Mobile-4DGS: Unified Static-Dynamic Real-time Mobile Gaussian Splatting

Mobile-4DGS. Unified Static-Dynamic Real-time Mobile Gaussian Splatting. By Xiaobiao Du and five others, from the University of Technology Sydney, Yale University, and three other institutions. Published on the fourth of October. Its code repository is worked out: about six hundred and fifty source files, with a readme.

3D Gaussian Splatting has become remarkably good at novel view synthesis, but getting it onto a phone is another matter. Resource-constrained mobile devices struggle with both static and dynamic Gaussian representations, for three reasons the authors name: heavy storage, redundant primitives, and costly per-frame computation. Each of those is a cost paid on the device, and dynamic scenes pay the per-frame one over and over during playback.

Mobile-4DGS is a single lightweight framework meant to handle both cases, static and dynamic, at high fidelity and in real time on mobile platforms. Its strategy is to attack each of the three costs with its own component: make appearance compact, prune what is redundant, and avoid recomputation that a moving scene does not actually need.

For appearance, a Monte Carlo Specular Energy Aggregator compresses high-order radiance residuals into first-order Spherical Harmonics. It works alongside an Attribute-Conditioned SH Enhancement module, whose predicted offsets are pre-baked before inference, so that work is done ahead of time rather than on the device. Against redundancy, a Multi-View Alpha-Based Densification and Pruning strategy suppresses unnecessary primitives while keeping the scene consistent across viewpoints.

Dynamic scenes get a compact explicit 4D representation. It is built from second-order Gaussian motion, learnable temporal support, and a binary partition separating static from dynamic content. Together these allow continuous-time modeling with no runtime deformation network at all. The partition then pays off a second time. A Depth-Order Certificate selectively reuses depth orders that were already committed, which cuts down on re-projection, sorting, merging, and index-buffer updates during playback.

Extensive experiments cover both static and dynamic scenes. The reported result is that Mobile-4DGS substantially reduces storage and rendering overhead while maintaining competitive visual quality. The abstract frames the outcome as an enabling one: real-time 3D and 4D Gaussian Splatting running on mobile devices. The authors state that code has been released.

Read as a whole, the design is a series of decisions about when work happens. Appearance detail is folded into a compact form and enhancement offsets are baked in beforehand. Motion becomes an explicit representation instead of a network evaluated each frame. And once the scene knows which parts are static, ordering decisions already made can be kept rather than redone. Each choice moves cost away from the moment of rendering, which is the moment a phone can least afford it.