acceptodds
Under review as a conference paper at ICLR 2027

VanishGS: Dual-Path Modeling of Primitive Disappearance for Pruning 3D Gaussian Splatting

12 views

Abstract

While 3D Gaussian Splatting (3DGS) enables high-quality real-time novel view synthesis, its reliance on millions of Gaussian primitives leads to substantial storage and computational overhead. Existing post-pruning methods typically estimate primitive importance from rendering contribution, heuristic attributes, or generic parameter sensitivity, which do not directly measure the rendering distortion caused by removing a Gaussian, potentially leading to inaccurate importance estimation. In this paper, we propose VanishGS, a post-pruning framework that estimates Gaussian importance by the rendering distortion induced by its disappearance. To capture this effect, we model Gaussian disappearance through two paths, opacity dissolution and geometric collapse, characterizing its rendering impact from both appearance and geometric perspectives. We estimate the directional distortion along these pruning perturbations using the Gauss–Newton matrix and efficiently approximate the required diagonal entries through randomized projections. By measuring the cost of removing each Gaussian, VanishGS provides an importance criterion that is better aligned with the pruning objective. Extensive experiments across multiple benchmarks and pruning ratios demonstrate that VanishGS consistently preserves higher rendering quality. Moreover, VanishGS can be integrated into existing post compression pipelines, further demonstrating its generality as a plug-and-play pruning strategy.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Post-Pruning, Rendering Distortion, Importance Estimation.

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_H7wzOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

VanishGS scores Gaussian primitives for pruning by combining Gauss-Newton estimates of rendering distortion along opacity-dissolution and scale-collapse paths. The diagonal terms are estimated through random projections. The method is evaluated across various pruning ratios, post-pruning fine-tuning, and integration into existing compression pipelines.

Critical strengths
  • The empirical advantage is supported under several different protocols.

  • The score can replace the pruning criterion without changing the scene representation or renderer. This work can be integrated into existing LightGaussian and MesonGS pruning framework.

Critical weaknesses
  • The connection between the proposed score and actual removal distortion is not sufficiently validated. Eq. 14 applies a local diagonal approximation to a finite, simultaneous change in all three scales, Eq.17 then adds estimates from two alternative disappearance paths. Neither operation establishes that the sum better approximates the distortion of removing one primitive. Additionally, retaining the within-primitive scale cross terms need not require constructing the full matrix. It can be achieved by using the same random-projection gradients, averaging the squared sum of the three log-scale gradients estimates the corresponding local directional energy.

  • The evidence for the specifically dual-path contribution is limited. Tab. 5 evaluates one dataset at one pruning ratio: scale-only obtains 21.235 dB and the combined score 21.329 dB (+ 0.094dB).

  • Some experiment settings need clarification. How many views and independent projections are used for importance estimation?

Questions for authors
  1. On a representative subset of primitives, how do opacity-only, scale-only, and combined scores correlate with the measured rendering distortion after actual deletion?

  2. Can you compare Eq.14 with the estimator obtained by squaring the sum of the three log-scale random-projection gradients? It uses the same backward passes while retaining the local scale cross terms.

  3. How many views and independent projections are used for importance estimation, and how stable are rankings and final quality across these choices?

6 comments
AuthorsOpenReview ↗

Response for Comments

We sincerely thank the reviewer for the insightful and constructive feedback and the specific technical suggestions. We address each concern point by point below, with further explanations and additional experiments.

Response for W1, Q1 and Q2:

W1.1 Connection between estimated importance and actual removal distortion (Q1)

Directly measuring the distortion caused by removing every Gaussian requires extensive repeated rendering. To make this validation feasible, we conducted a controlled experiment across all 13 scenes. For each scene, we selected 100 visible Gaussians from the first training view, removed them individually from the original model, and measured the actual rendering distortion (MSE). We then computed the Spearman rank correlation between the measured distortion and the estimated importance scores.

DatasetSceneOpacity-OnlyScale-OnlyDual Path
MipNeRF360bicycle0.9110.9110.958
MipNeRF360flowers0.9000.8860.965
MipNeRF360garden0.8180.8680.917
MipNeRF360stump0.9170.9100.949
MipNeRF360treehill0.9190.8490.955
MipNeRF360room0.9290.9310.951
MipNeRF360counter0.9390.9350.954
MipNeRF360kitchen0.9130.9420.961
MipNeRF360bonsai0.8620.8700.938
DeepBlendingdrjohnson0.8780.9130.946
DeepBlendingplayroom0.9060.9220.946
TanksAndTemplestruck0.8720.8530.944
TanksAndTemplestrain0.9330.8980.934
Average0.9000.8990.948

The Dual-path Score achieves the highest correlation in all 13 scenes, with an average Spearman correlation of 0.948, compared with 0.900 for Opacity-only and 0.899 for Scale-only. These results show that the dual-path score achieves a higher rank correlation with measured removal distortion than either single-path score, supporting its effectiveness for importance ranking under practical approximations.

Furthermore, our original experiments on zero-shot pruning, post-pruning fine-tuning, and downstream compression pipelines consistently demonstrate the effectiveness of VanishGS in practical applications, further supporting the reliability of our importance estimation.

(2) Justification for the additive dual-path formulation

As described in Sec. 3.2, we therefore model Gaussian disappearance through two complementary paths: Opacity Dissolution captures the impact of removing radiometric contribution, while Geometric Collapse captures the rendering sensitivity of spatial support.

Regarding the additive formulation in Eq. (17), both paths are evaluated using the same Gauss–Newton rendering-distortion measure. The combined score can be interpreted as the trace of an appropriately rescaled Gauss–Newton submatrix, aggregating the sensitivities of both disappearance directions. Although this formulation is not theoretically guaranteed to be optimal, it provides a simple and efficient way to combine complementary information without additional weight optimization.

The ablation results in Table 5 support this design. Compared with Scale-only, the full method improves PSNR by 0.094 dB, SSIM by 0.007, and LPIPS by 0.004. Compared with Opacity-only, it improves PSNR by 0.991 dB, SSIM by 0.011, and LPIPS by 0.008. The correlation results presented above further validate the effectiveness of the dual-path design.

(3) Local approximation and cross-scale interactions (Q2)

As described in Sec. 3.1, we use scale perturbations associated with geometric collapse and a local Gauss–Newton approximation to efficiently estimate rendering distortion for importance scoring. Although this approximation does not guarantee exact distortion estimation for complete geometric collapse, it provides a computationally efficient importance measure whose effectiveness is supported by our experiments. Regarding cross-scale interactions, Eqs. (8)–(9) adopt a diagonal Gauss–Newton approximation for efficient importance estimation, and Eq. (14) follows the same approximation for the Scale Path.

To examine the effect of cross-scale interactions, we compare our trace-based diagonal score with a joint directional-sensitivity variant that retains the cross-scale terms.

MethodDataset_GroupratioPSNRSSIMLPIPS
VanishGS_jointMipNeRF3600.921.0110.6490.367
VanishGS_diagMipNeRF3600.921.3290.6420.371
VanishGS_jointTanksAndTemples0.919.2200.6900.329
VanishGS_diagTanksAndTemples0.919.6730.6880.334
VanishGS_jointDeepBlending0.925.9370.8380.329
VanishGS_diagDeepBlending0.925.8800.8380.329

As shown in the results, both variants achieve comparable overall rendering quality at 90% pruning. The diagonal estimator achieves higher PSNR on Mip-NeRF 360 (+0.318 dB) and Tanks & Temples (+0.453 dB), while the joint estimator achieves a slight PSNR improvement on Deep Blending (+0.057 dB) and modest SSIM/LPIPS gains on some datasets. These results support the practical reasonableness of our diagonal approximation.

Overall, the additional experiments support the agreement between our importance rankings and measured removal distortion, the effectiveness of dual-path fusion, and the practical utility of the diagonal approximation.

Response for W2

Although the full dual-path method improves PSNR by only 0.094 dB over Scale-only, it is worth noting that SSIM increases by 0.007 and LPIPS decreases by 0.004, showing notable improvements in structural similarity and perceptual quality.

These results suggest that the benefit of the dual-path design goes beyond reducing pixel-wise reconstruction error, helping better preserve image structure and perceptual quality. Different disappearance directions provide complementary local sensitivities under practical approximations, leading to more reliable importance rankings.

We have further emphasized the contributions of the dual-path design to structural fidelity and perceptual quality in the revised manuscript.

Response for W3 and Q3:

As described in Eqs. (6)–(9), we use all available training views for importance estimation, with one independent Rademacher projection per view. Thus, for a scene with (V) training views, we perform (V) independent projections in total. All view-based methods use the same training views for fair comparison.

Following the reviewer's suggestion, we further evaluated the sensitivity to the number of training views and random projections at 90% pruning.

(1) Sensitivity to the number of training views.

We evaluated VanishGS using 40%, 60%, 80%, and 100% of the training views, while keeping one projection per view.

MethodDataset_GroupTrain_View_FractionPSNRSSIMLPIPS
VanishGSMipNeRF3600.421.0600.6360.374
VanishGSMipNeRF3600.621.1900.6390.372
VanishGSMipNeRF3600.821.2600.6400.372
VanishGSMipNeRF3601.0 (Full)21.3290.6420.371
VanishGSDeepBlending0.425.2600.8300.337
VanishGSDeepBlending0.625.6000.8350.332
VanishGSDeepBlending0.825.7900.8380.330
VanishGSDeepBlending1.0 (Full)25.8800.8380.329
VanishGSTanksAndTemples0.419.4500.6830.339
VanishGSTanksAndTemples0.619.5500.6850.338
VanishGSTanksAndTemples0.819.6600.6880.336
VanishGSTanksAndTemples1.0 (Full)19.6730.6880.334

The results show that reducing the number of training views gradually decreases rendering quality, while VanishGS still maintains competitive performance. Notably, using only 80% of the training views results in PSNR drops of just 0.069, 0.090, and 0.013 dB across the three datasets, respectively, indicating good robustness to view sampling.

(2) Sensitivity to the number of random projections.

We further varied the number of independent projections per training view from 1 to 3 while using all training views.

Dataset_GrouprateNum_ProbesPSNRSSIMLPIPS
MipNeRF3600.9121.3290.6420.371
MipNeRF3600.9221.3430.6420.371
MipNeRF3600.9321.3390.6420.371
DeepBlending0.9125.8800.8380.329
DeepBlending0.9225.9370.8400.328
DeepBlending0.9325.9640.8400.328
TanksAndTemples0.9119.6730.6880.334
TanksAndTemples0.9219.6740.6890.334
TanksAndTemples0.9319.6820.6880.334

The results indicate that increasing the number of random projections provides only marginal improvements in rendering quality. In particular, using three projections instead of one improves PSNR by only 0.010, 0.084, and 0.009 dB across the three datasets, respectively.

Therefore, we adopt one projection per training view as an effective trade-off between estimation quality and computational efficiency, avoiding unnecessary additional backward passes.

We have included these experiments and clarified the view sampling and projection settings in the revised manuscript.

Dear Reviewer H7wz,

Thank you again for your valuable feedback. We have carefully addressed your concerns in our rebuttal and updated the manuscript accordingly, including additional experiments and clarifications.

We would greatly appreciate it if you could review our responses and the revised manuscript when convenient.

We would also be happy to provide further clarification if needed.

Thank you for your time and consideration.

Reviewer_w9K7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the issue of having too many Gaussians in 3DGS models. The authors propose a dual-path primitive importance approach to rank the Gaussians and remove the less important ones. They first propose a mathematical approximation to estimate how much the rendering result changes when a Gaussian is removed. They then score the Gaussians using opacity and scale and combine the two scores to obtain the final ranking. The paper compares the proposed method with existing approaches, including LightGaussian, MesonGS, PUP, RAP, REFINE, EAGLES, and an opacity-only approach, on Mip-NeRF 360, Deep Blending, and Tanks & Temples. The results show that the proposed method can improve PSNR at high pruning ratios, although it requires more pruning time than several competing approaches.

Critical strengths
  1. The proposed approach clearly outperformed other methods at higher pruning ratios.

  2. With its reported computational cost, the proposed approach could be easily adopted in existing 3DGS pipelines when pruning is needed.

Critical weaknesses
  1. In Table 4, it seems that the PUP results were taken from the original paper and obtained using the pretrained 3DGS model from that work. Since the authors may be using a different pretrained model, would these results change if they reran PUP on their own pretrained model?

  2. The ablation study shows that using scale alone already achieves a PSNR of 21.235, while adding opacity improves it by only 0.094. This makes me wonder whether the dual-path design is the main driver of the proposed approach. Although the authors mention this result, they do not provide much explanation for why this is the case.

  3. The authors approximate how opacity and scale affect rendering and use this approximation to score each Gaussian. I wonder whether it would be possible to conduct a small-scale experiment using a subset of Gaussians selected from the full 3DGS model to limit the experiment's size. The authors could remove each selected Gaussian individually, compute the resulting rendering distortion, and compare the results with the proposed scoring method.

  4. RAP already uses opacity and scale for pruning, so the proposed idea seems incremental relative to this previous work. I understand that there are differences, so please clarify them in the rebuttal and the manuscript.

  5. Equation 9 averages sensitivity estimates over all training views, including views in which a Gaussian is invisible. Could this underweight Gaussians that preserve important details visible from only a few viewpoints, particularly when camera coverage is uneven?

  6. The following closely related works are not cited or discussed in this paper:

[1] Ning, Zhenhua, et al. "Beyond Heuristics: Learnable Density Control for 3D Gaussian Splatting." ICML 2026.

[2] Du, Xiaobiao, et al. "Mobile-GS: Real-time Gaussian splatting for mobile devices." ICLR 2026.

[3] Xiong, Butian, et al. "Nanogs: Training-free gaussian splat simplification." ECCV 2026.

[4] Lee, Yongjae, Zhaoliang Zhang, and Deliang Fan. "Safeguardgs: 3d gaussian primitive pruning while avoiding catastrophic scene destruction." WACV 2026.

Questions for authors
  1. I believe all of the weaknesses can be addressed in the rebuttal. Please clarify whether this work used the same pretrained 3DGS model as PUP. If not, please provide a matched comparison.

  2. Please provide further analysis to address Weaknesses 2 and 3, and clarify the view-averaging trade-off raised in Weakness 5.

  3. Please clarify the novel contributions relative to RAP (Weakness 4) and discuss and cite the related work mentioned in Weakness 6.

5 comments
AuthorsOpenReview ↗

Response for Comments

We sincerely thank the reviewer for the insightful and constructive feedback. We have carefully considered each concern and provide point-by-point responses below, including additional experimental results and clarifications.

Response for W1

Thank you for pointing this out. To ensure a fair comparison, we reran PUP using the official pretrained 3DGS models used for VanishGS. Notably, all experiments in our paper use the official pretrained models provided by 3DGS, rather than independently trained models, to ensure fairness and reproducibility.

The rerun results show slight differences from the originally reported values, but the overall conclusion remains unchanged. The updated results are shown below. Compared with the original comparison, VanishGS achieves larger PSNR gains over the rerun PUP on Mip-NeRF 360 (0.28 vs. 0.18 dB) and Tanks & Temples (0.83 vs. 0.58 dB). Although the gain decreases on Deep Blending (0.17 vs. 0.56 dB), VanishGS still achieves higher PSNR and SSIM across all three datasets.

We have updated Table 4 in the revised manuscript.

DatasetMethodPSNRSSIMLPIPS
Mipnerf360PUP(Org)26.670.7860.272
Mipnerf360PUP(Reran)26.570.7860.268
Mipnerf360VanishGS26.850.7880.271
Tanks & TemplePUP(Org)22.720.8010.244
Tanks & TemplePUP(Reran)22.470.7990.248
Tanks & TempleVanishGS23.300.8100.242
Deep BlendingPUP(Org)28.850.8810.301
Deep BlendingPUP(Reran)29.240.8970.274
Deep BlendingVanishGS29.410.8990.267

Response for W2

Although the PSNR gain over Scale-only is modest (+0.094 dB), the additional Opacity Path provides complementary sensitivity information, reducing reliance on a single approximate importance estimate. This also leads to improvements in SSIM (+0.007) and LPIPS (−0.004), indicating better preservation of structural and perceptual quality.

These results suggest that the benefit of the dual-path design goes beyond reducing pixel-wise reconstruction error, helping better preserve image structure and perceptual quality.

We have further emphasized the contributions of the dual-path design to structural fidelity and perceptual quality in the revised manuscript.

Response for W3

Directly measuring the distortion caused by removing every Gaussian requires extensive repeated rendering. To make this validation feasible, we conducted a controlled experiment across all 13 scenes. For each scene, we selected 100 visible Gaussians from the first training view, removed them individually, and measured the actual rendering distortion (MSE). We then computed the Spearman rank correlation between the measured distortion and the estimated importance scores.

The Dual-path Score achieves the highest correlation across all 13 scenes, with an average of 0.948, compared with 0.900 for Opacity-only and 0.899 for Scale-only. These results demonstrate that combining both paths provides more reliable importance estimates.

DatasetSceneOpacity-OnlyScale-OnlyDual Path
MipNeRF360bicycle0.9110.9110.958
MipNeRF360flowers0.9000.8860.965
MipNeRF360garden0.8180.8680.917
MipNeRF360stump0.9170.9100.949
MipNeRF360treehill0.9190.8490.955
MipNeRF360room0.9290.9310.951
MipNeRF360counter0.9390.9350.954
MipNeRF360kitchen0.9130.9420.961
MipNeRF360bonsai0.8620.8700.938
DeepBlendingdrjohnson0.8780.9130.946
DeepBlendingplayroom0.9060.9220.946
TanksAndTemplestruck0.8720.8530.944
TanksAndTemplestrain0.9330.8980.934
Average0.9000.8990.948

Furthermore, our original experiments on zero-shot pruning, post-pruning fine-tuning, and downstream compression pipelines consistently demonstrate the effectiveness of VanishGS in practical applications, further supporting the reliability of our importance estimation.

Response for W4

Thank you for the comment. We emphasize that opacity and scale are fundamental attributes of 3DGS, and using the same attributes does not imply the same importance modeling approach.

Specifically, RAP uses 15 hand-crafted features based on Gaussian attributes and local statistics to predict importance through an MLP. Opacity and scale serve as input features, without explicitly establishing an analytical connection to the rendering changes caused by Gaussian removal.

In contrast, VanishGS starts from the physical effect of pruning and explicitly links Gaussian disappearance to parameter perturbations. As described in Secs. 3.1–3.2, we model disappearance through Opacity Dissolution and Geometric Collapse, and use a Gauss–Newton approximation to estimate the resulting rendering distortion. This provides a more direct and interpretable basis for importance estimation.

Response for W5

We clarify that the averaging across views in Eq. (9) is consistent with the multi-view rendering distortion objective defined in Eq. (2). Following the original 3DGS training protocol, which uniformly samples training views without assigning view-specific weights, all training views are equally weighted. Invisible Gaussians contribute zero gradients, so including these views does not bias the estimator with respect to the uniformly averaged multi-view objective in Eq. (2).

Moreover, our pruning objective is to remove redundant Gaussians with minimal impact on overall rendering quality. Although Gaussians visible in only a few views may contribute less to the overall distortion, those preserving important details can still receive high importance scores due to their strong local rendering sensitivity. We acknowledge that uneven camera coverage may affect the relative importance of different regions, reflecting a trade-off between the view distribution and the overall distortion objective.

Response for W6

Thank you for suggesting these relevant works. We have added the corresponding citations and discussions in the revised manuscript.

Although these methods all aim to improve 3DGS efficiency, they differ in focus and methodology. Specifically, LeGS uses reinforcement learning for density control during training; Mobile-GS focuses on real-time mobile rendering and compression; NanoGS targets rendering quality under extreme compression through local Gaussian merging; and SafeguardGS focuses on pixel-level importance ranking and protection mechanisms under aggressive pruning.

In contrast, VanishGS focuses on importance estimation for post-training pruning. Starting from the rendering impact of Gaussian disappearance, we explicitly model pruning-aligned parameter perturbations through Opacity Dissolution and Geometric Collapse, and use Gauss–Newton information to estimate the corresponding rendering sensitivity.

We have further clarified the connections and differences between these works and VanishGS in the revised manuscript.

Dear Reviewer w9K7,

Thank you again for your valuable feedback. We have carefully addressed your concerns in our rebuttal and updated the manuscript accordingly, including additional experiments and clarifications.

We would greatly appreciate it if you could review our responses and the revised manuscript when convenient.

We would also be happy to provide further clarification if needed.

Thank you for your time and consideration.

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

The proposed VanishGS employs a post-training pruning strategy to achieve lightweight 3D Gaussian Splatting (3DGS). Motivated by the observation that rendering contribution and local parameter sensitivity may not directly characterize the effect of primitive removal, the authors formulate Gaussian importance in terms of rendering distortion induced by primitive disappearance. Specifically, the authors propose two disappearance paths: opacity dissolution, which removes a Gaussian by setting its opacity to zero, and geometric collapse, which removes its spatial support by shrinking its scales to zero. The corresponding distortion is approximated using diagonal Gauss–Newton information, estimated via randomized Rademacher projections. The two terms are combined into a unified importance score used to rank and prune Gaussians. Experiments on mutlple datasets demonstrate improvements over other pruning baselines. T

Critical strengths

1. Well-motivated problem formulation. The distinction between rendering contribution and removal-induced distortion is meaningful. Designing pruning criteria around the consequences of removing a primitive is an intuitive and potentially useful direction.


2. Strong empirical performance. The zero-shot results show substantial PSNR improvements, especially at 90% pruning. For example, on Mip360, VanishGS achieves 21.32, compared with 19.9 for EAGLES. The evaluation includes multiple pruning ratios, three datasets, and several competitive baselines.

Critical weaknesses

1. Insufficient theoretical justification for the dual-path formulation.

Under standard alpha compositing, the rendered color is affine in an individual Gaussian's opacity when other parameters are fixed. Thus, setting its opacity to zero gives

=-\alpha_i\frac{\partial C}{\partial\alpha_i}.$$ Consequently, the opacity-based distortion already characterizes the actual removal distortion under this model: $$D_i=\alpha_i^2 H_{\alpha_i}.$$ It is therefore unclear why geometric collapse is additionally needed to model removal-induced distortion, or why summing the distortions from two alternative disappearance paths is theoretically justified. **2. Potential inaccuracy of the geometric-collapse approximation.** The geometric path applies $\Delta s_i=-s_i$, which represents a large perturbation, whereas the Gauss–Newton approximation relies on local linearization. Moreover, since $s_i=\exp(s_i')$, collapsing the scale to zero corresponds to $s_i'\to-\infty$, rather than a finite perturbation in log-scale space. The proposed formulation also ignores cross-parameter interactions: $$D_s\approx\Delta s^\top H_{ss}\Delta s \neq\sum_k H_{s_k s_k}(\Delta s_k)^2.$$ The accuracy of this diagonal approximation for complete geometric collapse remains insufficiently validated. **3. Limited novelty relative to existing second-order pruning methods.** The proposed approach is closely related to PUP and REFINE, which already leverage second-order rendering sensitivity or removal-aware importance estimation. The main novelty appears to lie in combining opacity and scale sensitivities, while the Gauss–Newton approximation and randomized diagonal estimation are established techniques. **4. Lack of direct validation of importance-estimation accuracy.** Although the experimental results demonstrate effective pruning, they do not establish that the proposed scores accurately predict actual removal-induced distortion. The authors should directly measure the rendering distortion caused by removing individual Gaussians and evaluate the correlation between predicted importance and measured distortion, e.g., using Spearman rank correlation. **5. Limited ablation and efficiency analysis.** The scale-only variant already achieves 21.2dB on Mip-NeRF 360 at 90% pruning, compared with 21.3dB for the full method. The additional opacity path thus provides only a 0.09dB improvement. Moreover, the two terms are combined with equal weights without sufficient justification.
Questions for authors

please refer to the weaknesses

6 comments
AuthorsOpenReview ↗

Response for Comments

We sincerely thank the reviewer for the insightful and constructive feedback. We have carefully considered each concern and provide point-by-point responses below, supported by additional analyses and experiments.

Response for W1:

Our main motivation for introducing two paths is to obtain more robust importance estimates despite the approximations required in practice. Theoretically, both Opacity Dissolution and Geometric Collapse aim to eliminate a Gaussian's rendering contribution. If the final removal distortion could be computed exactly, both paths would yield the same distortion and importance ranking. As shown in the table below, varying the path weights while keeping both non-zero leads to only minor differences in pruning performance.

Dataset_GroupPrune_RateOpacity_WeightScale_WeightPSNRSSIMLPIPS
MipNeRF3600.90.50.521.3290.6420.371
MipNeRF3600.90.20.821.2830.6370.373
MipNeRF3600.90.80.221.2320.6440.370
MipNeRF3600.90.30.721.3030.6390.372
MipNeRF3600.90.70.321.2990.6440.370
MipNeRF3600.90.40.621.3240.6400.372
MipNeRF3600.90.60.421.3250.6430.370
MipNeRF3600.90.01.021.2350.6350.375
MipNeRF3600.91.00.020.3380.6310.379

However, efficient computation requires approximations, such as local Gauss–Newton and diagonal approximations, which may introduce errors in single-path importance estimates. We therefore evaluate Gaussian disappearance along two different parameter directions to reduce reliance on a single estimate. For example, Gaussians with similar opacity-based distortion estimates may exhibit different scale sensitivities. The two paths thus provide complementary information for importance ranking. The results above further support this design: dual-path combinations generally improve PSNR and consistently achieve better SSIM and LPIPS than either single path, demonstrating the benefit of combining both paths.

Regarding the additive formulation in Eq. (17), both paths are evaluated using the same Gauss–Newton rendering-distortion measure. Specifically, the combined score can be interpreted as the trace of an appropriately rescaled Gauss–Newton submatrix, aggregating the rendering sensitivities of both disappearance directions into a unified importance score. Although equal weighting is not theoretically guaranteed to be optimal, this trace-based formulation provides a simple and efficient way to combine complementary information without additional weight optimization. The results above further show that pruning performance remains stable across different non-zero weights, supporting the robustness of the dual-path combination. Therefore, we adopt equal-weight addition as a simple, effective, and robust choice without dataset-specific tuning. Exploring more advanced fusion strategies to further improve importance estimation accuracy is a promising direction for future work.

We have clarified the motivation for dual-path modeling and the interpretation of Eq. (17) in the revised manuscript.

Response for W2:

The scale perturbation serves as a local sensitivity proxy for Gaussian disappearance, rather than an exact finite step to zero in log-scale space. We acknowledge the potential inaccuracy of the geometric-collapse approximation. However, our goal is to strike a practical balance between computational efficiency and estimation reliability, rather than precisely estimate the distortion of complete collapse. The ablation results in Table 5 support this design. At 90% pruning on Mip-NeRF 360, the Scale-only variant achieves 21.235 dB PSNR, outperforming the strongest competing baseline (EAGLES, 19.930 dB in Table A5) by 1.305 dB. This demonstrates the practical effectiveness of our approximation for importance estimation. Exploring more accurate approximations to further improve pruning performance is a promising direction for future work.

Regarding cross-parameter interactions, as stated in Eqs. (8)–(9), we adopt a diagonal Gauss–Newton approximation to efficiently estimate parameter sensitivity for large-scale Gaussian models. Eq. (14) follows the same approximation for the Scale Path. To further validate this simplification, we additionally compare the original diagonal estimator with a variant (VanishGS_joint) that retains cross-scale terms.

MethodDataset_GroupratioPSNRSSIMLPIPS
VanishGS_jointMipNeRF3600.921.0110.6490.367
VanishGS_diagMipNeRF3600.921.3290.6420.371
VanishGS_jointTanksAndTemples0.919.2200.6900.329
VanishGS_diagTanksAndTemples0.919.6730.6880.334
VanishGS_jointDeepBlending0.925.9370.8380.329
VanishGS_diagDeepBlending0.925.8800.8380.329

As shown in the results, both variants achieve comparable overall rendering quality at 90% pruning. The diagonal estimator achieves higher PSNR on Mip-NeRF 360 (+0.318 dB) and Tanks & Temples (+0.453 dB), while the joint estimator achieves a slight PSNR improvement on Deep Blending (+0.057 dB) and modest SSIM/LPIPS gains on some datasets. These results support the practical reasonableness of our diagonal approximation.

We have included these additional results and clarified the approximation in the revised manuscript.

Response for W3:

As discussed in Sec. 2, a key limitation of existing importance estimation methods is the mismatch between the evaluated parameter perturbations and the actual effect of pruning, i.e., Gaussian disappearance. Such a mismatch may lead to inaccurate importance estimation.

Specifically, PUP evaluates general sensitivity using the log-determinant of the Hessian over position and scale parameters, without explicitly modeling perturbations associated with Gaussian disappearance. REFINE, although removal-aware, estimates removal-induced distortion using geometry, color, and opacity attributes with analytical approximations. However, these attribute-based perturbations are not explicitly aligned with the physical process of Gaussian disappearance, potentially introducing estimation errors.

In contrast, VanishGS starts from the physical effect of pruning and explicitly models Gaussian disappearance through Opacity Dissolution and Geometric Collapse. By constructing corresponding parameter perturbations and estimating their rendering impact using Gauss–Newton information, we establish a more direct and interpretable connection between importance estimation and the pruning objective.

Experiments further support this design. At 90% pruning on Mip-NeRF 360, VanishGS outperforms PUP and REFINE by 4.275 dB and 2.598 dB in PSNR, respectively.

Response for W4:

Directly measuring the distortion caused by removing every Gaussian requires extensive repeated rendering. To make this validation feasible, we conducted a controlled experiment across all 13 scenes. For each scene, we selected 100 visible Gaussians from the first training view, removed them individually, and measured the actual rendering distortion (MSE). We then computed the Spearman rank correlation between the measured distortion and the estimated importance scores.

The Dual-path Score achieves the highest correlation across all 13 scenes, with an average of 0.948, compared with 0.900 for Opacity-only and 0.899 for Scale-only. These results demonstrate that combining both paths provides more reliable importance estimates.

DatasetSceneOpacity-OnlyScale-OnlyDual Path
MipNeRF360bicycle0.9110.9110.958
MipNeRF360flowers0.9000.8860.965
MipNeRF360garden0.8180.8680.917
MipNeRF360stump0.9170.9100.949
MipNeRF360treehill0.9190.8490.955
MipNeRF360room0.9290.9310.951
MipNeRF360counter0.9390.9350.954
MipNeRF360kitchen0.9130.9420.961
MipNeRF360bonsai0.8620.8700.938
DeepBlendingdrjohnson0.8780.9130.946
DeepBlendingplayroom0.9060.9220.946
TanksAndTemplestruck0.8720.8530.944
TanksAndTemplestrain0.9330.8980.934
Average0.9000.8990.948

Furthermore, our original experiments on zero-shot pruning, post-pruning fine-tuning, and downstream compression pipelines consistently demonstrate the effectiveness of VanishGS in practical applications, further supporting the reliability of our importance estimation.

We have included these additional results and analyses in the revised manuscript.

Response for W5:

Although the PSNR gain over Scale-only is modest (+0.094 dB), the full method achieves notable improvements in SSIM (+0.007) and LPIPS (−0.004), indicating better preservation of structural and perceptual quality.

Regarding equal weighting, as discussed in our response to W1, Eq. (17) can be interpreted as the trace of an appropriately rescaled Gauss–Newton submatrix, aggregating the sensitivities of both disappearance paths. Although not guaranteed to be optimal, this simple formulation avoids additional weight tuning and achieves consistent improvements over either single path.

We have further clarified this design choice and the complementary contributions of the two paths in the revised manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Responses and Revisions

We sincerely thank all reviewers for their insightful and constructive feedback. We have carefully addressed the concerns and revised the manuscript accordingly, with all revisions highlighted in magenta for easy identification.

In response to the key issues raised by the reviewers, we have made the following improvements:

  1. Clarification of the dual-path formulation. We further explained the motivation for combining Opacity Dissolution and Geometric Collapse, the interpretation of the additive importance score, and the practical approximations involved. We also clarified the key differences between VanishGS and existing importance estimation methods, including PUP, REFINE, and RAP.

  2. Additional experimental validation. We conducted new experiments to evaluate the correlation between estimated importance and actual Gaussian removal distortion, achieving an average Spearman correlation of 0.948 across 13 scenes. We also investigated different dual-path weights, cross-scale interactions, and the effects of training view sampling and random projection counts.

  3. Fair comparison and improved experimental analysis. We reran PUP using the same official pretrained 3DGS models and updated Table 4 accordingly. We also clarified the contributions of both pruning paths and provided additional ablation visualizations.

  4. Expanded related work and implementation details. We added discussions and citations for LeGS, Mobile-GS, NanoGS, and SafeguardGS, and clarified the default training view and random projection settings.

These revisions and additional analyses provide further support for the effectiveness, reliability, and practicality of VanishGS. Detailed responses to each reviewer are provided below.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.