This page is an AI translation of the Japanese original. If anything differs, the Japanese version prevails.
SplatAnything Proposes Feed-Forward 3DGS That Normalizes Gaussian Spacing to Handle Arbitrary Scenes
zju3dv, a group linked to Zhejiang University, and colleagues have proposed GSNorm, which reduces the mismatch in training targets caused by scene scale and camera layout in feed-forward 3DGS. They say they have also built a benchmark that covers diverse data and extrapolated views.
This article was generated by AI and published automatically without prior human review.
- Announced
- 2026.11.05 (local time)
- Use cases
- Research, Production
- Sources
- 3 (en) (3 primary)
Key points
- The authors point out that existing methods predict Gaussian attributes in a space normalized by radial distance, so the spacing between neighboring Gaussians varies with scene scale and the supervision signal becomes ambiguous.
- The proposed GSNorm normalizes by a robust per-scene statistic of Gaussian spacing, aligning the spatial reference used for prediction.
- They built a benchmark covering diverse datasets, numbers of input views and extrapolated views, and the authors say the improvements are especially large under changes in scale and viewpoint.
Contents
What happened
According to the authors, radial-distance normalization only aligns the overall statistics of the reconstructed geometry, while the spacing between adjacent Gaussian centers varies widely with scene scale and camera configuration. As a result, the target values for Gaussian scale are systematically shifted, and the supervision signal during training becomes ambiguous.
How it works
For evaluation, they say they also built a systematic benchmark that includes multiple datasets, levels of input sparsity (number of input views) and extrapolated views (difficult viewpoints far from the inputs). Specific metric values, baselines and the network architecture cannot be confirmed from the available material.
Why it matters
The project page shows results in a SuperSplat-based viewer, and the HTML viewer is fetched from Hugging Face so users can interact with it. The setup appears to assume that outputs are checked on the web, but the material gives no details on the delivery format.
How to try it
- The code is listed on GitHub (zju3dv/splatanything). However, the README content could not be checked in the available material, so please check directly whether the code is fully released and whether weights are provided.
Caveats and open questions
- The body of the GitHub repository is not included in the material, so the release status of code and pretrained models, and the license, cannot be confirmed.
- "State-of-the-art" and "especially large improvements" are the authors' claims, and the material contains no figures or baselines.
- This article is based only on the paper abstract and the project page; the experimental results in the paper body have not been checked.
- The content of the comparison scenes on the project page and the evaluation conditions cannot be determined from the material.
Glossary
- Feed-forward 3DGS
- A method that outputs 3D Gaussians from input images through network inference alone, without per-scene optimization.
- GSNorm
- Gaussian Spacing Normalization. The method proposed in this paper, which normalizes by a scene-level statistic of the spacing between neighboring Gaussian centers.
- Extrapolated views
- Novel viewpoints far from the camera positions of the input images, which are difficult to reproduce.
Sources
Primary sources (publisher, paper, repository)
- SplatAnything: Gaussian Spacing Normalization for Feed-forward 3DGS across Arbitrary ScenesarXiv（arxiv.org）・en
- SplatAnything プロジェクトページzju3dv.github.io（zju3dv.github.io）・en
- zju3dv/splatanythingGitHub（github.com）・en
The X post that surfaced this story
This article was generated in Japanese by AI from the sources listed (primarily primary sources) and published automatically, without prior human review, after automated checks such as source matching (how we use AI). Figures and specifications are as stated by the publishers and have not been independently verified by us. Please consult the original sources. To report an error, use our contact form; we will review it promptly and correct it. This page is an AI translation of the Japanese original. If anything differs, the Japanese version prevails. Japanese original
From the operatorガウシアンスプラッティングとは（IZUTSUYA Inc.）