Paper review · November 10, 2026
SplatAnything: Gaussian Spacing Normalization for Feed-forward 3DGS on Arbitrary Scenes
SplatAnything: Gaussian Spacing Normalization for Feed-forward 3DGS across Arbitrary Scenes
The authors normalize scene scale by the median spacing between Gaussian centers (GSNorm), producing a feed-forward 3DGS model that generalizes well across different scene sizes and camera setups.
Problem
Feed-forward 3DGS uses the coordinate space produced by a geometry foundation model as is. This space is normalized by the mean radial distance of the points (RDNorm). As a result, the spacing between neighboring Gaussian centers varies widely with scene scale and camera setup. The authors report that even after RDNorm, the median Gaussian spacing differs by up to 4.5x across scene types. In effect, regions that look similar are given different ground-truth Gaussian sizes, which makes the training signal ambiguous. The result is holes or excessive blur in renderings.
Method
With one Gaussian per pixel, Gaussian spacing is roughly depth divided by focal length (D/√(fx·fy)). GSNorm computes this value at every valid pixel and then rescales the whole scene by the median. This sets the representative spacing to 1 while preserving relative spacing differences within the scene. Geometry is estimated with a frozen WorldMirror 2.0. The backbone of the same model is fine-tuned as the Gaussian backbone, and a lightweight GSHead is trained from scratch to predict per-pixel opacity, rotation, scale and SH coefficients. The losses are MSE and LPIPS. Training took 20,000 iterations on 64 H20 GPUs. The authors also built a new benchmark covering 7 datasets, 1–64 input views, and extrapolation viewpoints with input-target overlap ranging from 0.2 to 0.6.
Results
- In the pose-free setting with 4–64 input views, it achieved a DTU DISTS of 0.229, lower than the strongest baseline Mirror2-GS (0.347) (lower is better).
- In the same setting, DL3DV LPIPS was 0.156, lower than Mirror2-GS (0.188) and DA3-GS (0.228). It ranked first in DISTS and LPIPS on most of the 7 datasets.
- In the posed setting as well, it reached a Mip-NeRF 360 LPIPS of 0.440, better than the pose-conditioned version of Mirror2-GS (0.506) and ReSplat (0.525).
- With single-view input, it achieved a DL3DV DISTS of 0.279, better than SHARP (0.335). However, SHARP was better on ScanNet, NRGBD and Mip-NeRF 360, so the results were mixed.
Limitations
ETH3D LPIPS (pose-free) was 0.049, worse than QuerySplat (0.031) and YoNoSplat (0.033), so it does not lead on every metric. In the single-view setting, it trails the monocular-only model SHARP on some datasets. The Gaussian spacing calculation relies on a first-order approximation that depth is locally constant, and performance depends on the depth quality of the frozen geometry model (WorldMirror 2.0). The evaluation benchmark was designed by the authors themselves, and training requires 64 GPUs.
Why it matters
Techniques that build a 3D scene from a few images in seconds are used for large-scale 3D reconstruction and content creation. This paper improves generalization across scenes of different scales, such as indoor, outdoor and object scenes, by changing only the normalization criterion without changing the architecture. It can easily be tried in pipelines of the same family.
Key terms
- 3DGS(3D Gaussian Splatting)
- A method that represents a scene as a set of many 3D Gaussians (position, size, color, opacity) for fast rendering. The feed-forward approach predicts the Gaussians in a single network pass instead of optimizing each scene separately.
- Gaussian spacing
- The distance between Gaussian centers back-projected from neighboring pixels. To cover a surface without gaps, the Gaussian size must match this spacing.
- RDNorm / GSNorm
- RDNorm is the existing normalization that sets scene size by the mean radial distance of the points. GSNorm is the normalization proposed in this paper, which sets scene size by the median Gaussian spacing.
- Extrapolation viewpoints
- New viewpoints that lie outside the range of the input images rather than between them. This is a difficult evaluation condition that exposes the accuracy of the reconstructed geometry more clearly.
Original
Authors: Hao Yu, Zhenwei Wang, Yongzhen Hu, Yuning Gong, Junta Wu, Xiaowei Zhou, Hujun Bao, Tengfei Wang, et al. · Posted 2026-11-05 · License arXiv 비독점 배포 라이선스 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.