Gaussian Point Splatting: Scaling Novel View Synthesis to Hundreds of Millions of Gaussians

Gaussian Point Splatting enables real-time rendering of massive scenes

Gaussian Point Splatting is a stochastic rendering method designed to scale novel view synthesis to scenes containing hundreds of millions of Gaussians. By sampling pixel-sized, opaque points from Gaussians and splatting them to a framebuffer using 64-bit atomics, the method achieves high-performance rendering that remains faithful to the original Gaussian splatting technique while maintaining real-time speeds.

Stochastic sampling and GPU acceleration

The core technical innovation of Gaussian Point Splatting is the transition from traditional Gaussian rendering to a stochastic point-based approach. Instead of rendering the full Gaussian distribution, the system samples opaque points that are approximately the size of a pixel. These points are then splatted to the framebuffer using parallel programming primitives, allowing the workload to be distributed across millions of GPU threads.

To ensure the resulting image is faithful to the original Gaussian splatting, the researchers formalized and solved the problem of determining the optimal number of points to sample per Gaussian and how to distribute those points to maintain correct opacity. This process prevents the stochastic nature of the sampling from distorting the visual representation of the scene.

Performance optimization through culling and atomics

To achieve real-time performance with hundreds of millions of Gaussians, the method employs several optimization strategies:

  • 64-bit Atomics: The use of 64-bit atomics allows multiple threads to splat points to the same pixel independently without causing race conditions or requiring expensive synchronization.
  • Hierarchical Culling: The system implements both hierarchical frustum culling and occlusion culling to avoid processing Gaussians that are not visible to the camera.
  • Parallelism: The workload is distributed across millions of threads, ensuring that GPU resources are fully utilized.

Visual trade-offs and quality considerations

While Gaussian Point Splatting enables unprecedented scale, it introduces specific visual artifacts compared to traditional 3D Gaussian Splatting (3DGS). The primary differences are the introduction of slight noise and differences in aliasing.

Community discussion highlights that the stochastic nature of the method—similar to Monte Carlo rendering—introduces noise that may require significant sampling to resolve. One observer noted that the interactive viewer suggests approximately 128 samples per pixel (spp) may be required to match the quality of 3DGS. However, the researchers suggest that more sophisticated denoisers could mitigate these issues, and the current implementation includes temporal reprojection to reduce noise.

Comparison to existing techniques

Gaussian Point Splatting differs from traditional point splatting (a 1990s rendering technique) and mesh splatting. While mesh splatting uses triangles to represent sharp features more effectively than Gaussians, Gaussian Point Splatting focuses on the scalability of the rendering pipeline for massive datasets.

Furthermore, the method is orthogonal to Level-of-Detail (LoD) rendering. This means Gaussian Point Splatting can be combined with LoD techniques in the future to further increase performance gains for large-scale scene rendering.

Sources