Continuous Scene-Scale 3DGS Representation

Continuous scene-scale rendering requires comparable scales across viewing conditions. Common resize-based settings overlook focal-length and depth differences, assigning the same scale to observations with different sampling densities and compromising supervision and evaluation. We establish unified scene scale by combining image resize factors with robust, scene-normalized projection-density estimates, yielding a shared continuous coordinate that approximately aligns sampling density across cameras. Building on this formulation, we introduce ContinuousGS, a 3D Gaussian Splatting (3DGS) framework that adapts one representation across scene scales. Target-first sampling balances supervision over supported log-scale intervals, while anchored appearance and opacity residuals enable continuous adaptation with exact recovery of base attributes at the global finest supported scale. Matched-scale multi-view evidence regularizes appearance residuals and guides detail allocation. Our evaluation protocol aggregates cross-view performance along the shared scale axis, reporting full-range and high-support results separately. Experiments on four datasets demonstrate that ContinuousGS improves rendering quality across scene scales compared with representative methods while largely preserving native-resolution quality.