installation.md

Installation

Prerequisite

• NVIDIA GPU with at least 6GB VRAM. The more memory you have, the more methods and higher resolutions you can try.
• NVIDIA Driver whose version is higher than the Minimum Required Driver Version of CUDA Toolkit you want to use.

Install CUDA Toolkit

You can skip this step if you have installed sufficiently new version or you use Docker.

Install CUDA Toolkit.

• Example for Ubuntu 22.04:
  • Run command for CUDA 11.8 Ubuntu 22.04
• Example for Ubuntu on WSL2:
  • sudo apt-key del 7fa2af80
  • Run command for CUDA 11.8 WSL-Ubuntu

Git Clone

git clone https://github.com/threestudio-project/threestudio.git
cd threestudio/

Install threestudio via Docker

1. Install Docker Engine. This document assumes you install Docker Engine on Ubuntu.
2. Create docker group. Otherwise, you need to type sudo docker instead of docker.
3. Install NVIDIA Container Toolkit.
4. If you use WSL2, enable systemd.
5. Edit Dockerfile for your GPU to speed-up build. The default Dockerfile takes into account many types of GPUs.
6. Run Docker via docker compose.

cd docker/
docker compose build  # build Docker image
docker compose up -d  # create and start a container in background
docker compose exec threestudio bash  # run bash in the container

# Enjoy threestudio!

exit  # or Ctrl+D
docker compose stop  # stop the container
docker compose start  # start the container
docker compose down  # stop and remove the container

Note: The current Dockerfile will cause errors when using the OpenGL-based rasterizer of nvdiffrast. You can use the CUDA-based rasterizer by adding commands or editing configs.

• system.renderer.context_type=cuda for training
• system.exporter.context_type=cuda for exporting meshes

This comment by the nvdiffrast author could be a guide to resolve this limitation.