Splaxel framework speeds up 3D scene training 7.6x via pixel-level communication

Splaxel addresses a core bottleneck in scaling 3D Gaussian Splatting (3DGS) to large-scale scenes: the explosion of inter-GPU communication when synchronizing hundreds of millions of Gaussians. Existing distributed approaches either partition scenes into isolated regions (causing global inconsistency) or rely on global Gaussian-level exchanges that quickly dominate iteration time. Splaxel introduces a pixel-level paradigm: each GPU renders its local Gaussian subset, then exchanges only partial pixel values with other GPUs. This keeps communication volume constant regardless of scene size, while preserving mathematical equivalence to full global rendering.

Splaxel further reduces pixel-level redundancy via geometric and transmittance visibility prediction, and improves GPU utilization through conflict-free camera-view consolidation. In experiments on large-scale datasets with up to 120M Gaussians, Splaxel achieves up to 7.6x speedup over the previous state-of-the-art distributed 3DGS framework while maintaining high reconstruction quality. The framework's communication efficiency makes it practical for real-world applications like autonomous driving, digital twins, and AR/VR, where large environments must be reconstructed quickly and accurately. The paper is available on arXiv and includes 17 pages with 25 figures detailing the architecture and benchmarks.