Synthetic Data
Generate photorealistic training data from simulation using NVIDIA Cosmos world foundation models. This section covers the SDG pipeline architecture, Cosmos model integration, and workflow submission.
SDG Pipeline
The synthetic data generation pipeline chains three Cosmos capabilities:
| Stage | Model | Description |
|---|---|---|
| Transfer | Cosmos Transfer 2.5 | Convert simulation renders to photorealistic images |
| Predict | Cosmos Predict 2.5 | Generate plausible future frame sequences |
| Reason | Cosmos Reason 2 | Assess and curate data for training quality |
🏗️ Architecture
synthetic-data/
├── workflows/ # OSMO and AzureML SDG job definitions
│ ├── osmo/ # OSMO workflow YAML (Jinja templates)
│ └── azureml/ # AzureML job YAML (commandJob schema)
├── cosmos/ # Cosmos model integration
│ ├── transfer/ # Cosmos Transfer 2.5
│ ├── predict/ # Cosmos Predict 2.5
│ ├── reason/ # Cosmos Reason 2
│ └── configs/ # Model configuration templates
├── examples/ # Pipeline examples
└── specifications/ # Domain specifications