Omniverse Usd Performance Tuning

Discover asset boundaries

Use the intended workload to decide what must remain independently placeable, replaceable, configurable or movable. Factory-scale use is the default when no other workload is specified. The factory structuring guide provides the equipment, assembly and editable-interface model.

Inspect and select asset boundaries

Open the selected configuration with payloads loaded. Include hidden geometry; visibility can change with animation. Record missing dependencies as incomplete inspection. Do not enumerate alternate variants. Preserve the source layers and the variant selections recorded at the start of the run.

Tools summarize relevant prims children before parents, including referenced and instanced interiors. Reuse cached summaries for shared definitions while retaining occurrence context. Use schema, topology and valid bounds to narrow expensive comparisons. Missing bounds are unknown, not evidence that geometry is absent.

Boundary decision checklist

An asset boundary selects a prim whose assigned content will be extracted into a separately referenceable asset definition. Set disposition: "retain" to select that boundary; this does not preserve its internal hierarchy unchanged. Set disposition: "internal" to include the region within its enclosing asset. Preserving an editable prim or property does not itself select an asset boundary.

Use the candidate summary to answer this checklist. Record one boundary decision and concise reason per reviewed subtree or group; do not write an explanation for each mesh. Missing optional metadata does not prevent a decision.

Review parents before children. Select the complete equipment, meaningful service-system section or architectural element with retain. Use decision_scope: "prim" when internal child boundaries remain undecided and review them next; use subtree when interior review is complete. Record required internal edits in the same decision. If children were reviewed first, use evidence.enclosing_boundaries to select the complete equipment root before completing assessment. Select that root using the equipment evidence in the checklist and its required use as a unit; enclosing organizational Xforms do not become assets merely because they contain selected children. An asset without child assets is atomic. Atomic means no child assets, not one mesh or one physical part. Use internal/prim to continue through an organizational prim when a meaningful asset boundary has not yet been found. Do not automatically descend to individually named parts. Tools classify child assets before parents and handle placement, transform baking and geometry comparisons.

Interpret German and English CAD labels, including compounds and umlaut spellings (Tuer, Schraenke), using assembly context. Keep original product identifiers, dimensions and handedness significant; manufacturer and copy labels nominate related equipment, not identical geometry.

Combine labels with hierarchy context: opaque children beneath a product-labelled ancestor may be ordinary internal geometry; related product-labelled siblings may be separate occurrences. A level or discipline label does not establish equipment. Similar dimensions nominate repetition but do not prove matching geometry. Compare asset-local dimensions with the same units, purposes and USD time code, separately from world-space position. Rotated copies can have different world-axis-aligned bounds. Missing bounds mean unknown; do not demand a geometry scan to fill every row.

In candidates.json, evidence.product_review_groups batches members with shared product IDs, authored labels or alphanumeric label tokens. Page it with successor_assessment.py show --workspace OUTPUT --candidate ID --field evidence.product_review_groups. Use each group’s source_prim_paths for a common decision when its member evidence supports the same conclusion. Groups may overlap; submit each prim once and split exceptions into separate decisions. Interpret station and occurrence tokens from the observed naming convention while preserving model, dimensions, cut length, revision and handedness. Scripts keep complete labels and do not strip suffixes. evidence.members supplies the member context and children; repetition_nominations and confirmed_reuse distinguish candidates from confirmed matches. Schema-based review groups are explicitly marked review_batch_only_not_equivalence. Read each member's labels and exceptions; a representative's counts or children do not describe every member. Early reuse nominations are not confirmed matches. Fingerprints validate evidence freshness; they are not explanations of duplication. Request specific property values or spatial facts only when they could change a decision. Visual inspection is not required or available to the target agent.

For example, a factory-visualization workload could use this hierarchy. Labels describe structural type, not USD model kind:

Building - aggregate
└─ Level 01 - aggregate, if independently loaded or delivered
   ├─ Chiller - package
   │  ├─ Pump assembly - atomic
   │  └─ Internal geometry
   ├─ Complete robot - aggregate
   │  ├─ Base link - atomic
   │  └─ Moving links - atomic
   ├─ Piping section - atomic
   │  └─ Static pipe and fitting geometry
   ├─ Electrical conduit system - aggregate
   │  ├─ South service zone - atomic
   │  └─ North service zone - atomic
   ├─ Wall section - atomic
   ├─ Floor section - atomic
   └─ Ceiling section - atomic

A complete robot with geometry outside its selected child assets is a package.

Use existing meaningful service-system and architectural sections for factory visualization. Static pipes and fittings within a section are internal geometry unless the workload establishes a separate asset requirement. If a section also contains a valve selected as a child asset for required independent configuration, that section becomes a package. Apply the same criteria to conduit and duct sections. Do not invent partitions or regroup adequate subtrees to match the example. A level, discipline or product-family label alone does not establish an asset boundary; assess what its subtree represents in the workload.

For each selected asset boundary, record its independent use, supporting prim or property paths and a concise reason. Distinguish observed facts, user requirements and domain inference. Account for each service-system section and architectural element under its meaningful owner; do not collect unrelated infrastructure into one package merely because it shares a building ancestor.

Tools collect authored animation and addressed relationship/connection targets. Review unresolved semantic questions using the identified parts and their assembly context; record additional requirements by exact prim/property path and use. Use the boundary checklist to select mechanical bodies, including static poses. Authored animation remains preserved by the existing tools regardless of additional editability decisions. An internal attachment frame or required identification prim remains within its asset; geometry beneath it may be optimized while that prim stays addressable. Record an unresolved requirement when a required use cannot be satisfied. Asset authoring owns structural type, USD model kind and preservation behavior. Tools handle placement and animation preservation.

Reconstruct ownership

Keep the existing hierarchy when it already represents the chosen owners. Regrouping is not a prerequisite for optimization or library publication. Request reconstruction only when source organization prevents the intended geometry ownership:

For misplaced parts, inspect attachment relationships, assembly metadata and mounting context before spatial evidence. Bounds shortlist possible owners; proximity alone does not settle ownership. A fastener connecting two bodies has one geometry owner and can retain relationships to both bodies.

For regrouping, owner_prim_path names the physical owner. When that owner is missing, set create_owner: true in the reconstruction proposal to create one Xform immediately beneath an existing defined, editable parent. This does not create missing ancestor chains or overwrite a prim. Leave the flag absent for an existing owner. The tools preserve placement and inherited state, reject cycles and namespace collisions, then return the new owner for explicit boundary and structural classification. Creating an Xform alone does not establish an asset.

After submitting reconstruction decisions, run scripts/successor_reconstruction.py --workspace OUTPUT. The controller batches disjoint regroup requests into one authored output. Overlapping or dependent requests remain sequential; each request retains its outcome. It invokes native splitting or prim regrouping and returns a new assessment workspace. The regroup adapter calls regroup(source_layer, output_layer, memberships, create_owners=...), where memberships maps disjoint source prim roots to destination owner paths. It copies the local layer stack into an isolated output, retains references and payloads, and moves subtrees with Usd.NamespaceEditor; it does not flatten the composed stage. The result contains status, input/output paths, root_path_map and created_owner_paths. The controller translates the agent's proposals into these arguments; agents do not maintain descendant-by-descendant mappings. Continue in that returned workspace; the previous one is superseded. It retains the original source and ordered root-path mappings for profiling and remapping required interfaces. Regroup preservation is tested in development; execution does not produce or replay per-property preservation checks. Later phases check operation records and written composition as described in verification. Read the returned output paths and operation results. Do not assume native output names or use mesh-count reduction to judge a split. Disconnected pieces and material subsets do not automatically become separate assets.

In decision JSON, use operation: "splitMeshes" with parameters: {"splitOn": 0} for disconnected vertices, or parameters: {"splitOn": 1} for material GeomSubsets. The controller derives the prim paths from source_prim_paths. It operates on editable polygon meshes and preserves ancestor animation. Animated mesh data, subdivision surfaces, nested meshes and unresolved relationship remapping remain unsupported. Another native mode needs its own preservation checks; do not write a custom face-cutting algorithm. If the available modes cannot separate the intended parts, record the precise unsupported case after testing applicable options. Reinspect changed regions and remap required properties before proceeding. A failed reconstruction leaves its dependent work incomplete while independent regions can continue.

Confirm reuse

When persistent_indexes are supplied, inspect persistent_nominations beside candidate evidence. Each row names the source prim, matching product/model identifier and published layer/prim entrypoint. A child's nomination applies to that child, not its containing equipment. These are unconfirmed clues; retain independent boundaries and required edits. Record accepted product identity with exact source facts and a substitution reason when supported; otherwise preparation confirms published source identity regardless of category; see persistent reuse. If the workload explicitly permits small constant-colour rounding substitutions, set allow_constant_colour_rounding: true; preparation records the accepted difference and preserves occurrence placement. For source-content matching without that authorization, effective appearance must match exactly.

Authored clues nominate candidates. Accepted product identity authorizes representation substitution, not a claim of identical source geometry or corresponding internal mesh paths. Content-based reuse follows the existing source-identity checks and explicitly authorized enclosing-child normalization. See persistent reuse for confirmation requirements.

Resolve reuse of child assets selected by boundary decisions before merging the remaining interior geometry. Repeated meshes alone do not establish asset boundaries or exclude geometry from merging. Internal references and instances can remain inside an atomic asset.

Record decisions

Select the stage entry point as the root asset boundary. For a factory or building layout, record structural_type: "aggregate" with the layout reason: the root composes equipment and building assets without owning geometry. A spatial subset of a layout, such as one floor, roof or site zone, is still a layout. Assign its geometry to the physical owners selected above; do not change the layout root to a package just to absorb that geometry. An equipment-only source can instead have an atomic or package root according to its interior. Organizational children need no separate asset boundary. Record regrouping or splitting for misplaced geometry before marking its terminal wrapper internal.

Use internal for an organizational prim and retain for an asset. With decision_scope: "prim" (the default), its interior remains unresolved: review continues through its child groups. Use decision_scope: "subtree" with retain to select an asset and complete its interior review together, or with internal to complete a region inside its owner. Record the evidence and reason for that scope; do not use it to skip units whose independent use is still undecided. source_prim_paths can group disjoint subtrees with the same reviewed conclusion. Selected asset boundaries and protected interfaces still receive their own review. Required editable prims remain authored even when their boundary decision is internal. Terminal geometry wrappers require no separate mesh decisions.

Run scripts/successor_assessment.py inspect --source INPUT --workspace OUTPUT --request REQUEST_JSON. REQUEST_JSON is a JSON file containing the workload and operation policy. Record the workload and physical geometry budget, including any user override or explicit exact-geometry requirement. The tool selects the successor policy and writes the current source context, inventory, grouped boundary reviews, definitions, occurrences and unresolved decisions.

The CLI returns counts and artifact paths; complete evidence remains in candidates.json. List candidate groups with scripts/successor_assessment.py show --workspace OUTPUT --offset 0 --limit 10. Use show --workspace OUTPUT --candidate ID to list its fields, then --field evidence.all_members for its summary or --field evidence.members --offset 0 --limit 10 for members. Follow next_offset to inspect the remaining members; a page is not evidence for unreviewed members. Dot-separated fields can select nested evidence, including a member index.

Submit decisions with scripts/successor_assessment.py decide --workspace OUTPUT --decisions DECISIONS. Use decision-help.json to fill decision-template.json under decision-schema.json. Copy current source context and candidate fingerprints, then judge reviewed members; the template supplies no boundary answers. Submit a separate file to survive template regeneration. The adapter validates current evidence/configuration and preserves earlier decisions. During boundary review, unchanged source evidence is reused; current decisions and required edits are evaluated again. Progress reports the assessment pass, stage scan counts, and evidence reuse separately from boundary review. Tools own membership and generated paths. Agent records supply disposition, independent use, reason, evidence paths, semantic category or unknown, and required occurrence edits, geometry_representation for the selected operation order (omission uses merged_meshes), optional product_identity for accepted interchangeable products (key, reason, asset-root frame and exact evidence; format in decision-help.json), plus optional equipment_family when assigned using the checklist above. Unreviewed members stay pending. Source or configuration changes invalidate corresponding evidence.

Continue with asset authoring when ownership and interior review are complete. Classification cannot substitute for reviewing the contents.