SplatAnything releases inference code with Gaussian spacing fix
TL;DR
- SplatAnything identifies that feed-forward 3DGS predicts Gaussians in a radial-distance-normalized space whose actual spacing still drifts with scene scale and camera setup.
- The proposed fix, Gaussian Spacing Normalization or GSNorm, canonicalizes a scene-level spacing statistic as a more consistent spatial reference for Gaussian prediction.
- The authors claim state-of-the-art results with particularly strong gains under scene-scale and viewpoint shifts, and have released inference code and weights.
SplatAnything zeroes in on a specific failure mode in feed-forward 3D Gaussian Splatting: Gaussian attributes get predicted in a radial-distance-normalized space that holds one global statistic steady while "the spacing between neighboring Gaussian centers can still vary substantially across scene scales and camera configurations." The preprint argues this "mismatch induces systematic shifts in Gaussian-scale targets, leading to ambiguous supervision."
The proposed fix, Gaussian Spacing Normalization (GSNorm), "canonicalizes a robust scene-level statistic of Gaussian spacing to provide a more consistent spatial reference for Gaussian prediction." On a self-assembled benchmark across datasets, sparsity levels and extrapolated viewpoints, the authors report "state-of-the-art performance, with particularly strong gains under substantial variations in scene scale and viewpoint."
Inference code is live on GitHub, with example scenes spanning objects, humans, indoor, street, aerial, autonomous-driving rig captures and 360-degree panoramas; weights sit on Hugging Face. The abstract publishes no per-benchmark numbers.
Originally reported by paper
Read the original article →Original headline: SplatAnything Ships Inference Code for Feed-Forward 3DGS Across Arbitrary Scenes