AccelGS: An Acceleration Framework for Large-Scale 3D Gaussian Splatting Training

Large-scale 3D Gaussian Splatting (3DGS) currently suffers from challenges, namely slow training, heavy GPU memory footprint, and boundary artifacts. We propose AccelGS, an acceleration framework for large-scale 3DGS training. (1) We design a cache acceleration mechanism that reuses precomputed visibility and geometric data once densification stops, drastically reducing computation and communication. (2) We introduce a conditional gradient update strategy that selectively updates only the parameters essential for loss reduction, minimizing memory usage and computational overhead. (3) A distributed adaptive densification control is developed that dynamically terminates Gaussian densification based on loss reduction rates, avoiding unnecessary model growth. (4) A non-overlapping partitioning paradigm with cross-partition synchronization is adopted to eliminate boundary artifacts. Extensive experiments demonstrate that AccelGS achieves up to 12× speedup over 3DGS, up to 14.4× speedup over MomentumGS, and up to 78% speedup over the state-of-the-art GrendelGS, while maintaining rendering quality. Project page: https://aicodeclub.github.io/AccelGS.