Usd Scene Structure

USD Scene Structure Performance

How a partner structures their USD data determines loading speed, memory usage, and runtime performance. Authoring decisions compound across every asset in the scene.

Headless USD work? The USD Performance Tuning skill applies this authoring guidance directly to USD files. Hand off when the task is "fix the USD" rather than tune a running Kit app — see AGENTS.md → "When to hand off to the USD Performance Tuning skill."

Do / Don't

• ✅ DO use payloads for heavy geometry and shading on model assets. The single biggest fix for slow opens. Open with LoadNone to get a fast Model Hierarchy view; load payloads selectively.
• ✅ DO use binary .usdc for all data-heavy files. .usdc supports memory-mapped I/O so data is lazy-loaded; .usda cannot.
• ✅ DO keep layer counts in check — don't let them explode into the hundreds or thousands. Layers multiply across references; small per-asset counts compound across an aggregate scene. Do not publish a single threshold; the right number depends on workflow.
• ✅ DO mark referenced or payloaded repeated assets as instanceable=true. Critical caveat: this only helps when the same asset is referenced/payloaded multiple times. Setting instanceable=true on a duplicated Xform hierarchy where the geometry is itself duplicated provides no benefit.
• ✅ DO use UsdGeomPointInstancer for large counts of small repeated objects (bolts, fasteners, vegetation).
• ✅ DO pick the simplest composition arc that fits the purpose. Sublayers for workstream layering inside a published asset; references for asset aggregation; payloads for deferral. Do not publish a numeric cost ranking — cost depends on what the arc is doing in your stage.
• ❌ DON'T construct stages at runtime when you can pre-build them. Composition is expensive.
• ❌ DON'T use USDZ for runtime where load performance matters. USDZ front-loads everything and bypasses runtime caches. Packaging format only.
• ❌ DON'T use .usda for large data files.

Composition arcs — what to use when

LIVRPS strength ordering (strongest → weakest): Local > Inherits > VariantSets > References > Payloads > Specializes. This is opinion strength, not cost.

The payload pattern (recommend by default for model assets)

my_asset.usd                 ← Lightweight interface file (small)
  ├── defaultPrim metadata
  ├── AssetInfo
  ├── VariantSets (material choices, LOD levels)
  ├── Kind = "component"
  ├── Rest bounding box (for unloaded display)
  └── Payload → my_asset_geo.usd    ← Heavy content
                 ├── Geometry hierarchy
                 ├── Materials / shading networks
                 └── Texture references

Instancing strategy

The renderer-side flag is not the same as Isaac Lab environment instancing. It moves scene-graph instancing work into the renderer for the RTX backend.

Deduplicator caveat (so the user doesn't get surprised)

The Scene Optimizer deduplicator works at the mesh level — it deduplicates individual mesh data, not whole hierarchies. If a partner has many copies of the same hierarchy (e.g., a pallet with parts), the deduplicator collapses the per-pallet meshes but does not collapse the hierarchy itself.

For full-hierarchy reuse, recommend a small custom script that:

1. Identifies duplicate hierarchies (by display name from CAD or by hashing prim subtrees).
2. Picks one as the prototype.
3. Rewrites the others as internal references to that prototype.
4. Then runs the standard deduplicator on the prototype's meshes.

File formats

Convert with usdcat -o output.usdc input.usda.

Source guide section

docs/reference/usd-scene-structure.md.

External references: Maximizing USD Performance, Asset Structure Principles, Scenegraph Instancing, UsdGeomPointInstancer.