Synthetic Data

Synthetic Data Skill

Generate photorealistic training data using NVIDIA Cosmos world foundation models — Cosmos Transfer, Cosmos Predict, and Cosmos Reason.

Overview

The Synthetic Data domain provides SDG pipelines that transform simulation-rendered frames into photorealistic training data, predict future environment states, and curate output for training quality.

Pipeline Stages

| Stage | Model | Purpose |

|-------|-------|---------|

| Transfer | Cosmos Transfer 2.5 | Convert Isaac Sim renders to photorealistic images |

| Predict | Cosmos Predict 2.5 | Generate future frame sequences from observations |

| Reason | Cosmos Reason 2 | Assess data quality and filter training samples |

Workflow Submission

SDG workflows can be submitted via OSMO or AzureML:

• OSMO workflows: synthetic-data/workflows/osmo/
• AzureML jobs: synthetic-data/workflows/azureml/

Key Files

| File | Purpose |

|------|---------|

| synthetic-data/README.md | Domain overview and directory structure |

| synthetic-data/workflows/osmo/sdg-pipeline.yaml | End-to-end OSMO SDG pipeline |

| synthetic-data/cosmos/configs/README.md | Model configuration reference |

| synthetic-data/specifications/synthetic-data.specification.md | SDG pipeline specification |

| synthetic-data/specifications/cosmos-integration.specification.md | Cosmos model integration specification |

Environment Requirements

All NVIDIA Cosmos containers require:

| Variable | Value |

|----------|-------|

| ACCEPT_EULA | Y |

| PRIVACY_CONSENT | Y |

| NVIDIA_DRIVER_CAPABILITIES | all |

GPU Requirements

Each Cosmos model stage requires a minimum of 1x A100 (40 GB) GPU. H100 (80 GB) recommended for production workloads.