Gaussian spacing helps 3D scenes render across scales
SplatAnything: Gaussian Spacing Normalization for Feed-forward 3DGS across Arbitrary Scenes · arXiv:2611.08264
In plain words
SplatAnything is a method that builds a 3D scene from pictures by placing many small, soft-edged shapes in space. It first estimates the scene's geometry, then sets a shared scale based on the typical distance between neighboring shape centers before predicting their sizes and appearance.
Across seven varied datasets, its rendered views scored better on measures of appearance and detail than earlier methods in most tested settings.
The result is not universal: one image-error measure favored another method in several settings, whose renderings also showed artifacts at large viewpoint changes.
The research
A 3D reconstruction can have the right broad shape yet leave gaps or blur details when its predicted shapes are poorly sized for the scene.
A scale based on typical spacing between those shapes offers a way to make size predictions more consistent across different scenes and camera setups, a challenge reflected in the method's multi-dataset results.
SplatAnything is a feed-forward 3D scene reconstruction method that sets scale from typical Gaussian-center spacing before predicting Gaussian sizes.
The seven-dataset evaluation compares rendered target views with reference images across scenes and viewpoints, with and without camera poses.
DISTS and LPIPS measure image appearance and detail, with lower scores better; averaged across 4 to 64 input views, SplatAnything performs strongest on these measures in most datasets and settings.
A separate comparison finds that normalizing spacing before size prediction scores better than radial scaling or correcting sizes afterward. This supports the explanation that a consistent spacing reference helps the predictor generalize across the tested scenes.
The comparison covers seven datasets and selected extrapolated views. ReSplat has the highest average PSNR in several settings, though its renderings can show severe artifacts under large viewpoint changes. SplatAnything may also degrade with extremely sparse inputs or leave unseen areas incomplete.
What the evidence shows
- The scale prediction problem
SplatAnything is a feed-forward 3D Gaussian Splatting method: it predicts a 3D scene representation from images, with camera poses optional. The central issue is how to give the predicted Gaussian shapes sizes that cover surfaces across different scene scales and camera setups.
Existing radial-distance scaling uses a scene-level distance from a reference camera, which can leave the relevant center spacing inconsistent.
- A spacing-based reference
The diagnostic comparison shows the mismatch directly: after radial-distance scaling, median Gaussian spacing varies by up to 4.5-fold across scene categories. Similar-looking image regions can therefore correspond to different size targets.
GSNorm addresses that mismatch by choosing a scale from median Gaussian spacing itself, making the reference for size prediction more consistent across scenes.
- Results across varied scenes
The seven-dataset evaluation tests rendered target views across varied scenes and viewpoints, both with and without camera poses supplied to methods. DISTS and LPIPS compare rendered images with reference images, reflecting appearance and detail; lower values are better.
Averaged over 4 to 64 input views, SplatAnything performs strongest on these measures across most datasets and settings, supporting its generalization within the tested coverage.
- Why timing matters
A normalization comparison tests whether choosing the reference before prediction matters. At four input views, GSNorm's DISTS score is 0.198, compared with 0.223 for RDNorm under the same architecture and training settings.
The full comparison also finds RDNorm followed by size calibration consistently inferior to GSNorm, which supports the authors' view that changing predicted sizes afterward does not fix ambiguity in the original size targets.
- A bounded quality claim
The result depends on what is measured. ReSplat has the highest average PSNR in several settings, but its images can show severe artifacts under large viewpoint changes; the paper notes that PSNR can favor overly smooth images.
SplatAnything's advantage on appearance and detail measures is broad across the reported comparisons, not a win on every metric or setting.
Deep
- Original paper authors
- Hao Yu, Zhenwei Wang, Yongzhen Hu, Yuning Gong, Junta Wu
- Submitted to arXiv
- License
- Not specified by source