GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens

TL;DR

GlobalSplat introduces a novel 3D Gaussian Splatting framework that reconstructs scenes using global scene tokens instead of pixel-aligned primitives, achieving competitive rendering with only 16K Gaussians (4MB footprint) and 78ms inference. The approach avoids representation bloat via coarse-to-fine training and global scene awareness. Benchmarked on RealEstate10K and ACID, it is significantly faster and more memory-efficient than dense baselines.

• •Global scene tokens replace local pixel-aligned primitives, reducing representation bloat and improving multi-view consistency
• •Achieves 16K Gaussians with 4MB footprint and 78ms inference—significantly smaller and faster than dense baselines
• •Coarse-to-fine training curriculum prevents redundancy while maintaining competitive novel-view synthesis fidelity

Generated with AI, which can make mistakes.