search

dendenxu / fast-gaussian-rasterization

A geometry-shader-based, global CUDA sorted high-performance 3D Gaussian Splatting rasterizer. Can achieve a 5-10x speedup in rendering compared to the vanialla diff-gaussian-rasterization.
MIT License
257 stars 3 forks source link

python version #3

Closed EH-HU closed 4 months ago

EH-HU commented 4 months ago

I was wondering if it's not available in Python version 3.8.

dendenxu commented 4 months ago

I don't recall using any special features from 3.9 or newer so theoretically it should also work on 3.8. I listed 3.9 in pyproject.toml since I only tested on 3.9 or newer versions of Python.

I'll test on more python-pytorch version combinations in the future, but in the moment I don't have a readily available Python 3.8 environment at hand. Maybe you could clone the repo and manually update pyproject.toml's Python requirement to 3.8 and check whether there's any issue? If all goes well, I can release a version that also marks Python 3.8 as supported.

dendenxu commented 4 months ago

ps: another way to install the package without installing all dependencies: pip install fast_gauss --no-build-isolation --no-deps

EH-HU commented 4 months ago

Thank you for your quick response. Do you know why the rendering time was longer when using that module?

dendenxu commented 4 months ago

Could be due to multiple factors, for example, the number of points (the lower the faster compared to the vanilla) and the target resolution (the higher the faster compared to the vanilla renderer). Another thing to note is that the speedup is the most visible when you're rendering to a screen, compared to maybe saving the rendered image to disk.

EH-HU commented 4 months ago

Thank you for your response.