Fast and Slim Splat: Efficient 3D Gaussian Splatting via Confidence and Opacity-Guided Caching and Sampling

3D Gaussian Splatting (3DGS) has garnered significant interest for its ability to deliver high-speed rendering and photorealistic 3D reconstruction. However, its performance is heavily influenced by the quality and density of the initial point cloud. Existing approaches primarily rely on COLMAP-based Structure-from-Motion (SfM) pipelines, which incur additional computational cost when handling dense multi-view inputs. Meanwhile, recently proposed feed-forward models based on DUSt3R are optimized for a limited number of views and face scalability issues as the number of input images increases. In this work, we adopt FASt3R, a feed-forward framework capable of processing numerous images in a single pass, to efficiently generate a dense and consistent point cloud. Based on this initialization, we introduce three strategies to enhance training efficiency: (1) confidence-based filtering to prioritize uncertain points during early optimization, (2) selective caching of high-confidence points to accelerate convergence, and (3) opacity-guided sampling to prune less expressive points during later training stages. The experimental results demonstrate that our method achieves a 24% reduction in training time while maintaining competitive reconstruction quality. This efficiency enables scalable and practical 3DGS across a wide range of multi-view input settings.