acceptodds
Withdrawn from ICLR 2027

Exploration Matters for Escaping the Blur Trap in 3D Gaussian Splatting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

3D Gaussian Splatting (3DGS) employs Gaussian primitives for explicit scene representation, facilitating real-time, high-fidelity reconstruction and novel view synthesis of complex scenes. However, the explicit modeling inherent in 3DGS introduces a gradient bias during optimization, rendering its non-convex optimization process highly susceptible to convergence toward local suboptimal solutions. This constitutes a fundamental limitation in 3DGS optimization, which we term the Blur Trap. To address this limitation, we integrate simple explicit exploration into the 3DGS optimization framework. First, through rigorous mathematical analysis of the 3DGS optimization formulation, we identify the underlying optimization bias responsible for the Blur Trap and categorize it into two distinct subtypes: the Far-Side Blur Trap and the Near-Side Blur Trap. Subsequently, we propose two highly straightforward exploration strategies (Random Seeding and Random Splitting) to mitigate the far-side and near-side blur traps, respectively. Experimental validation demonstrates that the incorporation of these exploration operators effectively and complementarily overcome the Blur Trap, achieving high-quality rendering performance across multiple datasets.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Exploration, Gradient Bias, Non-convex Optimization, Neural Rendering

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_3QAiOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the problem of blur issue of 3D Gaussian Splatting (3DGS), which called "blur trap" in this paper. The blur trap comes from two main reasons: (1) the orthogonality between the dominant position gradient and the viewing ray, which limits depth optimization, particularly for distant Gaussian primitives; and (2) gradient attenuation caused by alpha-blending, which weakens the densification signal for occluded Gaussian primitives. To solve the problem, this paper proposed two simple exploration strategies: Random Seeding and Random Splitting. Random Seeding generates some random gaussian primitives at randomly sampled 3D positions. Random Splitting randomly selects existing Gaussian primitives for splitting, with a preference for larger ones. The experiment result shows this paper achieves overall better novel view synthesis result compared with other baselines.

Critical strengths
  1. The paper provides a detailed analysis of potential causes of blurring artifacts in 3DGS, supported by both theoretical derivations and experimental observations.
  2. The paper conducts extensive experiments across multiple datasets and demonstrates competitive or improved novel view synthesis performance compared to existing methods.
Critical weaknesses
  1. Though this paper devotes substantial effort to analyzing the reasons of blur, the proposed solution essentially relies on randomly introducing or splitting Gaussian primitives during training. However, neither Random Seeding nor Random Splitting explicitly identifies or targets the specific failure modes discussed in the theoretical analysis. Random Seeding does not identify regions suffering from insufficient depth gradients, and Random Splitting does not explicitly account for gradient attenuation caused by occlusion. In my opinion, the paper somewhat overstates the relationship between its theoretical findings and methodological contributions. Given the simplicity of the proposed approach, I find the methodological novelty and overall contribution somewhat limited.
  2. Adding more gaussian primtives would naturally increase the performance. Therefore, it is important to distinguish improvements resulting from the proposed exploration strategies from those caused by changes in the number and distribution of Gaussian primitives. Although Table 9 reports the final number of Gaussian primitives and shows that the proposed method does not consistently increase the primitive count, the paper also modifies the Gaussian pruning threshold, which may affect both the number and distribution of the retained primitives. Further experiments under the same pruning settings and comparable Gaussian budgets are needed to better validate the effectiveness of the proposed strategies.
Questions for authors

Please refer to "critical weakness".

Reviewer_eR5hOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates why 3DGS tends to produce blur in distant and occluded regions. It links these artifacts to weak depth-directed optimization signals and occlusion-attenuated gradients that hinder densification. The authors introduce two simple exploration strategies:

  • random seeding, which adds gaussian points at new locations,
  • random splitting, which splits existing gaussians independently of gradient thresholds.

Experiments across five datasets demonstrate improved novel-view rendering quality, although these gains can come with increased training time and Gaussian counts.

Critical strengths

The proposed method itself is intuitively simple and easy to follow. The heuristic random splitting/seeding tactics do not require introducing new networks, and might be easy to integrate to other methods.

Critical weaknesses

One of my major concern is that, the results of this paper is not very convincing. While I like how the authors utilize the gradients to generate new points, this method introduces significantly more gaussian pts compared to traditional methods. In addition, it is a relatively trivial insight that adding # of pts will be benifical to PSNR.


Also, the reported results of this paper are not good enough. The major contribution of this paper is about tactics which densify/redistribute the gaussians during training. The authors should be aware that there are already many works working on this issue. However, there seems to be no comprehensive baseline comparision in this paper.

For example, PCGS: Deblurring 3D Gaussian Splatting with Patch Comparison (ICML 2026) also focuses on how to densify 3DGS. I compared their result with this paper, and it seems that PCGS has better results on Mip-NeRF360, Tanks & Temples, and Deep Blending. Additionally, PCGS keeps the # of GS smaller than original 3DGS, and trains faster than original 3DGS.

However, I understand that this paper may have different experiment settings than PCGS paper, so for now I won't say this paper is not good enough. So I would like to suggest that, the authors may want to compare their method to PCGS (and more baselines, if possible), or at least discuss why their method is better than PCGS in some aspects.

For now, I am not sure if the results of this paper is solid/good enough to be accepted by ICLR. I will appreciate it if the authors could provide more experments. I will have a more positive view of this paper if the authors could better validate the advantage of their method.

Questions for authors

See Critical Weaknesses.

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

This paper investigates two reconstruction failures in 3D Gaussian Splatting (3DGS): the Far-Side and Near-Side Blur Traps. Through gradient analysis, the authors show that 3D positional updates are dominated by screen-space gradients, which provide no direct depth optimization signal, while alpha-blending suppresses densification gradients for occluded Gaussians. To address these issues, they propose two stochastic exploration operators: Random Seeding, which injects Gaussians into unexplored regions, and Random Splitting, which probabilistically splits Gaussians based on their scales. Experiments cover five static 3D reconstruction benchmarks and dynamic scenes using 4DGS.

Critical strengths
  1. Clear Diagnostic Insights: The gradient decomposition and transmittance-fixing experiments clearly illustrate how screen-space dominance and alpha-blending impact optimization.
  2. Extensive Benchmarking: Comprehensive evaluations across a wide range of static datasets and dynamic scenes.
Critical weaknesses
  1. Confounded Gains from ADC Reset: The near-side fix relies on a hidden side effect. Running Random Splitting at split interval = 1 resets the standard adaptive density control (ADC) accumulator every step, which drives most of the performance gains rather than stochastic splitting itself.
  2. Contradictory Operator Contributions: Region-wise analysis shows that Random Splitting alone outperforms Random Seeding even in far-field regions, undermining the claim that both operators are complementary and uniquely tailored to separate traps.
  3. Geometric Flaw in Random Seeding: Under zero parallax, distant views suffer from scale-depth ambiguity (producing identical 2D footprints at varying distances). It is unclear whether Random Seeding recovers meaningful far-field geometry or merely improves appearance fitting using large background Gaussians.
Questions for authors
  1. What are the metrics if Random Splitting is applied without resetting the standard ADC gradient accumulator? Conversely, how does standard 3DGS perform if xyz_gradient_accum is simply accumulated/averaged only over views where the primitive's transmittance Tm−1>τTT_{m-1} > \tau_T (or using max-pooling across views)?
  2. What is the performance of Standard 3DGS with only the 10×10\times pruning threshold relaxation, but without Random Seeding? Additionally, out of the 2,900 seeds injected during training (Nseed=20N_{seed}=20 every 100 steps from 500 to 15,000), what fraction survives until iteration 30,000?

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

Related papers

Discussion (0)

Sign in to comment.