acceptodds
Under review as a conference paper at ICLR 2027

ProbeGS: Closed-Form Scale for Error-Guided Gaussian Densification

Abstract

A small Gaussian budget makes 3D Gaussian Splatting (3DGS) viable on low-capacity devices, but only if every new Gaussian has the right position and size. The standard photometric loss averages over pixels and can leave small, high-residual regions uncovered, yet the residual itself shows how large each correction should be: a broad residual calls for a wide Gaussian, a thin streak for a narrow, oriented one. ProbeGS therefore reads each new Gaussian's scale directly from the residual. It places corrective 2D Gaussians on the residual and lifts them to 3D. Around each pixel sampled in proportion to the residual, we measure how fast the residual decays on rings at several radii and set the standard deviation of the 2D Gaussian placed there in closed form, with no scale search. A short refinement moves only centers and colors, and a covariance lift inserts an ordinary 3DGS primitive that reproduces the fitted 2D Gaussian in its source view. We evaluate on Tanks & Temples, Deep Blending, and Mip-NeRF 360 at three requested Gaussian budgets. ProbeGS achieves state-of-the-art quality at 100K Gaussians, where it raises Deep Blending PSNR by 0.50 dB and lowers Tanks & Temples LPIPS by 14.5% over the next-best method. Its gain in fine detail concentrates where few training views see the surface.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, novel view synthesis, densification, differentiable rendering, neural rendering

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_d9npOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ProbeGS introduces a densification method for resource-constrained 3D Gaussian Splatting that estimates new primitives' scales from local image-residual decay, refines their centers and colors, and lifts them to 3D using a depth backbone. Experiments on three benchmarks report improved reconstruction quality, particularly at a 100K Gaussian budget.

Critical strengths
  • This paper targets an important question: resource/budget-constrained densification. The key insight is that selecting a high-error location does not determine an appropriate primitive size. Estimating scale from the residual is a useful extension of error-guided insertion.

  • The low-budget improvements are meaningful. At 100K Gaussians, Table 1 improves Deep Blending PSNR from 29.44 to 29.94 dB over ConeGS and reduces Tanks & Temples LPIPS from the best baseline's 0.282 to 0.241.

Critical weaknesses
  • The free-shape ablation does not isolate the benefit of constraining shape. From Section 5.3, it seems that the authors change both the initialization, from the analytical field to a fixed isotropic default, and the parameters allowed to optimize. Thus, Table 2 cannot distinguish the value of the initialization from the value of keeping shape tied to the residual field.

  • The efficiency comparison with other residual-fitting approaches establishes a narrower result than a quality-cost comparison. Table 3 measures adapted fitters at the same proposal caps, but does not report the reconstruction quality they achieve under a matched time budget or a less frequent insertion schedule.

  • The comparisons do not fully establish the claimed state of the art under a limited Gaussian budget. Although Table 1 includes recent methods such as ConeGS and EDGS, SteepGS (https://arxiv.org/abs/2505.05587) is cited without a quantitative comparison, and Mini-Splatting2 (https://arxiv.org/abs/2411.12788) appears only at its native Gaussian counts. GS2(https://arxiv.org/abs/2604.01884) is another relevant compact-representation method not covered in the main comparison. The analytical scale/shape contribution also needs positioning against SAD-GS (https://vcai.mpi-inf.mpg.de/projects/SAD-GS/), which uses multiscale image structure for anisotropic densification, although its input-frequency guidance and larger native models differ from the proposed residual-guided insertion. Additionally, some of the baseline configurations need clarification. For example, the EDGS row enables densification, but original EDGS(https://arxiv.org/abs/2504.13204) is designed to eliminate it.

Questions for authors
  1. Can you compare constrained and free-shape fitting with identical analytical initialization, proposal sampling, and optimization budgets? How much quality remains with K=0?

  2. For Table 3, what quality do the adapted fitters, particularly AuGS, achieve at the reported runtimes or under the same total fitting budget?

  3. Table 2 shows only a 0.01 dB loss from removing anisotropy at 100K, compared with 0.11 dB at 1M. Are these differences consistent across scenes or just random noise?

Reviewer_B7qMOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper addresses test-time optimization of 3D Gaussian Splatting for reconstruction. Specifically, they address the densification required during reconstruction to adaptively refine the scene. The authors claim that their densification method achieves better novel view synthesis (both qualitatively and quantitatively) with a fixed number of Gaussians than previous methods, and report particularly large gains under low Gaussian budgets.

They adapt a previous approach called ConeGS, which renders error images and places new Gaussians at pixels with high errors. To place Gaussians in 3D (initialization and densification), they use a reconstructed iNGP model (as in both ConeGS and this work). The contribution is a new initialization method for the Gaussians at densification. They perform a 2D fit to the error image to better fit the new Gaussian and its parameters, thereby reducing the image error.

To validate their method, they run their reconstruction method on multiple datasets for multiple Gaussian budgets and report visual metrics and timings.

They show that their method, in most cases, achieves improved PSNR, SSIM, and LPIPS compared to previous methods with the same Gaussian budget.

Critical strengths

The gains are consistent and improve the baseline ConeGS. The paper is well written and has nice, helpful visualizations.

Critical weaknesses
  1. Paper Motivation and Evaluation

The paper motivates its research and method with "A small Gaussian budget makes 3D Gaussian Splatting (3DGS) viable on low-capacity devices" but fails to evaluate or report on these claims. A smaller number of Gaussians does not necessarily mean better rendering performance (e.g. overdraw is the biggest bottleneck in most cases). The paper fails to show that their reconstructions actually improve the speed-to-quality curve on "low-capacity devices". Also, it's unclear what they actually want to address. Is it rendering speed, memory consumption, or storage size? This definitely needs to be addressed/improved.

  1. Flawed reconstruction setup

The method presented in the paper, as well as its predecessor ConeGS, starts with an already reconstructed iNGP model and uses it to perform 3DGS reconstruction and densification. This makes the whole line of work more of a baking than an actual reconstruction method since the whole scene is reconstructed twice. Compared to methods that only reconstruct once, this seems like a flawed experimental setup and evaluation. In general, I dislike that the method presents itself as a 3DGS reconstruction/densification method but fails to clarify (especially early in the paper/abstract) that it starts from an already reconstructed scene. I find this line of work and the idea for densification interesting but see it as misplaced as long as it relies on an existing reconstruction. Exploring the use of monocular depth estimators or the model's rendered depth to replace the iNGP seems like a more promising direction to me.

  1. Missing ablations

One central claim of the paper is that the residual-derived scale drives gains in reconstruction metrics, but it fails to ablate this effect in isolation. The ablations that do exist do not isolate the individual contributions, and those that do show only minor gains in some cases (e.g., 0.09 for free-shape path), which could be attributed to noise. The scale heuristic seems to be weakly justified. Since the densification method relies on randomness, the authors should report the mean and standard deviation (or a confidence interval) of gains and metrics across multiple reconstruction runs.

Questions for authors

Why do you rely on a pretrained iNGP and not use a monocular depth estimator or the rendered expected or median depth of the model? How does your densification method affect rendering performance, and what is the goal of the method paper (FPS, VRAM storage)?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.