WebSplatter: Efficient and Faithful In-Browser
3D Gaussian Splatting across Devices via WebGPU


ACM Multimedia 2026


Yudong Han1    Chao Xu2    Xiaodan Ye2    Weichen Bi1    Xu Xuanhuai3    Zilong Dong2    Yun Ma1

1Peking University    2Alibaba Group    3Nanyang Technological University


ArXiv

Code

Paper

Demo

Interactive Demo

WebSplatter Van Gogh Room · 341,294 splats · 5.05 MiB SPZ-compressed GLB Click to focus · ` to capture mouse · WASD to move · Q / E to roll · Esc or ` to release · drag / right-drag / scroll when released Open full screen ↗

Abstract

3D Gaussian Splatting (3DGS) has become a core representation for immersive multimedia on the web, powering volumetric video, dynamic avatars, and interactive 3D experiences. While recent advances in compression and streaming bring 3DGS content closer to end users, existing web viewers naively port CUDA-based pipelines whose synchronization and memory patterns do not generalize across GPU architectures, causing rendering failures on non-NVIDIA and mobile devices. We present WebSplatter, a WebGPU-native rendering framework that eliminates these failure modes through a wait-free radix sort and a hardware-accelerated rasterization pipeline with opacity-aware quad sizing. Evaluated across eleven devices spanning five GPU architectures, WebSplatter is the first web-based 3DGS renderer that achieves cross-device reliability, rendering every benchmark scene on every tested device without failure. WebSplatter delivers 1.1 to 2.5 times speedups over the surviving baselines and near-lossless fidelity to the native CUDA renderer, with an average ΔPSNR of 0.07 dB.

Method

Overview of the WebSplatter rendering pipeline

WebSplatter adapts the complete 3DGS rendering pipeline to the WebGPU execution model. Its wait-free radix sort decomposes cross-workgroup dependencies into explicitly synchronized dispatches and uses bitmap-popcount ranking to preserve deterministic progress across GPU architectures. Its hardware-accelerated vertex-fragment rasterization pipeline removes memory-intensive tile-splat pair buffers, while opacity-aware quad sizing reduces overdraw without changing Gaussian evaluation.

Results

WebSplatter was evaluated on eleven consumer devices spanning NVIDIA, AMD, Apple, Qualcomm, and MediaTek GPU architectures. It rendered every benchmark scene on every tested device without failure, achieved 1.1 to 2.5 times speedups over the surviving web baselines, and matched the native CUDA renderer with an average ΔPSNR of 0.07 dB across 156 views from five scenes.

11

Devices

5

GPU architectures

1.1–2.5×

Speedup

0.07 dB

Average ΔPSNR