layer metadata path mappings. This allows reparenting and renaming without modifying original assets.

2. Content Accumulation

  • I (Inherits): Propagate opinions from base prims, using automatic upstream layer stack propagation
  • R (References): Bring in an external layer stack and add its opinions beneath the referencing prim in the composed stage. For example, if the defaultPrim in /Asset, when referenced at a particular prim results in the children of the defaultPrim composing at that prim. If the default prim is /Asset/MyRoot, which has the child opinion C/D under it, a composition under /A/B would result in /A/B/C/D.
  • P (Payloads): Work like references, but their content is only loaded when requested. They keep scenes lighter by deferring heavy assets until needed.
  • V (Variants): Apply conditional selection among authored alternatives using deferred evaluation

3. Global Hierarchy

  • S (Specializes): Create globally weak (base) opinions that can be overriden by all other composition arc opinions

The composition algorithm proceeds according to the LIVERPS strength ordering, with local opinions providing the strongest foundation. Inherits operate early in the process, propagating base prim opinions into derived prims. Variants require deferred evaluation—they are resolved only after all other composition arcs so they can make selections based on the fully composed opinion set. References and Payloads gather external layer stacks and with payloads optionally unloaded for scalability. Specializes establish globally weak opinion relationships where all specialized opinions are subordinate to their specializing counterparts. During the application of individual arcs, Relocates applies structural preprocessing, transforming namespace topology before opinion composition. The LIVERPS strength ordering thus ensures predictable resolution of conflicting opinions throughout the resolution process.

Local Foundation

Local — the L in LIVERPS — refers to “local” opinions from the current layer stack, which are stronger than all “remote” opinions introduced via composition arcs authored in that layer stack.

Layer Stacks establish the foundational substrate for all composition. Every layer stack begins with a “root layer” as the strongest layer. Layer stacks form ordered lists where earlier layers are stronger and later layers are weaker, creating the fundamental strength ordering system that governs how all other arcs resolve their contributions.

Sublayers extend the layer mechanism by allowing one layer to include other layers in its subLayers field. These sublayers are added in strong-to-weak order, where the first sublayer listed is strongest and the last sublayer is weakest. Each sublayer becomes weaker than the layer that included it, but stronger than any weaker sibling sublayers. Sublayers enable modular authoring workflows where complex stages are built from multiple files — a “shot” layer might sublayer separate files for lighting, animation, and modeling, allowing different departments to contribute opinions while maintaining predictable override behavior.

Sublayer Composition Safeguards: The composition algorithm includes error handling for malformed sublayers (which are ignored) and cycle detection (preventing infinite sublayer inclusion loops). Each sublayer may have associated time offsets that apply numerical scaling to all time values in that sublayer and its descendants.

Unlike composition arcs that operate at the prim level and create “remote” opinions, sublayers operate at the layer composition level, flattening additional authored content directly into the existing layer stack namespace to create “local” opinions with predictable strength relationships.

Relocates — the E in LIVERPS — enable non-destructive namespace transformation of prims from remote layer stacks. Unlike the composition arcs in the LIVERPS mnemonic that merge opinions for existing prims, relocates map prim paths from composition sources to new locations in the local namespace, allowing reparenting and renaming without modifying the original scene description.

Relocates address situations where direct editing would be destructive — affecting all other contexts that reference the same assets. By specifying path mappings in layer metadata (relocates = { </source/path> : </target/path> }), you can reorganize referenced content while preserving the source assets unchanged.

During composition, relocates integrate with the -IV-RPS strength ordering by providing additional namespace mappings that modify other arcs’ standard mappings. When a prim matches a relocates target path, USD composes opinions from the mapped source path, removes ancestral opinions (except variant arcs), and applies the namespace transformation to all descendant prims and their relationships. Relocates are resolved before references and payloads bring in external scene description, ensuring that the remapping applies to those arcs as well.

Relocates transform “where things are” rather than “what things contain,” enabling flexible asset organization patterns while maintaining composition arc semantics within the transformed namespace structure.

Content Accumulation

Inherits

Inherits — the I in LIVERPS — establish persistent relationships between “derived” prims and “base” prims, creating base/derived relationships similar to inheritance patterns in programming. Unlike references which bring in external layer stacks and assets, inherits work within the same layer stack to provide semantic layering by propagating opinions from base prims to derived prims.

Inherits enable mass editing by propagating reusable composed opinions from base prims to all derived prims. Multiple inherits can be stacked, allowing you to build “classes” of prims (e.g., a default Prop definition) and specialize them in many places. The derived prim receives all of the overrides applied to the base prim anywhere in the composition. This makes inherits ideal for establishing shared attributes and conventions across multiple prims within the same layer stack.

Use references and payloads for cross-asset structural composition (bringing in geometry and full subtrees from external files), and inherits for semantic layering within your existing layer stack structure. For example, define “all props should have collision geometry enabled” through inherits on base class prims, while the actual prop geometry comes through references from external asset files.

// Simple example:
class Prop {
    bool enableCollision = true  // Base prim opinion
}

over "MyChair" (
    inherits = </Prop>       // Inherits collision setting
    references = @chair.usd@ // References geometry
)

Note that inherit arcs introduce opinions during composition and therefore inherit arcs within the same layer can require individual recomputation of composed opinions even when pulling from the same base prim. Accordingly, for efficiency, it’s good practice that inherit arcs pull from other layers.

VariantSets

VariantSets — the V in LIVERPS — enable conditional selection among alternate authored opinions within a prim. “Variant Sets” are collections of “variants”. Unlike references or payloads that introduce external content, variants switch between different versions of content already authored within the same prim.

Variants function as conditional branches in the composition process. When a variant set contains multiple options (e.g., “summer” and “winter” versions of a tree), the composition engine selects one branch based on the current variant selection, then proceeds as if only that branch had been authored.

Variants operate entirely on pre-authored content — all variant options must be written into the layer beforehand. This makes them ideal for authored alternatives like different materials, LOD levels, or seasonal variations of the same asset.

Variant evaluation is deferred until after all other composition arcs have been processed. This ensures that all sources of opinions are available when computing which variant should be selected, allowing the variant selection itself to be influenced by the complete composed opinion set. Once a selection is made, composition continues into the variant set; nothing within the selected variant can retroactively change a previous selection choice.

Variant branches can themselves contain other composition arcs. References, payloads, or sublayers authored within a variant branch only take effect when that variant is selected, while arcs outside the variant set apply unconditionally to whichever branch gets chosen.

def Xform "Tree" {
    variantSet "Season" = {
        "Summer" {
            def Sphere "Leaves" { float size = 1.0 }
        }
        "Winter" {
            def Sphere "Leaves" { float size = 0.2 }
        }
    }
}

References

References — the R in LIVERPS — are a key mechanism for bringing external content into a stage. A reference introduces a complete prim hierarchy from another layer or asset, creating a new subtree in the namespace while preserving the source hierarchy.

References may point to an entire layer (via its defaultPrim metadata) or directly to a specific prim path—such as /Asset/B/C. This makes them versatile for assembling complex scenes. A single prop, environment section, or character can be referenced multiple times at different points in the stage. Each reference is processed individually, with the composition algorithm executed separately for each reference.

References are unconditional: once authored, they always participate in stage composition. This makes references essential for scene construction workflows where external assets must be integrated into the composed stage.

def Xform "World" {
    def Scope "Props" {
        def Xform "Chair" (
            references = @chairAsset.usd@
        )
    }
}

Payloads

Payloads — the P in LIVERPS — are primarily a scalability mechanism for managing heavy or bulk data that doesn’t always need to participate in composition. Unlike other composition arcs that are always active, payloads enable on-demand loading of content like dense geometry, simulation caches, or texture atlases.

Payloads provide conditional content accumulation. While references always bring in their target content, payloads maintain structural presence in the namespace but can selectively load or unload their opinion contributions based on workflow needs. This makes them ideal for scenarios where you need the option to work with lightweight proxies during editing but access full-resolution data for final rendering.

def Xform "World" {
    def Xform "Environment" (
        payload = @heavyForestAsset.usd@
    )
}

Payloads support scalable production pipelines where different stages require different levels of detail. Animators might work with unloaded payloads (seeing only proxy geometry), while lighting artists load the full payload data for accurate rendering.

Global Hierarchy

Specializes

Specializes — the S in LIVERPS — provide a mechanism for establishing base class relationships, similar to class inheritance in object-oriented programming. Specializes create “is-a” relationships where a prim can specialize another prim, inheriting its base characteristics.

Specializes arcs are weaker than inherits arcs in the composition strength ordering, meaning inherited opinions will override specialized ones when they conflict. This creates a clear hierarchy: specializes establish foundational base classes, while inherits provide more targeted opinion injection that can override those base definitions.

Specializes are particularly useful for creating foundational class hierarchies. For example, you might have a base “Vehicle” class that defines common attributes, and specific vehicle prims specialize from this base. Then, inherits can be used to apply more specific overrides or modifications to individual vehicles while maintaining the specialized base relationship.

The strength relationship Inherits > Specializes ensures that more targeted inherited opinions take precedence over general specialized base classes, providing a predictable override mechanism where specific modifications can fine-tune general base definitions.

LIVERPS Composition Strength Ordering

USD’s composition system resolves opinion conflicts through the LIVERPS strength ordering, where each composition arc type has a specific position in the hierarchy. LIVERPS governs the strength ordering of composition arcs within a single LayerStack. Across layers, the usual sublayer strength rules apply: the strongest opinions come from the last sublayer in the stack.

It helps to think of sublayer strength as the outer frame, and LIVERPS as the inner ordering within each frame.

  1. Local - Direct Specification — Opinions authored directly in the current layer stack
    • Establishes the base prim definition and local overrides
    • Local opinions always win
  2. Inherits - Class based composition — Copies primspecs from inherited class prims into the target prim
    • Adds properties, metadata, and composition arcs from base classes
    • Strong early binding, inherit opinions become part of the inheriting prim’s definition
  3. VariantSets — Conditional Composition
    • Selects and incorporates primspecs based on variant selection state
    • Conditionally adds/modifies primspecs and properties based on variant choices
    • Applied after inherits, allowing variants to override inherited data
  4. rElocates — Namespace topology transformation
    • Relocates modify namespace topology during the composition of each arc
    • Restructures prim hierarchy without changing original layer content
    • Relocates are not a composition arc per se
  5. References — Asset inclusion
    • Incorporates primspecs from external layers
    • Referenced content can be overridden by higher-strength operations
  6. Payloads — Deferred loading
    • Conditionally loaded external content for memory management
  7. Specializes — Weakest opinions
    • Base class relationships with globally weak opinions
    • Incorporates primspecs from specialized base prims
    • Provides fallback values and definitions that can be easily overridden

This strength ordering ensures predictable composition behavior where local authoring takes precedence, inheritance relationships work intuitively, and structural transformations occur at the appropriate processing stage.

Notable Arc Relationships

  • Local > All Arcs: Direct authoring always wins over composed content
  • Inherits > Specializes: Targeted inheritance overrides foundational base classes, enabling specific modifications to fine-tune general base definitions
  • Variants > External Content: Variant selections can meaningfully modify content brought in by references and payloads
  • Preprocessing Sequence: Relocates transform the namespace topology before other arcs operate on the transformed space

Special Behaviors

  • Specializes as a Base Class: If prim A specializes prim B, then ALL opinions for prim A are stronger than ALL opinions for prim B, creating hierarchical inheritance.
  • Variant Deferred Evaluation: Variants are processed after all other arcs to ensure complete opinion availability for selection computation.

Property Resolution Within LIVERPS

For each property encountered during LIVERPS processing:

Value Resolution

  1. Opinion Gathering: Collect all property values from contributing primspecs in LIVERPS order
  2. Strength Ordering: Apply composition strength rules to determine winning opinion
  3. Value Composition: For array properties, apply list editing operations (prepend, append, delete)

Metadata Propagation

  • Property Metadata: Custom metadata, interpolation settings, connection validity
  • Composition Metadata: Time samples, value blocking, sparsity patterns

PrimVar Handling

  • Inheritance: PrimVars can be inherited down namespace hierarchy
  • Interpolation: Vertex, face-varying, constant, and uniform interpolation modes
  • Element Size: Array dimensions for structured data types

Composition Cache and Performance

The LIVERPS algorithm employs aggressive caching:

  • PrimIndex Cache: Memoizes composition results for each prim path
  • Layer Offset Application: Cached time transformations for referenced/payload content
  • Property Index Cache: Caches property resolution chains to avoid repeated traversal

This layered caching strategy enables high performance composition of complex scenes while maintaining deterministic composition semantics across the LIVERPS strength ordering.