Your first dataset

Create dataset.py beside a .blend file. This example produces display color plus lossless metric and semantic channels:

from pathlib import Path

import bpy

from blendgen.dataset import Dataset, DatasetOutputType
from blendgen.passes.alpha import AlphaPass
from blendgen.passes.base import ImageOutputType
from blendgen.passes.color import ColorPass
from blendgen.passes.depth import DepthPass
from blendgen.passes.index import RawMaterialIndexPass
from blendgen.passes.normal import NormalPass
from blendgen.passes.opticalflow import OpticalFlowPass
from blendgen.renderer import Background, Renderer
from blendgen.renderers import CyclesBackend, CyclesDevice
from blendgen.session import Session

OUTPUT = "/data/output/"

if bpy.context.scene.camera is None:
    raise RuntimeError("The scene has no active camera")
Path(OUTPUT).mkdir(parents=True, exist_ok=True)

passes = [
    ColorPass(prefix="color", output_type=ImageOutputType.PNG),
    AlphaPass(prefix="alpha", output_type=ImageOutputType.PNG),
    DepthPass(prefix="depth", output_type=ImageOutputType.EXR),
    NormalPass(prefix="normal", output_type=ImageOutputType.EXR),
    OpticalFlowPass(prefix="flow", output_type=ImageOutputType.EXR),
    RawMaterialIndexPass(prefix="material_index", output_type=ImageOutputType.EXR),
]

renderer = Renderer(
    backend=CyclesBackend(samples=8, device=CyclesDevice.GPU),
    background=Background.ALPHA,
    resolution_x=640,
    resolution_y=360,
    output_base_path=OUTPUT,
    passes=passes,
)

dataset = Dataset(
    prefix="",
    dataset_name="My BlendGen dataset",
    output_dir=OUTPUT,
    filename="blendgen_dataset",
    output_type=DatasetOutputType.JSON,
)

session = Session(
    renderer=renderer,
    dataset=dataset,
    output_dir=OUTPUT,
    frame_start=1,
    frame_length=1,
)

print(session.info)
session.run()

Run it against your scene:

docker run --rm --gpus all \
  -v "$PWD":/data -w /data blendgen:latest \
  blender --background --python-exit-code 1 /data/scene.blend \
  --python /data/dataset.py

For CPU or Eevee, remove --gpus all. For a CPU Cycles run, change the dataset file to CyclesDevice.CPU.

Expected output

output/
├── alpha/Image0001.png
├── color/Image0001.png
├── depth/Image0001.exr
├── flow/Image0001.exr
├── material_index/Image0001.exr
├── normal/Image0001.exr
└── blendgen_dataset.json

The exact ordering is unimportant. Every path recorded in blendgen_dataset.json must point to a real file.

Expected output

On this page