Dataset output

Dataset stores a name, creation timestamp, and ordered frame records. Each frame contains its timeline index as a string, a list of pass-path mappings, and a list of custom attributes.

{
  "name": "My BlendGen dataset",
  "date": "07/17/2026,12:00:00",
  "data": [
    {
      "index": "1",
      "passes": [
        { "ColorPass": "/data/output/color/Image0001.png" },
        { "DepthPass": "/data/output/depth/Image0001.exr" }
      ],
      "attributes": [
        { "camera_id": "Camera" }
      ]
    }
  ]
}

The exact pass keys come from each pass class's type. Treat recorded file paths as the output contract: a successful manifest entry must point to an existing individual pass file.

JSON or NPY

Use DatasetOutputType.JSON for portable, inspectable metadata. DatasetOutputType.NPY serializes the dataset object with NumPy and is Python/NumPy oriented. Image pass files remain separate in both cases.

Attributes-only runs

Renderer(render_images=False) allows a session to iterate frames and save annotations without writing image passes. This is useful only when the intended dataset is genuinely metadata-only; it must not be used to mask missing requested images.