PocketGS: On-Device Training of 3D Gaussian Splatting for High Perceptual Modeling

Researchers have introduced two new frameworks, TideGS and PocketGS, to improve the training of 3D Gaussian Splatting (3DGS) models, enabling scalable and efficient processing on single GPUs and mobile devices.

TideGS, presented in a paper on arXiv^[1], allows for the training of over one billion 3D Gaussian Splatting primitives on a single 24 GB GPU via out-of-core optimization. This is a significant improvement over previous systems, which were limited to tens of millions of Gaussians due to memory constraints. TideGS achieves this through three synergistic techniques: block-virtualized geometry, hierarchical asynchronous pipeline, and trajectory-adaptive differential streaming. According to the researchers, TideGS achieved the best reconstruction quality among evaluated single-GPU baselines on large-scale scenes. Meanwhile, another research team introduced PocketGS, a mobile scene modeling paradigm that enables on-device 3DGS training on mobile devices^[2]. PocketGS resolves the tension between training efficiency, memory compactness, and modeling quality through three co-designed operators and outperforms mainstream workstation 3DGS baselines under mobile budgets.