arnaudstiegler / torch-gaussian-splatting-rasterizer

Reproducing Gaussian Splatting

Minimal Gaussian Splatting Rasterizer

This is a torch implementation of the rasterization process for Gaussian Splatting.

Custom CUDA kernels are re-implemented in Torch, thereby loosing the CUDA parallelism. As a result, the implementation rasterizes Gaussians one at a time rather than distributing the rasterization per pixel (and parallelizing it with custom CUDA kernels). It is of course pretty suboptimal and not meant to be used for training/evaluation: rendering with this codebase will take about 5 minutes / image vs. less than a second for the original implementation.

render_video.mp4

Install

pip install -r requirements.txt

Additionally, if you want to generate the video, you will need ffmpeg installed.

How To Run

MipNerf 360 scenes can be found at: https://jonbarron.info/mipnerf360/

Trained Gaussian Splatting models can be found at: https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/datasets/pretrained/models.zip

To try out the rasterization, run the following:

python rasterize.py --input_dir {MIPNERF_360_PATH} --trained_model_path {GAUSSIAN_MODEL_PATH} --output_path {OUTPUT_PATH} --scene-index {SCENE_INDEX} --scale-factor 2 [--generate_video]