AdpSplit cuts 3D Gaussian Splatting training time by up to 22%

A major bottleneck in 3D Gaussian Splatting (3DGS) has been the rigid binary splitting used during adaptive density control, which requires many densification rounds to expose fine scene geometry. Researchers now introduce AdpSplit, an error-driven adaptive split operator that replaces the standard binary split. Instead of fixed-cardinality splitting, AdpSplit leverages L1-pixel-error region statistics to dynamically determine the number of child Gaussians and initialize their parameters. This enables fewer densification iterations, reducing training time while maintaining rendering fidelity.

Across three benchmark datasets—MipNeRF360, Deep-Blending, and Tanks&Temples—AdpSplit cuts training time by 9.2% to 22.3% when integrated as a drop-in replacement into multiple accelerated 3DGS pipelines like FastGS. With FastGS specifically, AdpSplit matches the full-schedule PSNR on MipNeRF360 while reducing training time by 16.4%, corresponding to a 12.6x acceleration over vanilla 3DGS. This work directly addresses the efficiency bottleneck in 3D scene reconstruction, enabling faster geometry discovery without sacrificing quality.