HybridGS: A Cross-Platform Architecture for Primitive and Non-Primitive Rendering using C++ and WebGPU

Abstract

The evolution of the web into a high-performance application platform, fueled by WebGPU and WebAssembly, challenges the historical dichotomy between global accessibility and the raw computational power of native C++ environments. Parallel to this architectural shift, 3D Gaussian Splatting (3DGS) has emerged as a transformative technique for photorealistic rendering, overcoming the geometric and labour-intensive limitations of traditional triangular meshes. While 3DGS enables real-time rendering, its heavy memory and computational demands pose significant challenges in resource-constrained environments. Some current solutions include Gaussian model compression or hybrid approaches that enable the use of both primitive-based meshes and non-primitive models in the same scene. Although recent literature establishes the feasibility of hybrid rendering using C++ and the Dawn WebGPU backend, there is a distinct lack of actionable architectural guidance on low-level implementation details, such as vertex generation strategies and memory alignment. Furthermore, there is a lack of rigorous performance validation concerning the trade-offs between reconstruction quality, computational cost, and memory usage.

This thesis addresses these gaps by presenting the design, implementation, and rigorous validation of HybridGS, a novel cross-platform rendering engine built with C++ and the Dawn WebGPU backend. To advance the development of high-performance web graphics, this research makes three primary contributions. First, we provide a comprehensive architectural blueprint for our rendering pipeline, explicitly detailing the geometric construction of splats as planar quadrilaterals, per-splat quantization strategies, and hybrid scene compositing. Second, we evaluate the performance benefits of primitive-based meshes versus non-primitive models, empirically confirming that traditional meshes remain vastly more efficient for structural geometry than their volumetrically converted counterparts. Finally, we present what is, to our knowledge, the first rigorous empirical benchmarking of the compressed SPLAT format. Through extensive evaluation on the Synthetic NeRF dataset, we compare the inherently 8-bit (U8x4) SPLAT format and third-degree Spherical Harmonic PLY format utilizing full-precision (F32x4), half-precision (F16x4), and 8-bit (U8x4) colour quantization, against the reference implementation baselines to quantify the precise trade-offs between the various techniques. We identify that F16x4 quantization for PLY models yields significant memory reductions with negligible impact on perceptual quality, offering an optimal balance. Conversely, while the SPLAT format achieves the highest computational and memory efficiency, our evaluation reveals a distinct perceptual cost that restricts its viability to strictly resource-constrained scenarios. Ultimately, this thesis delivers concrete, validated architectural guidance for deploying photorealistic, high-performance hybrid graphics on the modern web.