Tiny Gaussian Splatting Viewer

This is a simple Gaussian Splatting Viewer built with PyOpenGL / CUDARasterizer. It's easy to install with minimum dependencies. The goal of this project is to provide a minimum example of the viewer for research and study purpose.

News!

1/10/2024: The OpenGL renderer has faster sorting backend with torch.argsort & cupy.argsort. With cuda based sorting, it achieves nearly real-time sorting with OpenGL backend.

12/21/2023: Now we support rendering using the official cuda rasterizer!

Usage

Install the dependencies:

pip install -r requirements.txt

Launch the viewer:

python main.py

You can check how to use UI in the "help" panel.

The Gaussian file loader is compatiable with the official implementation. Therefore, download pretrained Gaussian PLY file from this official link, and select the "point_cloud.ply" you like by clicking the 'open ply' button, and you are all set!

Optional dependencies:

• If you want to use cuda backend for rendering, please install the diff-gaussian-rasterization following the guidance here. And also install the following package:

pip install cuda-python

• For sorting, we provide three backend: torch, cupy, and cpu. The implementation will choose the first available one based on this priority order: torch -> cupy -> cpu. If you have torch or cupy backend, turning on auto sort will achieve nearly real-time sorting.

  • If you want to use torch as sorting backend, install any version of PyTorch.

  • If you want to use cupy to accelerate sorting, you should install the following package:

  pip install cupy-cuda11x // for cuda 11
  pip install cupy-cuda12x // for cuda 12
  

Troubleshoot

The rendering speed of is comparable to the official CUDA renderer. If you're experiencing slow rendering, it's likely you are using integrated graphics card instead of a high-performance one. You can configure python to use high-performance graphics card in system settings. In Windows, you can set in Setting > System > Display > Graphics. See the screenshot below for example.

Limitations

• The implementation utilizes SSBO, which is only support by OpenGL version >= 4.3. Although this version is widely adopted, MacOS is an exception. As a result, this viewer does not support MacOS.

• The cuda backend currently does not support other visualizations.

• Based on the flip test between the two backends, the unofficial implementation seems producing slightly different results compared with the official cuda version.

TODO

• Add orthogonal projection
• Make the projection matrix compatiable with official cuda implementation
• Tighter billboard to reduce number of fragments
• Save viewing parameters