SplatAnything: Gaussian Spacing Normalization for Feed-forward 3DGS across Arbitrary Scenes
Normalizing the spacing between Gaussian centers gives 3D prediction a more consistent reference across scene scales and camera setups.
The authors identify a mismatch in existing feed-forward 3D Gaussian splatting methods: normalizing radial distance still leaves neighboring Gaussian centers unevenly spaced. That variation shifts the Gaussian-scale targets used in training, making supervision ambiguous. Their GSNorm method standardizes a robust scene-level spacing statistic to address this. Across a benchmark covering diverse datasets, sparse inputs and extrapolated viewpoints, the authors report state-of-the-art results, with particularly strong gains when scene scale and viewpoint vary substantially.
HF Daily Papers' note
The authors identify a mismatch in existing feed-forward 3D Gaussian splatting methods: normalizing radial distance still leaves neighboring Gaussian centers unevenly spaced. That variation shifts the Gaussian-scale targets used in training, making supervision ambiguous. Their GSNorm method standardizes a robust scene-level spacing statistic to address this. Across a benchmark covering diverse datasets, sparse inputs and extrapolated viewpoints, the authors report state-of-the-art results, with particularly strong gains when scene scale and viewpoint vary substantially.
HF Daily Papers' note
score 4