TideGS: Scalable Training of Over One Billion 3D Gaussian Splatting Primitives via Out-of-Core Optimization

Researchers introduced TideGS, an out-of-core training framework for 3D Gaussian Splatting.

It uses block-virtualized geometry, an asynchronous pipeline, and trajectory-adaptive differential streaming to move parameters across SSD, CPU, and GPU memory.

With this design, a single 24 GB GPU can train models with over one billion Gaussians.

The approach expands 3D scene reconstruction to much larger scales while improving quality on large scenes.