This paper addresses the architectural limitations of InstantSplat—a pose-free and lightweight framework for real-time 3D reconstruction—when scaled to dense multi-view settings. While InstantSplat is effective under sparse-view inputs, it suffers from memory inefficiency, optimization instability, and degraded rendering performance as the number of views increases. To overcome these challenges, we propose a modular set of architectural enhancements designed to maintain the core strengths of InstantSplat while improving its scalability. Our contributions include: (1) a feature-aware Gaussian initialization to reduce redundancy, (2) iterative pose refinement with geometric consistency, (3) view-aware pruning and clustering for Gaussian management, (4) log-domain scale optimization with regularization for stability, and (5) tile-based rasterization to boost rendering efficiency. Although this work is conceptual and lacks empirical evaluation, the proposed modules offer a flexible and extensible blueprint for adapting InstantSplat to high-density scenarios. We aim to bridge the gap between real-time neural rendering and large-scale 3D reconstruction using Gaussian Splatting.
