SplatAnything
Feed-Forward 3D Gaussian Splatting

SplatAnything

Gaussian Spacing Normalization for Feed-Forward 3DGS across Arbitrary Scenes

Authors
Affiliations
Paper Code HuggingFace Demo

TL;DR

Demo Video

demo_video.mp4
Demo video coming soon

Method

SplatAnything pipeline: input images pass through a frozen geometry model, Gaussian Spacing Normalization, and a trainable Gaussian model to produce a 3DGS reconstruction.

(a) Overall pipeline

Full size

Architecture of SplatAnything. Given N input images and optional camera parameters, a frozen geometry foundation model first estimates multi-view geometry. We convert the geometry from its native radial-distance-normalized space to the proposed Gaussian-spacing-normalized space and predict per-pixel Gaussian attribute through a Gaussian backbone and a lightweight GShead.

Radial distance normalization leaves different Gaussian spacings and inconsistent scale targets across scenes; Gaussian Spacing Normalization aligns a representative spacing statistic to alleviate supervision ambiguity.

(b) Normalization space for feed-forward 3DGS

Full size

Radial Distance Normalization. Fixing the mean radial distance to the reference-camera origin does not align Gaussian spacing. Similar image regions may therefore require different Gaussian-scale targets as scene scale or camera configuration changes, making supervision ambiguous.

Gaussian Spacing Normalization. Canonicalizing the median Gaussian spacing gives predictions a more consistent spatial reference across scenes. This alleviates ambiguity in Gaussian-scale supervision, helping avoid under-coverage and excessive overlap.

Qualitative Results

Scene 01
Drag to orbit · Scroll to zoom · Fly: WASD + Mouse, Esc to exitViewed with SuperSplat

Click Explore in 3D to download this scene’s HTML viewer from Hugging Face and explore it interactively.

Acknowledgements

Citation

splatanything.bib