SplatAnything: Gaussian Spacing Normalization for Feed-forward 3DGS across Arbitrary Scenes

SplatAnything addresses a fundamental limitation in existing feed-forward 3D Gaussian Splatting (3DGS) methods: although radial distances are often normalized, the actual spacing between neighboring Gaussian centers varies widely across diverse scene scales and camera layouts, causing systematic inconsistencies in scale targets and ambiguous supervision. To fix this, the authors propose Gaussian Spacing Normalization (GSNorm), which standardizes a robust scene-level statistic of Gaussian spacing to act as a consistent spatial reference. Evaluated across a newly established benchmark covering diverse datasets, sparse inputs, and extrapolated viewpoints, SplatAnything demonstrates state-of-the-art performance, especially under severe scale and viewpoint variations. As a preprint, the method offers practical utility for fields like autonomous driving and AR/VR, though independent replication and full ablation details should be verified from the complete paper text.