Splaxel Cuts Distributed 3DGS Training Communication 7.6x With Pixel-Level Exchange

Up to 7.6x speedup over the state-of-the-art distributed 3D Gaussian Splatting framework, while keeping reconstruction quality intact, with communication cost that doesn't bloat as you add more Gaussians. That's what the Splaxel paper delivers, and it solves the one bottleneck that made scaling 3DGS training to large scenes painful.

The Stupid Tax on Gaussian Synchronization

Existing distributed 3DGS approaches either partition scenes into isolated regions (which introduces global inconsistency) or synchronize all Gaussians globally. That global synchronization means every GPU talks to every other GPU about every Gaussian- and as you scale to hundreds of millions of Gaussians, that communication cost dominates iteration time. You end up spending more time moving data than computing. Splaxel's authors looked at that and said: why are we moving Gaussians at all?

Pixel-Level Local Rendering, Global Composition

Here's the clever trick: instead of synchronizing Gaussians, each GPU renders its local subset independently, then exchanges only partial pixel values with neighboring GPUs. The math stays consistent because the final rendered image is a composition of these local renders. Communication cost stays stable as the scene size increases because you're only shipping pixel values, not Gaussian parameters. The paper introduces two additional optimizations- geometric and transmittance visibility prediction to prune redundant pixel exchanges, and conflict-free camera-view consolidation to keep GPU utilization high.

Results That Scale Without Pain

Splaxel was evaluated on large-scale datasets containing up to 120 million Gaussians. Compared to the leading distributed 3DGS framework, Splaxel achieved up to 7.6x speedup while maintaining high reconstruction quality. That speedup doesn't come from cutting corners on math- it comes from cutting out the communication that was never necessary in the first place.

This approach opens the door to training 3DGS on scenes orders of magnitude larger than what's feasible today, without requiring exotic interconnect hardware or sacrificing consistency across regions.

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