What does this skill do, and when should you use it?
This NVIDIA skill orchestrates People Attribute Search workflows for the Physical AI data factory on OSMO. It handles flow selection, credential preflight, submit-time interpolation, monitoring, and output retrieval. The pipeline creates controlled clothing or appearance variations for person crops and generates attribute captions and search queries. It is best suited to teams that already operate OSMO with GPU capacity and the required inference endpoints.
Collects dataset, flow, platform, endpoint, and augmentation inputs; runs credential and readiness checks; selects e2e, augmentation, or auto_labeling workflows; submits OSMO jobs; uses the paidf-augmentation container for image-edit augmentation and MCQ verification; uses the paidf-auto-labeling container for person-attribute captioning; monitors workflow status and task logs; and lists or downloads outputs such as augmented images, captions, verification metadata, and structured datasets.
- A data engineer needs controlled clothing variations for person-crop data while generating attribute captions.
- A computer-vision team wants image-domain augmentation only, without captioning.
- An annotation team has pre-augmented person crops and needs batch-generated attribute captions or search queries.
- A Physical AI team needs a monitored and reproducible person-data workflow running through OSMO.
How do you install this skill?
- The skill handles person images, clothing, and attributes; production use should add privacy, authorization, retention, and access-control requirements.
- The default policy may automatically deploy cluster NIMs and submit long-running GPU-consuming workflows that write to object storage; user confirmation, cancellation, cleanup, and rollback should be specified.
- Scripts, NIM reference files, and independently reviewable execution artifacts are not supplied; the BENCHMARK.md claims alone should not be treated as proof of reproducibility.
- Chinese support and mainland-China network reachability are unspecified; external model endpoints, Hugging Face, OSMO, and image/storage services may block deployment.
- Shell / CLI
- Network access
- Local filesystem
OSMO CLIbashjqGPU poolNVIDIA NIM endpointsHugging Face token
Install the individual skill with the NVIDIA skills CLI: npx skills add nvidia/skills --skill physical-ai-people-attribute-search --yes. The supplied README does not document a separate manual installation procedure.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/physical-ai-people-attribute-search" ~/.claude/skills/
rm -rf "$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:
- Run PAS on OSMO to augment person crops and generate attribute captions.
After installing it in an Agent Skills-compatible client, use a prompt such as “Run PAS on OSMO to augment person crops and generate attribute captions.” Provide a person-crop dataset organized as <person_id>/<view>.jpg, plus OSMO CLI access, Hugging Face credentials, an available GPU pool, and the required inference endpoints. Choose e2e for augmentation plus labeling, augmentation for image editing only, or auto_labeling for labeling only. Consult the --help output of the scripts for exact arguments.
What are this skill's strengths and limitations?
- Covers the full PAS path from preflight through submission, monitoring, and retrieval.
- Supports full, augmentation-only, and labeling-only workflows.
- Requires storage_url to be derived from the active dataset backend instead of silently using a stale default.
- Documents commands for workflow status, task logs, data listing, and downloads.
- Requires OSMO, GPU capacity, credentials, and reachable inference endpoints; it is not a standalone local utility.
- No built-in demo dataset is provided in the source material.
- Runtime can be substantial: an e2e run for 100 images with three augmentations typically takes about 6–8 hours.
- The supplied sources provide no test-suite details or evidence of production success rates.
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 |
|---|---|---|---|---|
| PAS People Attribute Search Workflow this page ✓ NVIDIA · Official | 49 · Use with care | ★ 3.8k | today | Apache-2.0 |
| TAO Image Grounding ✓ NVIDIA · Official | 52 · Use with care | ★ 3.8k | today | Apache-2.0 |
| NVIDIA Physical AI Defect Image Generation ✓ NVIDIA · Official | 55 · Use with care | ★ 3.8k | today | Apache-2.0 |
| TAO AOI Image Mining ✓ NVIDIA · Official | 45 · Not recommended | ★ 3.8k | today | Apache-2.0 |
| Entity Registry | 60 · Recommended | ★ 3k | today | Apache-2.0 |
How did FollowSkills review this skill?
The skill requires credential preflight, prohibits secret disclosure, derives storage_url from the active backend, and limits external endpoints to explicit use. However, it processes person images and attributes without clear privacy, retention, least-privilege, final-confirmation, or rollback controls; automatic NIM deployment and output download effects are not fully disclosed, so points are deducted.
The workflow instructions, three YAML files, parameters, and monitoring commands are broadly consistent, and missing credentials, GPU capacity, or endpoints produce defined USER_INPUT_REQUIRED cases. But the scripts, NIM configuration, and referenced troubleshooting materials are absent, so key paths cannot be reproduced statically; abnormal-input, cancellation, and recovery behavior remain thin. The score stays within the static ceiling.
The name, triggers, input format, three flows, and the non-fit case of container-internal tuning are reasonably clear, and the skill requires user-provided data rather than demo assets. Chinese-language interaction is not addressed, and core operation depends on OSMO, cluster GPUs, NIM services, Hugging Face, and object storage without evidence of mainland-China reachability; boundary and false-trigger evidence is limited, so points are deducted.
The documentation covers purpose, prerequisites, flow selection, parameters, timing, monitoring, outputs, and cookbook structure. It declares CC-BY-4.0 AND Apache-2.0 and includes an NVIDIA owner, version, and review date. However, the validator reports a missing recommended Examples section and an invalid author format, while a clear changelog, maintenance contact, and update path are absent, so points are deducted.
The skill directly addresses PAS augmentation, auto-labeling, submission, monitoring, and output retrieval. The included evaluation report claims an overall pass and reports 95% skill-assisted effectiveness for Codex. Those results lack independently verifiable artifacts, while operation requires substantial external infrastructure and model-dependent outputs that still need review, so the static score is conservatively limited and reduced.
A pinned revision scope, complete workflow YAML, configuration files, evaluation tasks, and benchmark report provide auditable material. However, there is no committed test suite, CI coverage for key paths, execution logs, or independently reviewable artifacts; the benchmark figures are primarily report claims and cannot be independently reproduced from the supplied files, so only limited credit is given.
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 →