LiteVGGT addresses the computational and memory bottlenecks of the Visual Geometry Grounded Transformer (VGGT) for large-scale 3D reconstruction. It achieves significant speedups and memory reductions by introducing a geometry-aware cached token merging strategy that preserves critical geometric information.

C3G is a novel feed-forward framework that efficiently reconstructs and understands 3D scenes from sparse multi-view images by estimating a compact set of 2,048 3D Gaussians. This approach significantly reduces memory overhead and improves multi-view feature aggregation compared to dense, per-pixel methods, leading to superior performance in novel view synthesis and 3D open-vocabulary segmentation.