Creating a 3D scene from a handful of photos typically requires lengthy per-scene optimization. Feed-forward 3D Gaussian Splatting avoids this by predicting 3D blobs (Gaussians) from images with a single model, but can fail when scene size or camera setup changes, causing holes or blur.
The authors found existing normalization, which adjusts geometry based on average scene radius, fails to account for spacing between neighboring blob centers. Their Gaussian Spacing Normalization scales the scene based on a robust statistic of blob spacing, creating a consistent reference for prediction across scenes.
This method, implemented in their system SplatAnything, provides unambiguous learning targets for the model. It achieved top performance in rigorous standard tests, with strong gains under variations in scene scale and viewpoint, marking a step toward reliable instant 3D reconstruction.