Abstract
We analyze the performance of novel view synthesis methods in challenging out-of-distribution (OOD) camera views and introduce SplatFormer, a data-driven 3D transformer designed to refine 3D Gaussian splatting primitives for improved quality in extreme camera scenarios.