Search papers, labs, and topics across Lattice.
UWANLIndependent ResearcherRiceTexas A&MUHUniversity of GeorgiaUT ArlingtonJun 17, 2026arXiv:2606.18588This paper introduces Splaxel, a novel framework for distributed training of 3D Gaussian Splatting (3DGS) that addresses the challenges of scaling to large-scale scenes by optimizing communication between GPUs. By leveraging pixel-level local rendering and global composition, Splaxel minimizes inter-GPU communication while ensuring mathematical consistency, resulting in a significant reduction in iteration time. Evaluations demonstrate that Splaxel achieves up to 7.6脳 speedup over existing methods while maintaining high-quality scene reconstruction, making it a compelling solution for real-time applications.
Achieving a 7.6脳 speedup in distributed 3D scene reconstruction without sacrificing quality could redefine efficiency benchmarks in computer graphics.
3D Gaussian Splatting (3DGS) enables high-fidelity and real-time 3D scene reconstruction, but scaling training to large-scale scenes requires optimizing hundreds of millions of Gaussians across multiple GPUs. Existing distributed approaches either partition scenes into isolated regions, causing global inconsistency, or rely on global Gaussian-level exchanges, which lead to substantial growth in inter-GPU communication and quickly dominate iteration time. We propose Splaxel, a communication-efficient distributed 3DGS training framework based on pixel-level local rendering and global composition. Instead of synchronizing Gaussians, each GPU renders its local subset and exchanges only partial pixel values, maintaining mathematical consistency while keeping communication cost stable as the scene size increases. Splaxel further reduces pixel-level redundancy through geometric and transmittance visibility prediction and improves GPU utilization via conflict-free camera-view consolidation. Evaluated on large-scale datasets with up to 120M Gaussians, Splaxel achieves up to 7.6$\times$ speedup over the state-of-the-art distributed 3DGS framework while preserving high reconstruction quality.
