skilld
All skills
nvidia avatar

/physical-ai-video-data-augmentation

@0482ebc
by NVIDIA Corporationnvidia/skills3.5k stars
424

Use when running video data augmentation and auto-labeling workflows on OSMO: flow selection, preflight, submit-time interpolation, monitoring, and output retrieval. Trigger keywords: video data augmentation, data enrichment, auto labeling, VDA demo, OSMO workflow, pseudo labeling.

Ask your Agent

Use this Skill: https://skilld.dev/gh/nvidia/skills/physical-ai-video-data-augmentation

This session only. Nothing lands on disk.

assetscookbooksTUNING_GUIDE.md

≈520 tokens on demand. Your agent reads this file only when SKILL.md points to it.

Shared VDA Tuning Guide

This guide centralizes common parameter behavior used across cookbook scenes. Keep scene READMEs focused on deltas that are unique to that scenario.

Augmentation (workflow_config.yaml)

  • n_augmentations: number of augmented outputs per source video.
  • Variable weights: rebalance toward underrepresented conditions; each variable distribution must sum to 1.0.

Augmentation (augmentation/augmentation.yaml)

  • cosmos.parameters.sigma: larger values increase appearance drift from source.
  • cosmos.parameters.num_steps: larger values increase quality and runtime.
  • cosmos.parameters.guidance: larger values enforce prompt intent more strongly.
  • video_captioning.parameters.fps: raise for fast motion, lower for static scenes.
  • video_captioning.parameters.max_tokens: raise for visually dense scenes.
  • video_captioning.parameters.temperature: lower for deterministic captions.
  • pipeline.retry: retries for the full augmentation chain.
  • template_generation.parameters.retry: retries for template extraction only.
  • template_generation.parameters.retry_policy: strategy for retry behavior.
  • hallucination_check.threshold: stricter checks at lower values.

Auto-labeling (auto_labeling/auto_labeling_config.yaml)

  • detection_and_tracking.classes: keep only classes relevant to the scene.
  • detection_and_tracking.threshold: tune precision vs. recall trade-off.
  • detection_and_tracking.max_age: track persistence through occlusion.
  • vlm_json.frame_fps: analysis temporal granularity.
  • vlm_json.resolution: quality vs. token cost trade-off.
  • vlm_json.max_tokens: event-output budget.
  • vlm_json.timeout: endpoint timeout window.
  • mcq_generation.window_metadata_extraction.{vlm_max_tokens,llm_max_tokens}: MCQ extraction token budgets.
  • mcq_generation.window_metadata_extraction.window_frames: per-window span.
  • mcq_generation.window_metadata_extraction.sampling_fps: keep aligned with vlm_json.frame_fps.
  • super_resolution.enabled: enable only when fine detail is needed.

Source: SKILL.md on GitHub

Third-party checks

2 warnings3mo3 checks · Risk SAFE
  • Gen Agent Trust Hub3mo

    The skill is a workflow orchestrator for video data augmentation and auto-labeling on the NVIDIA OSMO platform. It manages the end-to-end pipeline, including credential verification, configuration generation, worker execution, and result retrieval. No security issues or malicious patterns were detected; all external dependencies and network operations are associated with trusted vendors and the skill's primary functionality.

  • Socket3mo

    3 alerts: gptAnomaly

  • Snyk3mo

    Risk: MEDIUM · 1 issue

Provenance

Signed by skilld at 0482ebc. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub yesterday.

Activeupdated 4 months ago

Capability

Other metadata
metadata
{
  "owner": "NVIDIA",
  "service": "data",
  "version": "1.0.0",
  "reviewed": "2026-05-26",
  "author": "NVIDIA",
  "tags": [
    "physical-ai",
    "video-data-augmentation",
    "auto-labeling",
    "cosmos"
  ]
}

README badge

README badge for nvidia/skills/physical-ai-video-data-augmentation