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 and 2 others
Zhejiang University · Tencent Hunyuan · Causa Robotics
Published November 5, 2026 · on the feed November 9, 2026 · ▲ 70 · arXiv 2611.08264
The abstract
We identify a key limitation of existing feed-forward 3DGS methods: Gaussian attributes are typically predicted in a radial-distance-normalized space that canonicalizes a global radial statistic of the reconstructed geometry, while the spacing between neighboring Gaussian centers can still vary substantially across scene scales and camera configurations. This mismatch induces systematic shifts in Gaussian-scale targets, leading to ambiguous supervision. To address this, we propose SplatAnything with Gaussian Spacing Normalization (GSNorm), which canonicalizes a robust scene-level statistic of Gaussian spacing to provide a more consistent spatial reference for Gaussian prediction. To rigorously assess our method, we further establish a systematic benchmark spanning diverse datasets, input sparsity levels, and challenging extrapolated viewpoints. Extensive experiments demonstrate that SplatAnything achieves state-of-the-art performance, with particularly strong gains under substantial variations in scene scale and viewpoint.
The Brief
1:53A short podcast summary of the paper. Written and voiced by AI.
Transcript
SplatAnything. Gaussian Spacing Normalization for Feed-forward 3DGS across Arbitrary Scenes. By Hao Yu and nine others, from Zhejiang University, Tencent Hunyuan, and one other institution. Published on the fifth of November.
Feed-forward 3DGS methods usually predict Gaussian attributes in a space normalized by radial distance, which canonicalizes one global radial statistic of the reconstructed geometry. The authors point out what that leaves loose. The spacing between neighbouring Gaussian centers can still vary a great deal across scene scales and camera configurations. That mismatch shifts the targets for Gaussian scale systematically, and the supervision becomes ambiguous.
SplatAnything introduces Gaussian Spacing Normalization, or GSNorm. Instead of a radial statistic, it canonicalizes a robust scene-level statistic of Gaussian spacing itself. The aim is a spatial reference for Gaussian prediction that stays more consistent from one scene to the next.
The change is in what gets normalized. Under the radial scheme, two scenes can look alike by the global statistic while their Gaussian centers sit at very different distances from one another, and the network is asked to predict scales against that inconsistent backdrop. GSNorm makes the spacing between Gaussian centers the thing that is canonicalized, so the reference the scale targets are measured against stays steadier. To test this rigorously, the authors also build a systematic benchmark that spans diverse datasets, different levels of input sparsity, and challenging extrapolated viewpoints.
Extensive experiments report state-of-the-art performance for SplatAnything, with particularly strong gains when scene scale and viewpoint vary substantially. The benchmark's mix of sparse inputs and extrapolated views is where those variations are put under pressure.
Those are the conditions the diagnosis singled out, since spacing drift was tied to changes in scale and camera setup. The contribution is a choice of reference frame: normalize by the statistic that the prediction actually depends on.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
One network, any scene: a cup or a city. Here is what keeps tripping it up.
You are building a network that turns a handful of photos into a cloud of 3D Gaussian blobs in a single pass, with no per-scene optimisation. It should work on a teacup, a living room or a city street.
The road ahead
- 1 · The problem ~6 min Rescaling scenes by average distance to the camera leaves the gaps between neighbouring Gaussian blobs uneven, so the network cannot learn consistent blob sizes.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Rescale each scene so its median gap between neighbouring Gaussian centres equals one, giving blob sizes a consistent reference across scenes.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min A frozen geometry model gives depth, GSNorm rescales it by median spacing, and a fine-tuned network predicts per-pixel Gaussians, trained on rendered-image losses.
- 6 · The experiment ~6 min You’ll guess: Where did SplatAnything's gains over prior feed-forward methods turn out strongest?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.