Gaussian Point Splatting
Gaussian point splatting method scales to hundreds of millions of Gaussians using GPU atomics and parallel workload distribution.
Rijsdijk et al. propose a stochastic rendering method that samples pixel-sized opaque points from Gaussians and splatts them using 64-bit atomic operations. The approach distributes work across millions of GPU threads while maintaining faithfulness to original Gaussian splatting through formal solutions for opacity and point distribution. Hierarchical frustum and occlusion culling further accelerate rendering. The method produces slight noise and aliasing differences compared to standard Gaussian splatting but achieves real-time performance on scenes with hundreds of millions of Gaussians. Published in ACM Transactions on Graphics (SIGGRAPH 2026).
What HN community is saying
Commenters noted the technique resembles stochastic progressive point cloud rendering used in VFX for over 15 years, though the Gaussian-specific opacity correction and atomic-based workload distribution appear novel. The core contribution lies in making Gaussian splatting feasible on modern GPU hardware rather than inventing the fundamental approach. Discussion clarified that millions refers to logical GPU threads, not OS threads, and that avoiding explicit sorting of Gaussians appears to be a key efficiency gain over conventional 3DGS pipelines.