Gaussian splats part 3: There’s always a bigger scene
Summary
Magnopus documents the ongoing Gaussian splats project, focusing on breaking the 16 million splat cap by moving from GPU Scene to an indirect-instanced rendering pipeline. It details streaming LOD (streamed SOG), asynchronous loading, per-splat data on the GPU, and a memory-budgeted approach that allows tens to hundreds of millions of splats to render adaptively. The article also covers practical challenges, memory management, and future work to improve performance and scalability.