acceptodds
Under review as a conference paper at ICLR 2027

SplatAnything: Gaussian Spacing Normalization for Feed-forward 3DGS across Arbitrary Scenes

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.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Feed-Forward 3D Reconstruction

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_svmkOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the normalization space used for Gaussian attribute prediction in feed-forward 3D Gaussian Splatting. The authors argue that radial distance normalization does not canonicalize the spacing between Gaussian centers, which can make appropriate Gaussian scales vary across camera configurations and scenes. They propose Gaussian Spacing Normalization (GSNorm), using the median of depth divided by the geometric mean of the pixel-space focal lengths as a scene-level normalization factor. The reported results show broad improvements in perceptual quality, supported by normalization and training-design ablations.

Critical strengths
  1. A useful perspective on normalization for Gaussian prediction. The paper revisits the common practice of predicting Gaussian attributes in the normalized coordinate space inherited from a geometry backbone. Its observation that canonicalizing radial distance does not necessarily canonicalize Gaussian spacing is relevant to feed-forward Gaussian prediction. Since appropriate Gaussian scales depend on the spacing between their centers, using a spacing-based normalization statistic provides an intuitive, geometrically motivated alternative.

  2. Broad empirical evaluation and informative visualizations. The evaluation covers seven datasets, both pose-free and pose-conditioned settings, and varying numbers of input views. The normalization and training ablations provide useful evidence for the effectiveness of the proposed system, while qualitative comparisons across different scene configurations help illustrate the practical effects of the normalization choice.

Critical weaknesses
  1. The role of physical scene scale requires a more precise explanation. The paper places considerable emphasis on generalization across scene scales, but absolute physical size alone cannot explain variations in Gaussian spacing after RDNorm. If scene geometry and camera translations are jointly scaled by the same factor, while camera intrinsics and image sampling remain unchanged, both local spacing and the radial-distance normalization factor scale proportionally. The normalized spacing therefore remains unchanged. Although the paper also discusses camera configurations, it does not clearly distinguish the effect of physical scene size from correlated factors such as camera placement and viewing distance. An operational definition of scene scale and a clearer account of these relationships would make the motivation more precise and clarify when GSNorm should be most beneficial.
Questions for authors
  1. Which scene and camera properties determine the limitations of RDNorm? Please clarify what "scene scale" means operationally and how it relates to normalized Gaussian spacing, given the invariance to joint scaling of scene geometry and camera translations described above. More specifically, which measurable properties of the scene or capture configuration predict when RDNorm becomes inadequate and when GSNorm offers the greatest benefit? A more explicit characterization would strengthen the paper's contribution beyond demonstrating that an alternative normalization works well.

  2. Can a focused controlled analysis better distinguish the sources of the improvement? A useful complement to the existing experiments would be a within-scene comparison that varies camera trajectory extent or the spatial distribution of observations, while controlling other settings where feasible. For example, uniformly distributed observations could be compared with observations concentrated in a subset of scene regions. Reporting both normalized spacing statistics and rendering performance would help connect the proposed explanation to the observed gains. A median radial-distance baseline would also help distinguish the benefit of normalizing a different geometric quantity from that of replacing the mean with a more robust statistic. These targeted comparisons would address my concern more directly than additional evaluations across scene categories.

1 comment
Reviewer_svmkOpenReview ↗

Office Comment by Reviewer

Thanks for the rebuttal. My concerns have been addressed. I will raise my score to weak acceptance.

Reviewer_VEH2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SplatAnything, which introduces Gaussian Spacing Normalization (GSNorm), which canonicalizes a robust scene-level statistic of Gaussian spacing to provide a more consistent spatial reference for Gaussian prediction. By normalizing Gaussian spacing rather than radial distance, the method provides a more consistent spatial reference across different scene scales and camera configurations. This paper also establishes a benchmark covering diverse datasets, input view counts, and extrapolated viewpoints. Experiments demonstrate state-of-the-art performance, particularly under significant variations in scene scale and viewpoint.

Critical strengths
  1. From Tables 1, 3, and Figures 5, 6, and 7 of the paper, it can be seen that the simple design changes have brought significant benefits, especially the impressive visualization effect.
  2. The evaluation includes several datasets that are relevant to real-world generalization, further demonstrating the effectiveness of the proposed method.
Critical weaknesses
  1. Gaussian Spacing Normalization is more like adjusting the point cloud output of the geometry foundation model, so it's just normalizing the initialization of the Gaussian center, right? This is not a normalization of the final Gaussian space. This makes this naming convention less accurate and reasonable.
  2. The comparison method is a bit unfair, for example, some methods in Table 1 (AnySplat, YoNoSplat, tokenGS, QueryPlat, etc.) were only trained on a small number of datasets, so the benefits cannot be attributed to your GSNorm. It is also possible that you used 14 datasets to train the performance growth is due to this data scaling. From the ablation experiments in Table 3, it appears that the performance improvement is not particularly significant.
Questions for authors
  1. Is the efficiency of this GSNorm high, and will it incur additional time and space costs
  2. Why is the PSNR performance of the method proposed in the paper not good enough, and what are the reasons for it?
Reviewer_s96FOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper observes that feed-forward 3DGS methods built on geometry foundation models predict Gaussian attributes in the radial-distance-normalized space (RDNorm) of the geometry backbone, where the spacing between neighboring per-pixel Gaussians varies widely across scene scales and camera configurations (up to 4.5× across scene types), making Gaussian-scale supervision inconsistent. It proposes Gaussian Spacing Normalization (GSNorm), which rescales each scene so that its median Gaussian spacing is the same across scenes before Gaussian attributes are predicted. To evaluate the resulting model, SplatAnything, the authors also introduce a seven-dataset benchmark that controls input sparsity and extrapolation difficulty, and report state-of-the-art LPIPS and DISTS in pose-free, pose-conditioned and single-view settings.

Critical strengths
  • Clear diagnosis. The paper identifies a concrete, under-discussed issue: Gaussian scales are predicted in a space (RDNorm) where neighbor spacing varies about 4.5× across scene types (Fig. 4) and also with resolution and FoV.
  • A simple, cheap fix. GSNorm is one scalar per scene, has no extra parameters, and could be added to other feed-forward 3DGS pipelines.
Critical weaknesses
  • The per-pixel footprint parameterization in prior work is not discussed or compared. pixelSplat (cited) and MVSplat (not cited) already multiply the predicted Gaussian scale by each pixel's depth divided by the focal length. In the notation of Eq. 4, their scale is σ_i(u) ∝ m_i(u) · d_i(u), where m_i(u) is a bounded multiplier predicted by the network. In other words, each Gaussian's base size is set to its own pixel footprint, and the network only predicts a correction relative to it. The paper's core insight, that Gaussian scale should be commensurate with local spacing, is therefore already built into earlier methods, and GSNorm can be seen as a per-scene version of the same idea: it normalizes by the median spacing of the scene rather than by each pixel's own spacing. The paper should compare GSNorm with RDNorm + per-pixel footprint scaling, and explain why a scene-level factor would be preferable.

  • A substantial part of the improvement comes from the training recipe rather than GSNorm. The GSNorm row of Table 3 matches the averaged "Ours (Pose Cond.)" rows of Table 1 exactly, so the ablation appears to use the pose-conditioned benchmark, which the paper should state. Comparing Mirror2-GS (Pose Cond.) in Table 1 with the RDNorm and GSNorm rows of Table 3 at 4–32 views, retraining with RDNorm improves DISTS by 0.024–0.029 and LPIPS by 0.022–0.025, and GSNorm adds a further 0.025–0.028 DISTS and 0.033–0.036 LPIPS. GSNorm therefore accounts for only about half of the total gain over Mirror2-GS; the other half comes from the training recipe (14 training datasets, the view-sampling strategy and the fine-tuned Gaussian backbone). At 64 views GSNorm's share drops further (DISTS 0.007 vs. 0.033 from the recipe; LPIPS 0.021 vs. 0.037), which the paper doesn't explain. Reporting the RDNorm-retrained model in Table 1 and giving per-dataset ablation results would clarify how much GSNorm itself contributes.

  • PSNR results are weaker, and the explanation is not supported. Table 10 shows that without poses YoNoSplat outperforms SplatAnything on PSNR at 4–32 views (e.g., 18.34 vs. 17.85 at 32 views), and with poses ReSplat outperforms it at 4–32 views by about 0.5–1.0 dB. SplatAnything is best only at 64 views. The authors attribute this to PSNR favoring overly smooth predictions, but this explanation is not quantified. Table 10 also reports only PSNR averaged over datasets; per-dataset PSNR and SSIM should be reported to show where the gap comes from.

Questions for authors
  • How much of the gain over Mirror2-GS comes from GSNorm vs. the data and training recipe? Please add the RDNorm-retrained model to Table 1, and explain why the GSNorm gain shrinks at 64 views.
1 comment

I thank the authors for their detailed rebuttal and additional experiments. It helped me to clarify my understanding of the paper and addressed my concerns. I have a few remaining suggestions.

  • Per-pixel footprint parameterization, I suggest that the authors briefly discuss existing approaches (e.g., pixelSplat) in the paper and clarify how they differ from the proposed method.

  • PSNR evaluation: I find the authors' explanation reasonable. However, for clarity and transparency, I encourage the authors to report per-dataset PSNR (and SSIM, if possible) for each view configuration in the appendix. It would also be helpful to include additional qualitative examples illustrating why PSNR alone may not adequately reflect rendering quality under view extrapolation.

Given the clarifications, I have raised my rating to 3.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.