Training Acceleration

FastGS

FastGS is a CVPR 2026 Highlight project focused on accelerating 3D Gaussian Splatting training, targeting scene training in about 100 seconds while preserving comparable quality.

Stage: TrainingFast experimentsResearch comparisonAcceleration studies

Open official linkProject Homepage->

What It Does

FastGS targets one of the main pain points in Gaussian Splatting: iteration speed. Faster training matters when you need to evaluate many captures, tune parameters, or compare methods.

It is best treated as a research and advanced pipeline tool. For casual capture, a desktop or cloud product may be easier; for method work, FastGS is worth studying closely.

How To Use It In 3DGS

Prepare a COLMAP or compatible dataset as you would for vanilla 3DGS.
Run the provided standard or high-quality training scripts.
Compare output quality, Gaussian count, memory behavior, and training time against your baseline.
Use SuperSplat or a browser viewer for fast visual inspection after training.

Things To Watch

The headline speed is workload-dependent, so benchmark on your own scenes.
Research repositories often inherit setup complexity from the original 3DGS stack.
Faster convergence does not fix weak capture coverage or bad poses.

Official And Related Links

Project Homepage GitHub Repository Original 3DGS Repository

Current Research Status

FastGS is marked as a CVPR 2026 Highlight and reports 2-7x acceleration across varied backbones, with the project headline of about 100-second scene training.
Released modules include Fast-D3DGS for dynamic scenes, Fast-DropGaussian for sparse-view reconstruction, and Fast-PGSR for surface reconstruction.
Hardware notes still resemble the original 3DGS stack: CUDA-ready GPU, compatible compiler/CUDA versions, and high-VRAM cards for paper-quality runs.