InceptionGS: Generative Bootstrapping for Large-Scale Gaussian Splatting under Unstructured View Sampling

Achieving truly immersive large-scale scene digitization ne-cessitates consistent and visually pleasing rendering across all possibleviewing perspectives. However, collecting multi-view images covering ev-ery fine detail of a large-scale scene is prohibitive due to scene complexity,capture cost, negligence, or accessibility constraints. As a result, the sam-pled views tend to be highly unstructured – the majority of the sceneis well covered yet certain regions inevitably lack sufficient observations.Existing reconstruction based methods are vulnerable to view scarcitywhile generation based approaches suffer from generalization, controlla-bility, and 3D consistency issues. To address this challenge, we proposeInceptionGS, which bootstraps Gaussian splatting by subtly balancingreconstruction and generation. Starting from an initial Gaussian splat-ting, InceptionGS reasonably rethinks and repairs problematic regionscaused by view scarcity while preserving the quality elsewhere, by softlyincorporating scene- and view-adaptive generative priors. Extensive ex-periments on real-world large-scale scenes demonstrate the superiorityand broad applicability of our approach in handling unstructured im-agery and boosting high-fidelity Gaussian splatting.