Splaxel: Efficient Distributed Training of 3D Gaussian Splatting for Large-scale Scene Reconstruction via Pixel-level Communication

2026-06-16

AI Summary

This work addresses the challenges of high inter-GPU communication overhead and difficulty in maintaining global consistency during large-scale 3D Gaussian splatting training. The authors propose a distributed training framework based on pixel-level local rendering and global composition. Their approach introduces a novel pixel-level communication mechanism that replaces full Gaussian parameter synchronization by exchanging partial pixel values, combined with geometry- and transmittance-based visibility prediction and conflict-free camera view merging. This strategy preserves mathematical consistency and reconstruction fidelity while substantially reducing communication costs. Experiments demonstrate that the method achieves up to a 7.6× speedup over the current state-of-the-art on scenes containing up to 120 million Gaussians, with strong scalability.

Abstract

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.

Problem

• 3D Gaussian Splatting
• distributed training
• large-scale scene reconstruction
• communication efficiency
• multi-GPU

Innovation

• distributed training
• 3D Gaussian Splatting
• pixel-level communication
• large-scale reconstruction
• communication efficiency