Omniverse Usd Performance Tuning

Classify and author assets

Classify selected assets children before parents, after their workload-required independent use has justified a boundary under structure assessment. Having child assets and no owned geometry determines aggregate structure; it does not justify making an organizational prim a separate asset. These workflow conventions use the terminology of annotated asset structures:

Every geometry occurrence belongs to one atomic asset or package. An aggregate's referenced descendants are expected; unassigned geometry is not. An atomic asset may use internal geometry references and shared material libraries.

Model hierarchy and occurrence opinions

Choose kind for the composed role, separately from structural type. Equipment can be a component even when its structure is aggregate. Significant parts inside a component are subcomponent occurrences; their standalone definitions can remain components. Organizational ancestors of models need group or a derived kind such as assembly, without becoming separate published assets. Authoring clears obsolete copied kind metadata on copied internal prims, including protected coordinate frames without geometry, then applies the selected asset and interface kinds. Preserving a prim and its transforms does not require preserving its imported model kind. Do not turn every imported model label into subcomponent. Selected child assets and referenced contents keep their declared roles; source layers remain unchanged.

Keep editable mechanisms in the equipment hierarchy, with references to atomic geometry below their interfaces. Preserve occurrence placement, visibility, authored animation and domain relationships. Shared geometry must not turn a required editable descendant into an instance proxy. Instance eligible atomic asset occurrences even when the current assembly uses a definition only once. Placement remains editable on the occurrence root; parents with independently editable children remain non-instanceable.

Imported references and instances describe source composition, not required asset boundaries. Preparation resolves selected interiors into editable definitions while preserving selected child assets and required internal prims. Assessment merge groups are provisional; preparation recomputes them on those editable inputs before optimization. Accepted product identities are checked against supplied libraries before preparation. Confirmed parent-child groups reuse published definitions directly. Explicit variant contexts are applied during ordinary input preparation. Other definitions continue through preparation and the existing reuse checks. The workflow records how many worker inputs were avoided.

The authoring and optimization tools handle transform composition and placement. They may bake static internal transforms into merged geometry; those original transform attributes need not survive. Retain transform operations and their order where authored animation or a required editable attribute depends on them, including reset-stack behavior and authored time samples. The agent records the required movement/edit paths, not a transform-preservation checklist.

Record each edit's existing prim/property path, human-readable use and edit_class (placement, visibility or identification). Placement and visibility require the corresponding USD capability and an editable output prim. Identification preserves the named prim/property for its selection or lookup use and existing identifying opinions; it imposes no transform-frame or editability requirement. Occurrence labels can differ while sharing geometry. Geometry beneath a required prim can merge within it. A Mesh used as a placement frame keeps that frame; fitting remains available beneath a preserved Xform. Preserve empty attachment frames without adding geometry. Detailed pose/visibility trials and unrelated-occurrence isolation checks belong in development fixtures, not repeated whole-scene runtime trials. Unsupported required interfaces remain incomplete.

Shared materials and semantic libraries

Resolve effective material bindings in occurrence context, including GeomSubsets, inherited bindings and binding purposes. Keep shader connections, primvars and resources intact. The workflow creates the shared material library automatically before geometry workers run: it extracts used networks and their shader dependencies, removes unused networks, and invokes native optimizeMaterials for compatible duplicates. Asset definitions reference the resulting material prims; do not create separate copies of their shader networks for each asset. Distinct material properties and required metadata remain distinct. One writer coordinates this work.

Publication remaps those references to the delivered library. The published manifest's output.material_library and the report identify its final location. Published material-library filenames use materials_ followed by the existing unique publication suffix. Preserve source material prim names and displayName labels when present. A CAD colour label such as Color_D2D2FF identifies colour; it does not establish a physical material such as steel or plastic. Keep those labels when colour is the only supplied identification. Material names and labels do not determine equivalence; native material optimization compares the networks. Check the reported location when inspecting the delivery, rather than the temporary working/materials.usdc file. Resource localization follows the persistent-library requirements; this does not imply independent lookup of equivalent materials across unrelated scenes.

Publish multiple separately referenceable definitions in a library for each established semantic category: robots, fences, pumps, chairs and similar types. Use the general library when equipment type cannot be established. Optional equipment_family labels group supported product/model relationships within a category; omit them when evidence is absent. They do not change library partitioning, entrypoints or geometry-equivalence decisions. Category, structural type and kind are independent fields. Distinct models in a category retain distinct entrypoints. Keep the consumer assembly separate from category libraries.

For persistent reuse, follow the library index requirements. Confirm exact cache hits before optimization across the supplied indexes, independent of category. The lookup helper returns nominations requiring confirmation; equal names or bounds do not authorize reuse. Keep inherited entries read-only and preserve their local coordinate frames. Changed geometry creates a new output version instead of silently changing other consumers. Use remote access for resolver and localization work.

Working layers and publication

Express geometry in the selected asset's local coordinate frame and retain its occurrence transform. Replace the original rendered subtree with its reference; do not leave a second rendered copy. Sdf.CopySpec copies authored specs, so the authoring tool must first resolve the selected composition. Give each definition units, axis and a usable default prim.

Author an immutable input and separate output for each unique definition. Parent working layers reference immutable child inputs, so a worker never reads another worker's partially written output. These temporary files do not determine delivery packaging. Record output entrypoints before dispatch.

Category planning separates independent computation from publication. After native execution and output checks pass, publication rewrites references to verified outputs and writes each category through one writer. Within that unpublished layer, the native sharing helper can share identical remaining geometry across definitions with different child poses. It resolves explicit owned occurrence paths to editable meshes or existing sources beneath each asset's Prototypes class, submitting each authored source once. Native deduplicateGeometry and scoped organizePrototypes place confirmed sources under the category's /Library/Prototypes class. The publisher updates occurrence path mappings and the written library fingerprints. Editable ancestors, addressed mesh interfaces, material bindings and independently referenced entrypoints remain intact. These internal geometry references do not introduce asset boundaries. This pass compares eligible whole meshes within one writable category layer. It does not recover repeated parts inside a merged mesh or share geometry across category libraries. Inherited entries remain read-only. For a common library, publish unique category files under assets/, materials under materials/, and supporting layers under dependencies/. Preserve existing files and atomically replace the cumulative index.json after new files are installed. Keep each station’s after.usdc, reports and working files in its output directory. References are relative to final file locations. A failed publication leaves the previous index usable; unindexed new files may remain. Standalone publication commits its output directory together. Reopen the consumer and check references, binding remaps, placement and ownership; a successful copy alone does not establish valid composition.

For portable delivery, follow the library dependency requirements. Move a successfully published dependency snapshot and reopen it to check references, materials and resources. Otherwise retain resolvable persistent references and report them.