Title: HiGS: A Hierarchical Rendering Architecture for Real-Time 3D Gaussian Splatting

Abstract:

3D Gaussian Splatting (3DGS) has emerged as the benchmark for real-time novel view synthesis on consumer-grade GPUs. However, its current pipeline binds spatial partitioning and rasterization to a single tile size, creating a fundamental conflict: partitioning—which involves binning and depth-sorting gaussians—becomes more efficient with larger tiles, whereas rasterization benefits from smaller ones. While previous acceleration methods have lowered the cost of individual stages, they maintain this rigid single-scale constraint, resulting in performance bottlenecks where a small number of dense tiles dictate the overall frame time.

To address this, we introduce Hierarchically Tiled Gaussian Splatting (HiGS), a system that assigns distinct scales to each process. In HiGS, partitioning operates across coarse macro-tiles, while rasterization executes over finer render tiles contained within them. Furthermore, rasterization workload is distributed based on the density of gaussians within each macro-tile, rather than being fixed per tile. This approach allows dense regions to leverage multiple parallel processing units simultaneously, avoiding serialization through a single unit. Evaluations across various test scenes demonstrate that HiGS achieves rendering speeds up to 15.8 times faster than the original 3DGS implementation. It surpasses all other evaluated rasterizers while maintaining exact front-to-back alpha compositing.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC