PocketGS: High-Fidelity On-Device Training for 3D Gaussian Splatting
PocketGS enables high-quality, end-to-end 3D Gaussian Splatting reconstruction directly on commodity smartphones. It resolves conflicts between training speed, memory limits, and quality by using three co-designed operators: a geometry prior builder (G), a parameter initializer (I), and a hardware-aligned optimization engine (T). This allows for practical capture-to-render workflows on mobile devices.