View a markdown version of this page

Build a NeRF Studio conda package for Deadline Cloud - Deadline Cloud

Build a NeRF Studio conda package for Deadline Cloud

The nerfstudio rattler-build recipe packages NeRF Studio plus the following:

  • The Splatfacto in the Wild external model.

  • NeRF Studio dependencies that are not yet available on conda-forge.

  • The nerfstudio/gsplat package's simple_trainer.py example, available as the command gsplat_simple_trainer.

  • The KevinXu02/splatfacto-w package's export_script.py, available as the command splatfactow_export.

To build this recipe, deploy the CUDA farm CloudFormation template to create a Deadline Cloud farm with a CUDA fleet for the build, then submit the package build job:

./submit-package-job nerfstudio

To reuse conda environments between jobs, attach the conda_queue_env_improved_caching.yaml queue environment to your queue. Because the dependency closure of NeRF Studio contains many gigabytes of packages, caching saves significant time and bandwidth.

For a job bundle that uses this package, see Train 3D Gaussian Splatting from video on Deadline Cloud.