Skip to content
Book Open access

BubbleSplat: Forget Swollen LODs with Low-Redundancy, Inflation-Driven 3D Gaussian Streaming

Nov 2026 · Proceedings of the 34th ACM International Conference on Multimedia · 0 citations · 29 references

Abstract

Streaming large-scale 3D Gaussian Splatting (3DGS) scenes to bandwidth- and memory-constrained clients is an open systems problem. Replacement-based level-of-detail (LOD) pipelines bound client memory but re-transmit overlapping content across levels, wasting 57-62% of bandwidth on redundancy. Vanilla additive streaming transmits each primitive once yet produces incoherent early views with no path to memory-efficient rendering. We present BubbleSplat, the first Sculpted Additive progressive 3DGS streaming system. BubbleSplat is entirely training-free: a one-time offline stage partitions a pretrained scene into spatial tiles, constructs append-only coarse-to-fine levels via hybrid saliency-spatial sampling, and applies support-aware inflation that enlarges sparse early Gaussians to produce visually coherent partial views. At runtime, a value-per-byte scheduler prioritizes tiles by viewpoint saliency and completion, while inflated coarse levels serve as a persistent LOD cache that enables bounded-memory rendering without re-transmission. Across four benchmark scenes spanning 2.5M to 81M Gaussians, BubbleSplat reduces bandwidth redundancy to 11.1%, requires 3.2× fewer bytes than the strongest replacement-based baseline to reach 27dB PSNR, and sustains real-time rendering on consumer mobile devices---including an 810M-splat stress test at 60FPS on mobile phones and 22FPS under a 128MB VRAM budget.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.