What does this skill do, and when should you use it?
This is NVIDIA Omniverse’s top-level workflow skill for USD performance diagnosis and optimization. It covers slow loading, high memory use, low FPS, GPU crashes, conversion-quality triage, and validation failures. The workflow gates on runtime context, preserves evidence, and routes work through structure assessment, validation, restructuring, Scene Optimizer operations, profiling, and reporting. It fits teams that already have access to the relevant Omniverse or USD runtime components and need a controlled optimization process.
It collects scene and runtime context, requires setup-usd-performance-tuning to verify Kit, standalone, or partial runtime paths, and then plans and routes USD structure assessment, validation, restructuring, Scene Optimizer operations, baseline and after profiling, and optimization-report. During execution it may read logs, CSV files, and summary JSON while keeping raw artifacts on disk. When mutation is authorized, it produces an optimized USD stage plus structured JSON and Markdown optimization reports.
- An Omniverse user has a USD scene with slow loading and needs composition assessment and baseline profiling first.
- A digital-twin team has low interactive FPS or excessive GPU memory use and wants scoped, evidence-based optimization iterations.
- An engineer has validation failures or conversion-quality issues after CAD-to-USD processing and needs the correct validation route.
- A technical artist needs to evaluate restructuring or Scene Optimizer changes before authorizing mutations.
How do you install this skill?
- This is a source-only static review; no runtime, script, authentication, or optimization workflow was executed.
- Key capabilities depend on multiple downstream skills and NVIDIA Omniverse runtimes whose contents are not supplied; missing dependencies may reduce the workflow to structural diagnosis or blocking.
- The BENCHMARK marks the overall result FAIL and reports PII findings, duplicated content, and missing Tier 3 evidence; these should be remediated and re-evaluated before publication.
- For Shell, Write, remote assets, and destructive Scene Optimizer operations, verify permissions, output paths, backups, and rollback procedures in the target environment.
- Shell / CLI
- Network access
- Local filesystem
NVIDIA Omniverse KitScene OptimizerAsset ValidatorUSD Pythonomniverse:// authentication
Install the skill from NVIDIA/skills with the skills CLI:
npx skills add nvidia/skills --skill omniverse-usd-performance-tuning --yes
You can target a client explicitly:
npx skills add nvidia/skills --skill omniverse-usd-performance-tuning --agent codex
The supplied README does not document the exact local installation directory for this individual skill.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/omniverse-usd-performance-tuning" ~/.claude/skills/
rm -rf "$tmp"Generated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- This USD loads slowly; triage what to check first.
- Route a low-FPS CAD scene through the performance workflow.
After installation, give the agent a concrete USD performance request such as “This USD loads slowly; triage what to check first.” or “Route a low-FPS CAD scene through the performance workflow.” Provide the scene path, problem type, workload type, runtime status, and whether in-place mutation or a separate output is allowed. omniverse:// assets require authentication before first opening.
What are this skill's strengths and limitations?
- Coordinates runtime gating, structure assessment, validation, profiling, optimization, and reporting in one workflow.
- Requires baseline evidence and validation before and after processor execution.
- Plans three scoped iterations by default for broad optimization requests while preserving milestone order.
- Supports diagnosis-only work and authorized generation of an optimized USD stage.
- It is an orchestration workflow, not a standalone USD performance profiler, and depends on downstream skills and Omniverse runtime components.
- SKILL.md does not provide installation instructions for Kit, Scene Optimizer, Asset Validator, or USD Python.
- Full Kit FPS, frame-time, Hydra, and RTX profiling requires the external NVIDIA/omniperf skills.
- It gives no fixed thresholds or guaranteed performance gains; results depend on scene evidence and runtime conditions.
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| Omniverse USD Performance Tuning this page ✓ NVIDIA · Official | 52 · Use with care | ★ 3.8k | today | Apache-2.0 |
| Omniverse Realtime Viewer ✓ NVIDIA · Official | 55 · Use with care | ★ 3.8k | today | Apache-2.0 |
| React Native Performance Guide | 60 · Recommended | ★ 1.7k | 20d ago | MIT |
| CAD to SimReady Workflow ✓ NVIDIA · Official | 49 · Use with care | ★ 3.8k | today | Apache-2.0 |
| Browser DevTools Testing | 59 · Recommended | ★ 112k | 15d ago | MIT |
This skill is an orchestration and routing workflow rather than a standalone profiler. Full Kit runtime profiling is delegated to NVIDIA/omniperf, while runtime probing and selection are owned by setup-usd-performance-tuning.
How did FollowSkills review this skill?
The skill specifies a runtime-context gate, remote-asset authentication, user approval for destructive operations, artifact retention, and protection against unauthorized in-place mutation; its authentication guidance also forbids requesting or persisting passwords and tokens. However, it declares Shell and Write access, writable output paths, and omniverse:// access across many delegated phases, without a complete skill-local permission boundary, unified rollback procedure, or full data-flow disclosure, so points are deducted.
Instruction ordering, runtime gates, blocked codes, failure routing, and report requirements are relatively specific. However, many critical behaviors are delegated to downstream references not supplied for this target, while BENCHMARK reports an overall FAIL and no Tier 3 results. Static review cannot reproduce key paths, so the score is capped below 10 and reduced for limited test and diagnostic-failure evidence.
Use cases, entry-skill ownership, local-versus-remote authentication boundaries, and the negative viewer-construction case are fairly clear. Chinese-language support, mainland-China network reachability, and variation across installation modes are not addressed sufficiently; core capability depends on Kit, Scene Optimizer, Asset Validator, USD Python, and runtime probing, so trigger and environment fit remain incomplete.
The skill provides versioning, an Apache-2.0 license, author metadata, tags, examples, prerequisites, limitations, troubleshooting, reference navigation, and structured-output requirements. It lacks a clear changelog, named maintenance responsibility, and update path. The supplied BENCHMARK also reports 10 static findings and 17 deduplication findings, indicating documentation maintenance noise and repetition.
The intended workflow, phase chain, baseline/after comparison, structured report, and neutral handling of successful no-op operations are clearly defined and could cover diagnosis through reporting. Actual usefulness depends on downstream skills, runtime availability, and external tooling; BENCHMARK is an overall FAIL with no usable Tier 3 evidence. Direct usability therefore remains unverified and the score is capped below 7.
The supplied evals.json, BENCHMARK static-validation results, referenced report schema, and artifact-retention rules provide limited auditable evidence. There is no available third-party execution evidence, Tier 3 result, committed test suite covering key paths, or independent reproduction in the supplied material, so the score is capped below 5.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →