3D Gaussian Splatting (3D-GS) has emerged as a powerful representation for reconstructing 3D scenes from posed image collections, enabling high-quality novel view synthesis at real-time rendering rates. However, standard 3D-GS reconstructions remain difficult to use in downstream settings: they are often too large to store efficiently, slower than necessary to render, and limited by the pixel resolution of the input images.
This defense presents three methods that make 3D-GS models smaller, faster, and more flexible. The first is a principled post-hoc pruning method, PUP 3D-GS, that scores the sensitivity of each Gaussian and enables aggressive compression, removing 90% of Gaussians while preserving high visual fidelity. The second is a rendering acceleration framework, Speedy-Splat, that corrects an inefficiency in Gaussian localization and integrates pruning directly into training, achieving an average rendering speed-up of 6.71x across real-world scenes. The third is a selective super-resolution framework, SplatSuRe, that generates higher-than-pixel-resolution 3D-GS reconstructions by applying super-resolved supervision only where scene geometry and camera pose indicate that additional high-frequency detail is useful.
Together, these contributions advance 3D Gaussian Splatting toward practical downstream use as an efficient, compact, and high-quality 3D scene representation.