infinigen-1.19.0

Infinigen conda package recipe

Package contents

This directory holds a rattler-build recipe for Infinigen, a procedural 3D scene generator from Princeton Vision & Learning Lab. If you build this package, you can use it to generate photorealistic indoor and outdoor scenes with ground truth labels on AWS Deadline Cloud.

The package you build from this recipe will include the following:

1. Infinigen v1.19.0: the procedural scene generation engine.
2. bpy 4.2.0: Blender as a Python module for headless rendering via Cycles. No standalone Blender installation is needed at runtime.
3. Terrain C++ shared libraries (CPU + optional CUDA) compiled from Infinigen's source for landscape generation.
4. A patched gin-config resource reader to support namespace package path resolution in non-editable installs.
5. setuptools<70 pinned for pkg_resources compatibility required by landlab.

Building the package on Deadline Cloud

You can build this package on a Deadline Cloud farm that is configured for package build jobs that update an S3 conda channel and has a CUDA fleet to run the build. See the AWS Deadline Cloud developer guide Create a conda channel using S3 for setup instructions, and the CUDA farm CloudFormation template for a quick way to deploy a CUDA-capable farm.

Submit a build job from the conda_recipes directory of the Deadline Cloud samples repository as follows. If you need to customize this, see the conda recipe samples README.

$ ./submit-package-job infinigen-1.19.0

The build requires a GPU worker (for CUDA terrain compilation) and takes ~6 minutes.

Building the package locally

You can also build the recipe directly with rattler-build on a Linux host, optionally with CUDA available. From the conda_recipes directory:

rattler-build publish infinigen-1.19.0/recipe/recipe.yaml \
    --to s3://amzn-s3-demo-bucket/Conda/Default \
    -c conda-forge \
    --build-number=+1

If your machine has no CUDA GPU but you still want to build the package, the terrain compilation step falls back to a CPU-only build automatically. You will need a GPU at runtime to use the nature (terrain) scene type.

Use with the companion job bundle

See the infinigen_scene_gen job bundle for submitting scene generation jobs against the package built from this recipe.

The job template's CondaPackages parameter defaults to infinigen jinja2 pyyaml setuptools. The extra packages are transitive dependencies that are not bundled in the conda package itself.

Licensing

Infinigen is distributed under the BSD-3-Clause license. The bpy Python module that Infinigen depends on at runtime is GPL-licensed (Blender). This recipe builds Infinigen from upstream source and pulls bpy from PyPI at install time. No GPL binaries are redistributed in this samples repository.

Note on the package you build from this recipe: the resulting conda package physically bundles bpy (and its Blender components), which are licensed under the GNU GPL. The built package artifact is subject to GPL terms, even though its license metadata records BSD-3-Clause (which reflects Infinigen's own source license only). Building and using the package to render is unrestricted; however, if you redistribute the built package to third parties, you are responsible for complying with the GPL, including making the corresponding source available. Consult your own legal/open-source guidance before redistributing.

Known issues

• Infinigen v1.19.0 has a bug in reptile.py where reptile_postprocessing() is called with the wrong arguments. Workaround: disable snakes via -p SceneOverrides="populate_scene.populate_snake_enabled=False" when submitting from the infinigen_scene_gen job bundle.
• Nature scenes without simple.gin take 30-60+ minutes due to full asset generation.
• The intermediate .blend files show placeholder bounding boxes in the Blender viewport, which is by design. Full geometry is only generated during the Cycles render pass.

Contributing this package recipe to conda-forge

Both Infinigen and bpy are open source. Infinigen and its dependencies are good candidates to contribute as conda package recipes to conda-forge. See the conda-forge documentation about contributing packages to learn more about the process. The recipe provided here is a good starting point, but is not ready to contribute as-is. Recipes in conda-forge feedstocks must follow stricter conventions to inter-operate with the full set of conda-forge packages.