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

PocketGS is a research attempt to move 3D Gaussian Splatting training from workstation-class hardware onto mobile devices without blowing up memory or turnaround time. The paper’s core idea is to co-design the training pipeline around mobile constraints instead of just shrinking the model: geometry-faithful point-cloud priors, local surface statistics to initialize anisotropic Gaussians, and a backpropagation scheme that caches intermediates and scatters gradients efficiently.

For game developers, the interesting part is not just faster training, but the possibility of a practical capture-to-render loop on-device. That could matter for AR/VR content creation, rapid environment capture, or tools that let artists and technical artists prototype real-world scenes without hauling data back to a workstation. The authors say PocketGS outperforms a mainstream workstation 3DGS baseline under mobile budgets, though this is still an arXiv paper rather than a shipping engine feature.