SplatAnything: Gaussian Spacing Normalization for Feed-forward 3DGS across Arbitrary Scenes
Hao Yu, Zhenwei Wang, Yongzhen Hu, Yuning Gong, Junta Wu, Xiaowei Zhou, Hujun Bao, Tengfei Wang, Chunchao Guo, Sida Peng
cs.CV
2026-11-06
SplatAnything swaps radial-distance normalization for median Gaussian spacing in feed-forward 3DGS, leading on all seven benchmark datasets and reconstructing 8-view scenes in 0.76 s.
Feed-forward 3DGS targets the painful part of conventional 3DGS: per-scene optimization that takes tens of minutes. A network emits all Gaussian parameters in one forward pass, in seconds. The dominant recipe runs a geometry foundation model (Depth Anything 3, WorldMirror and the like) to get multi-view depth and poses, then predicts the remaining per-pixel Gaussian attributes inside that model's coordinate system.
The trouble is the coordinate system itself. These models normalize with the mean radial distance of all reconstructed points to the reference camera (RDNorm). That canonicalizes how far points sit from the camera, not how far apart neighboring Gaussian centers sit, and the second quantity is what sets Gaussian size: covering a surface requires the scale to match local spacing. The paper measures up to 4.5x variation in median Gaussian spacing across scene types even after RDNorm, so identical local observations map to different scale targets in different scenes. A predictor biased small leaves holes at coarse spacing; one biased large over-smooths dense regions. That supervision ambiguity is the bottleneck for cross-scene generalization.
GSNorm in one sentence: derive the normalization factor from a representative statistic of Gaussian spacing, not from a global radial one.
At the system level, a frozen WorldMirror 2.0 supplies depth, intrinsics, and poses; after conversion to the GSNorm space, the unprojected points serve as fixed Gaussian centers. A Gaussian Backbone initialized from the geometry backbone (fine-tuned) plus a lightweight GSHead predicts per-pixel opacity, rotation, scale, and SH coefficients. Training spans 14 datasets with about 22,600 scenes, 20k iterations on 64 H20 GPUs, 4-32 input views, and resolutions sampled from 504 to 1024, with an AnySplat-style voxel merge handling overlap redundancy. The accompanying benchmark covers seven datasets, filters target views by symmetric co-visibility into a 0.2-0.6 overlap band, and stresses extrapolated views that interpolated-view protocols never test.
Averaged over 4-64 views across seven datasets:
| Setting | Dataset | Metric | Best baseline | SplatAnything |
| pose-free | DTU | DISTS | 0.347 (Mirror2-GS) | 0.229 |
| pose-free | ETH3D | DISTS | 0.188 (Mirror2-GS) | 0.156 |
| pose given | DTU | DISTS | 0.532 (Mirror2-GS) | 0.364 |
| single view | ETH3D | LPIPS | 0.133 (SHARP) | 0.104 |
The ablation isolates GSNorm's own share: same architecture, same training, LPIPS at 4 views falls from 0.278 (RDNorm) to 0.245, and at 64 views from 0.250 to 0.229. Camera-distance normalization collapses to DISTS 0.793 in the single-view setting, and depth normalization lands at 0.307, both behind GSNorm's 0.294. On one H20, the forward pass takes 0.76 s at 8 views and about 11 s at 64, more than three times faster than QuerySplat (39.6 s) and DA3-GS (35.9 s).
Two footnotes: in the pose-conditioned setting, ReSplat posts the highest PSNR throughout and keeps ETH3D LPIPS (0.027 vs 0.037), but its extrapolated views carry heavy artifacts, which is why the paper leads with LPIPS and DISTS. In the monocular setting the race with SHARP splits, with SplatAnything taking DISTS on 4 of 7 datasets.
Coordinate normalization is infrastructure: everyone inherits it, nobody audits it. This paper shows it decides whether Gaussian-scale supervision stays consistent across scenes, and the fix is one divisor, nearly free. Any pipeline shaped like a geometry foundation model plus a Gaussian head can adopt it, and that portability is worth more than SplatAnything itself. For practitioners, seconds-level reconstruction, 1 to 64+ views, optional poses, and coverage from objects to outdoor scenes move feed-forward 3DGS one step closer to a general-purpose tool; the benchmark, with controlled view density and explicit extrapolation difficulty, has standalone reuse value.
The system-level lead is not all GSNorm: the strong frozen geometry base, 22,600 training scenes, and fixed Gaussian centers each contribute. The isolated GSNorm gain is about 10% relative LPIPS, smaller than the system's margin over Mirror2-GS.
The authors' own list: performance degrades when the geometry model fails on extremely sparse or hard inputs, and only observed regions are reconstructed, leaving holes that generative priors would have to fill.
Open questions from reading: the benchmark protocol is the authors' own, so the strongest-method verdict arrives on home turf, even though all methods are evaluated identically. Average PSNR sits at 17-19 dB, which says extrapolated-view synthesis is far from solved for every method. The single-view race with SHARP is a split decision. And the 64-H20 training run sets a high replication bar: GSNorm transfers cheaply, the full system does not.