3D Gaussian Splatting (3DGS) has been one of the most prominent technical approaches in the 3D reconstruction field in recent years. However, it faces a critical bottleneck: VRAM.
Each Gaussian primitive carries a large attribute vector. When the number of primitives scales to tens of millions, the parameter table exceeds GPU capacity. Previously, on consumer-grade single-GPU hardware, systems could handle at most a few tens of millions of Gaussians.
The TideGS paper, accepted as a Spotlight at ICML 2026, shatters this ceiling by reaching 1 billion+ Gaussians—all using just a single 24GB GPU.
Core Insight: 3DGS Training is Inherently Sparse
The team's starting point is clever: 3DGS training is inherently sparse and trajectory-conditioned.
Each iteration only activates the Gaussians visible from the current camera batch. This means GPU VRAM doesn't need to serve as persistent parameter storage; instead, it can act as a working set cache—loading only the subset of Gaussians currently needed.
Three Synergistic Technologies
TideGS manages parameters through a three-tier SSD-CPU-GPU storage hierarchy:
1. Block-Virtualized Geometry
Aligns spatial locality with SSD architecture. It organizes 3D space into blocks, ensuring that physically adjacent Gaussians are also adjacent in storage, thereby reducing I/O fragmentation.
2. Hierarchical Asynchronous Pipeline
Overlaps I/O and computation. While training the current batch on the GPU, it prefetches the next batch of required Gaussian data from the SSD, ensuring neither process blocks the other.
3. Trajectory-Adaptive Differential Streaming
Transfers only the working set increments between iterations. Instead of fully reloading data every time, it calculates which Gaussians have changed state and streams only the deltas.
Performance & Scale
The comparison numbers speak for themselves:
- Standard in-memory training: ~11 million Gaussians
- Previous out-of-core baselines: ~100 million Gaussians
- TideGS: Over 1 billion Gaussians
On large-scale scenes, TideGS also surpasses evaluated single-GPU baselines in reconstruction quality.
Why It Matters
3D reconstruction is transitioning from labs to real-world applications—autonomous driving, digital twins, and AR/VR all require processing city-scale, large-scale scenes. TideGS enables single-GPU training for billion-level Gaussians, significantly lowering the hardware barrier for large-scale 3D reconstruction.
Paper: arXiv:2605.20150