Researchers Joris Rijsdijk, Christoph Peters, Michael Weinnman and Ricardo Marroquim introduced Gaussian point splatting, a stochastic rendering technique that samples pixel‑sized opaque points from Gaussian primitives. By writing each point to the framebuffer with 64-bit atomics, the method distributes work across millions of GPU threads, enabling real‑time rendering of massive point clouds. It avoids costly per‑pixel blending and scales linearly with size.

The team formalized how many points each Gaussian should emit and how to balance opacity, solving a long‑standing ambiguity in splat‑based pipelines. Hierarchical frustum and occlusion culling prune invisible contributions, keeping the workload tight even when scenes contain hundreds of millions of Gaussians. Such pruning keeps memory bandwidth within typical GPU limits. The authors presented the work at SIGGRAPH 2026 in July.

Benchmarks show the approach matches visual fidelity of traditional Gaussian splatting while delivering interactive frame rates; the only artifacts are minor noise and altered aliasing patterns. Source code and demo videos are released on GitHub and a Shadertoy page, allowing practitioners to adopt the technique with minimal code changes in novel view synthesis and scene reconstruction pipelines.