gaussian-point-splatting

Gaussian Point Splatting

We propose Gaussian point splatting, a stochastic method for rendering massive 3DGS scenes. By sampling pixel-sized opaque points, splatting them atomically, and applying stochastic transparency, we eliminate the need for sorting. Our approach distributes workload evenly across GPU threads, enabling the real-time display of hundreds of millions of Gaussians efficiently.

Citation

If you find this code useful for your research, please consider citing our paper:

@article{Rijsdijk2026GaussianPointSplatting,
  title = {Gaussian Point Splatting},
  author = {Rijsdijk, Joris and Peters, Christoph and Weinnman, Michael and Marroquim, Ricardo},
  journal = {ACM Trans. Graph.},
  volume = {45},
  number = {4},
  publisher = {Association for Computing Machinery},
  year = {2026},
  doi = {10.1145/3811272}
}

How to Run

Prerequisites

• CUDA Toolkit (developed with version 12.6)
• CMake (>= 3.27)
• All other third-party packages are included directly in /packages
• Note: This application requires an NVIDIA GPU to work. On systems with multiple GPUs (e.g., laptops with integrated graphics), you must ensure the application is configured to run on your dedicated NVIDIA GPU. If the executable defaults to the integrated GPU, the CUDA-OpenGL interoperability will fail to initialize.

Installation

• Clone the repository
• Build the project using CMake

Example Usage

./gaussian_splatting --model-path ./data/scene_01 --resolution 1920 1080 --samples-per-pixel 4

Command Line Arguments

Argument	Description
--model-path <path>	Path to the trained Gaussian point model. (Either a .ply file, or the root folder of a 3DGS output.)
--resolution <w> <h>	Sets the render window resolution (e.g., --resolution 1920 1080).
--supersampling-factor <n>	Overrides the default supersampling factor, (2 -> 2x2 samples per output pixel)
--max-number-of-points <n>	Overrides the default limit on the number of points per frame, (250e6). If the limit is too low, some Gaussians might not render. Higher limits use more VRAM.
--samples-per-pixel <n>	Overrides the default number of samples per pixel, (1)
--fovy <degrees>	Overrides the default camera field of view (45)
--near-view-plane <distance>	Overrides the default camera near view plane (0.2)
--far-view-plane <distance>	Overrides the default camera far view plane (10000)
--camera-path <path>	Load a .json camera trajectory exported by 3DGS. If --model--path points to a 3DGS output directory, it automatically loads the camera path.
--disable-hierarchical-culling	Disables hierarchical culling optimizations.
--disable-occlusion-culling	Disables occlusion culling.
--sort-morton-order	Enables Morton order sorting for faster processing and lower (system) RAM usage.
--one-point-per-thread	If enabled, the renderer uses K = 1 (as described in the paper). By default, K = (supersampling factor)^2. Enabling this flag enables mathematically more correct rendering, at the cost of (much) higher frame times.
--play-camera-path	Automatically play the loaded camera path on start.
--disable-gui	Disables the UI overlay.
--cull-small	Enables culling for small Gaussian splats from Splatshop. This does not make a big difference in frame time for our renderer.
--background <r> <g> <b>	Sets the RGB background color.

Controls

• Use WASD (horizontal movement), space bar (vertical up), ctrl (vertical down) and the mouse (rotation) to control the camera.
• Use backspace to toggle the visibility of the UI.
• Use escape to toggle mouse pointer lock.
• Use F5 to take a screenshot, saved at ./out/screenshots.

---

Important Files

The main contributions of the paper can be found in the following files:

File	Description
src/main.cu	Render loop and program entry
src/config.h	Defines compile-time constants, for instance to toggle spherical harmonics and compression.
src/core/rendering/passes/gaussian_point_splatting.cu	Implementation of the main method, including kernels for preprocessing, splatting, and combining of samples.
src/core/rendering/passes/post_processing_kernels.cuh	Implementation of temporal reuse.
src/core/random/workload/workload_distributor.cuh	Handles the workload distribution algorithm as discussed in the paper.
src/core/gaussians/gaussian_types.cuh	Implementation of our compressed Gaussian format.
src/core/rendering/occlusion/gaussian_bvh.cu	Computes the hierarchy and handles hierarchical culling.
src/core/rendering/occlusion/depth_mip_chain.cu	Computes the depth mip chain and exposes a device function to check whether an AABB at a certain depth is occluded or not.
assets/cameras/*.json	The camera paths used to run our experiments and render our supplemental video. The file format is consistent with 3DGS.

---

Evaluation Scenes

The following scenes were used to test the implementation:

• Opole campus (25 M Gaussians),

• St. Sebastian church (30 M Gaussians)

• Jastrzebia Góra (106M Gaussians)

• The Tanks and Temples, Mip-NeRF360 and Deep Blending 3DGS scenes can be found here, or linked on the 3DGS repository

---

License

This project is licensed under the BSD 3-Clause License - see the LICENSE file for details.

Acknowledgments

This work was supported by the Dutch Research Council (NWO) under the project VR Retrofit-4U (Grant ID: 10.61686/EIHMV70145).

The large-scale scene (Jastrzebia Góra) used in our teaser was provided by Andrii Shramko, Teleportour.