acceptodds
Under review as a conference paper at ICLR 2027

Recoverability-Aware Gaussian Pruning for 3D Gaussian Splatting Compression

12 views

Abstract

Gaussian pruning for 3D Gaussian Splatting (3DGS) typically retains primitives according to importance or estimated removal effect at the model state. Yet pruning quality is evaluated only after the remaining Gaussians are further optimized, leaving a key factor unmodeled: how well the retained representation adapts to removed information. We call this post-pruning recoverability. We study recoverability through controlled counterfactual interventions. Across matched recovery trajectories, recoverability exhibits reproducible structure across recovery schedules, with more consistent cross-schedule agreement at longer recovery horizons, while freezing retained Gaussians with strong geometric responses reduces subsequent recovery. Directly using this response as a pruning criterion, however, degrades reconstruction quality, showing that counterfactual compensation evidence complements rather than replaces primitive importance. Based on this observation, we introduce RAGP (Recoverability-Aware Gaussian Pruning), a fixed-budget refinement framework for Gaussian importance criteria. RAGP uses counterfactual pruning previews to measure the response of the surviving representation to localized information removal. The resulting geometric response is whitened and fused with the host importance prior through bounded host-relative calibration, while preserving the native host-score construction, Gaussian budget, and post-pruning optimization protocol. Thus, RAGP changes only the retained Gaussian identities, adding no parameters or inference-time rendering operations. Across five OMG pruning budgets and 13 scenes, RAGP improves mean PSNR by – dB under matched Gaussian counts, consistently outperforming the pruning pipeline at the same compression budget.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting Compression

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_tRWxOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper argues that the contribution of a Gaussian to the current scene's state is different from the recoverability if removing this Gaussian, and proposes Recoverability-Aware Gaussian Pruning (RAGP), which designed a workflow to perform counterfactual probes during training, which can get Gaussians' response score. Together with a selected existing importance score, this responese score help merge into a final score that can be used to select Gaussian under a given budget.

Experiments show the effectiveness of this proposed score computing workflow by improving PSNR under the budget, while introducing approximately 36% training overhead.

Critical strengths
  1. the motivation is reasonable and clear, and the paper proposes to probe the scene by masking Gaussians to test their recoverability, and utilized it to compute a score that can effectively prune Gaussians while improving rendering quality.

  2. this method is universal and can be applied on top of various existing pruning methods.

Critical weaknesses
  1. The readability is poor, the symbol R^{geo}_i appears in Section 3 without being defined, causing great confusion when reading this article. And the "Counterfactual Evidence" in Section 3 seems like a lot experiments with some findings, further confusing me, as the main method is only introduced later in Section 4; In Section 4, the F_mid function is not defined,

  2. How do you select which part of Gaussians to be masked? This is one of the most important question but seems not be mentioned or explained. Without knowing this, I really could not fully understand the working mechanism.

  3. There are enumerous formulas to derive the final score, and one has to sort them out with the help of AI, as many of them are in the Appendix and so complex designed. This makes it difficult to figure out the validity of each component, even though some designs are ablated.

Questions for authors
  1. Could you provide a clearer, self-contained explanation of the full mapping from gradient contrasts to the final response score?
  2. Could you elaborate how to select the cells/ Gaussians to be masked?
4 comments
AuthorsOpenReview ↗

Response part 1

We thank the reviewer for the constructive feedback. We have revised the manuscript to make the response-score computation and masking procedure clearer and more self-contained.

[Q1] Full mapping from gradient contrasts to the final response score

Response. We have clarified the distinction between the motivation in Section 3 and the actual score construction in Section 4.2. Specifically, RigeoR_i^{\mathrm{geo}} is a pre-pruning geometric compensation response, rather than a direct estimate of the retrospective, multi-step recoverability measure CTC_T.

The complete mapping consists of four steps:

  1. Paired gradient contrasts: Normal and Masked branches start from the same Gaussian state and camera. We compute Scale and Rotation gradient differences, while only the Normal branch updates model parameters.

  2. Covariance-tangent mapping: The contrasts are normalized using Normal-branch gradient magnitudes, with Rotation projected onto the unit-quaternion tangent space. Both are mapped into a six-dimensional covariance-tangent representation.

  3. Response construction: The geometric response is whitened using Normal-branch statistics, normalized by removed contribution mass, and aggregated across valid interventions using variance- and observation-count-aware reliability penalties to obtain LiL_i.

  4. Final score construction: We apply a tie-aware empirical midpoint CDF over Gaussians with at least four valid observations. Each score is assigned the fraction of eligible scores below it plus half the fraction tied with it. The resulting response is Rigeo=2F^mid(Li)−1R_i^{\mathrm{geo}}=2\widehat F_{\mathrm{mid}}(L_i)-1. Gaussians with fewer than four valid observations receive zero response.

Revision. We explicitly define RigeoR_i^{\mathrm{geo}} at its first occurrence in Section 3 and clarify the transition from counterfactual evidence to the actual algorithm. Section 4.2 now presents the complete gradient-to-response mapping and explains the midpoint CDF, making the score construction self-contained. Supporting implementation details are provided in Appendix A.2.

[Q2] Selection of cells and Gaussians for masking

Response. RAGP masks spatially grouped Gaussians rather than randomly selected individual primitives. At the frozen 15K state, Gaussian identities are fixed and their normalized centers are partitioned into occupied Fine cells. Starting from the resolution whose occupied-cell count is closest to 128, we examine progressively finer grids under the frozen training-camera schedule.

A cell is eligible if it contains at least one Gaussian with positive screen-space radius and positive accumulated rendering contribution in at least one scheduled training view. We select the finest feasible resolution for which every eligible cell receives at least two actual Masked observations during 15.5K–20K. The scheduler prioritizes under-observed cells approaching their last visible opportunities.

For each selected cell-camera pair, all Gaussians within the cell are temporarily attenuated using a shared contribution-aware multiplier. The perturbed Gaussians are excluded from response aggregation; only Gaussians outside the selected cell serve as compensation responders. Normal and Masked branches share the same initial state and camera, while Masked computes read-only gradients and only Normal updates the model.

Revision. We clarified cell eligibility, adaptive resolution selection, temporary attenuation, and responder selection in Section 4.2, with supporting implementation details in Appendix A.1.

AuthorsOpenReview ↗

Response part 2

[W3] Component effectiveness and design complexity

Response. The score construction is a sequence of transformations applied to one counterfactual response, rather than a collection of independent scoring modules. Each stage has a specific role: paired gradient contrasts capture intervention-induced changes; covariance-tangent mapping represents geometric variation; whitening and reliability-aware aggregation normalize and stabilize the response; and midpoint-CDF normalization produces scene-relative compensation evidence.

Core component ablation (Appendix G.1, Table 19). We evaluate the principal components and design choices at budget X. All configurations share the same 20K source state, native host prior, exact-K budget, recovery schedule, and evaluation protocol. The table reports 13-scene mean changes relative to the paired Official OMG branch.

ConfigurationΔPSNR ↑ΔSSIM ↑ΔLPIPS ↓
A2: Whitened Fine only+0.10592+0.000656-0.000444
A4: HRNB only (w/o Fine)+0.01233+0.000499-0.000802
Gradient norm + Host-rank+0.09530+0.000133+0.000313
Covariance tangent + Host-rank+0.10748+0.000564-0.000249
A1: Whitened + Uniform+0.09715-0.000931+0.002017
Whitened Fine w/o reliability+0.09223+0.000495-0.000316
A3: Whitened Fine + Odds-5/View+0.10478+0.001226-0.001612
Full RAGP+0.12083+0.001289-0.001609

Whitened Fine provides the primary improvement (+0.10592 dB), while HRNB alone contributes +0.01233 dB. Combining them increases the gain to +0.12083 dB, an additional +0.01491 dB over Fine alone. Full RAGP achieves the highest PSNR and SSIM in this comparison, with LPIPS comparable to the best alternative.

Geometric response representation (Appendix G.4, Table 22). This independent matched diagnostic compares alternative response representations.

Response constructionΔPSNR ↑ΔSSIM ↑ΔLPIPS ↓
Gradient-norm response+0.2153+0.00113-0.00147
Covariance tangent, no whitening+0.2140+0.00148-0.00197
Covariance tangent + whitening+0.2200+0.00154-0.00207

Covariance-tangent whitening achieves the best results across all three metrics in this study, supporting the selected geometry-aware, optimization-normalized response.

Aggregation strategy (Appendix G.6, Table 25). This separate matched diagnostic compares Fine-only with multi-scale aggregation alternatives under the same Fine/Medium/Coarse hierarchy.

Aggregation strategyΔPSNR ↑ΔSSIM ↑ΔLPIPS ↓
Fine only+0.2421+0.00164-0.00212
Arithmetic mean+0.2193+0.00144-0.00198
Reliability-weighted mean+0.2249+0.00151-0.00209
Weighted median+0.2238+0.00140-0.00189

Fine-only outperforms all evaluated multi-scale alternatives across PSNR, SSIM, and LPIPS, supporting the simpler single-scale design. The results in Tables 22 and 25 are interpreted within their respective matched diagnostic studies and are not compared numerically with Table 19.

Revision. Section 4.2 organizes the construction into four sequential stages. Component and design-choice evidence is reported in Appendix G.1 (Table 19), G.4 (Table 22), and G.6 (Table 25).

Thank you for the clarifications and additional ablations. The revised explanations of masking and response-score construction have largely addressed my concerns, and I will raise my rating to Weak Accept. I encourage the authors to add concise end-to-end pseudocode to the main text, making the computation easier to follow without repeatedly consulting the appendix.

Thank you for your positive feedback and for reconsidering our work. We appreciate your helpful suggestion. We will add concise end-to-end pseudocode to the main text in a subsequent revision to improve clarity and make the computation easier to follow. Thank you again for your constructive comments and encouragement.

Reviewer_ZvQCOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Recoverability-Aware Gaussian Pruning (RAGP), a method for choosing which Gaussians to keep when compressing a 3D scene. It temporarily weakens small groups of Gaussians and observes how the gradients of the remaining Gaussians respond. It combines this information with an existing importance score to decide which Gaussians to retain. The key idea is to keep Gaussians that are important to the scene (this is common) and those that could help recover visual quality after pruning (this is novel). The method retains the same number of Gaussians and uses the same refinement process as the baseline. The results improve visual quality, but the method introduces substantial training-time overhead.

Critical strengths
  • The paper proposed a pruning technique from the interesting observation. The motivation is therefore clear and convincing.
  • The results show consistent PSNR improvements with broad testing of importance criteria
Critical weaknesses
  • Equations (1) and (2) do not guarantee that the masked loss remains higher than the normal loss. Although the condition D0>δD_0 > \delta ensures positive initial damage for included interventions, DTD_T may become negative or exceed D0D_0, so CTC_T can lie outside [0,1][0,1]. Please clarify these cases, report how many interventions are excluded by the initial-damage condition, and distinguish relative loss-gap closure from absolute improvement in the masked branch. Actually, how straightforward is it to define masks that satisfy D0>δD_0 > \delta?
  • Section 3 shows that CTC_T is meaningful, but only indirectly argues that RigeoR^{geo}_i predicts or reflects CTC_T. The link between tehm is established mainly through empirical observations rather than mathematical analysis. So, Section 3 is conceptually convincing but not fully rigorous.
  • Section 4.3 explains how HRNB is constructed, but its necessity is less clear. Why should projected-area normalisation improve the integration of counterfactual response with the host score? Table 18 shows that most of the improvement is achieved without HRNB. Please strengthen the motivation and compare against a simpler normalisation or score-combination approach.
  • RAGP uses additional computation for counterfactual probing and calibration, while the baseline follows the same number of training iterations. Please also compare at equal total running time, allowing the baseline to use the additional time for further optimisation. This would clarify whether RAGP’s quality gains justify its reported training-time overhead.
Questions for authors

Please see weaknesses above.

6 comments
AuthorsOpenReview ↗

Response part 1

Thank you for your careful evaluation and recognition of RAGP's motivation and consistent improvements across pruning criteria. We address W1 and W2 together, as both concern the relationship between post-pruning recoverability and geometric compensation response.

[W1] Interpretation and screening of recoverability

Response. We define recoverability as signed relative closure of the initial Masked–Normal loss gap: CT=1−DT/D0C_T=1-D_T/D_0, where D0=LM0−LN0>δD_0=L_M^0-L_N^0>\delta and DT=LMT−LNTD_T=L_M^T-L_N^T. The condition D0>δD_0>\delta ensures a positive initial gap but does not constrain DTD_T. Thus, CTC_T is not restricted to [0,1][0,1]: CT<0C_T<0 indicates a widening gap; CT=0C_T=0 no closure; 0<CT<10<C_T<1 partial closure; CT=1C_T=1 complete closure; and CT>1C_T>1 gap reversal.

To distinguish relative gap closure from absolute improvement, define GM=LM0−LMTG_M=L_M^0-L_M^T and GN=LN0−LNTG_N=L_N^0-L_N^T. They satisfy D0CT=GM−GND_0C_T=G_M-G_N. Thus, CTC_T compares the two branches' loss reductions, not the Masked branch's improvement alone.

Mask screening. Candidate localized interventions are generated from scheduled Fine-cell/camera pairs using our fixed contribution-controlled masking protocol, independently of the recoverability criterion. We then screen them using D0>δD_0>\delta.

CandidatesValidExcluded
10466 (63.5%)38 (36.5%)
Recovery stepsNegativeBetween 0 and 1Above 1GMG_MGNG_N
1004557+0.05947−0.00164
5005529+0.06781−0.00487

The CTC_T counts cover the 66 valid interventions; GMG_M and GNG_N are 13-scene mean loss reductions. The Masked branch decreases in loss on average, while the Normal branch slightly increases. This diagnostic set is distinct from the 23 interventions in the recovery-schedule study.

Revision. We clarified the signed interpretation and its distinction from absolute branch-wise changes in Section 3, and reported signed-gap screening statistics in Appendix F.

[W2] Mathematical connection between geometric response and recoverability

Response. At a shared pre-update state and camera, let Dm=LmM−LmND_m=L_m^M-L_m^N be the signed Masked–Normal gap. Paired Scale/Rotation contrasts differentiate this gap with respect to Gaussian parameters. For a=Δgi,msa=\Delta g_{i,m}^{s}, b=Δgi,mqb=\Delta g_{i,m}^{q}, and Pi=I−qiqiTP_i=I-q_iq_i^T, the negative normalized Scale and tangent-projected Rotation directions yield Dm′(0)=−∥a∥22/νms−bTPib/νmq≤0D'_m(0)=-\|a\|_2^2/\nu_m^s-b^TP_ib/\nu_m^q\leq0. Because PiP_i is positive semidefinite and both normalizers are positive, this direction locally decreases the signed gap underlying CTC_T (Appendix A.2, Eq. 16).

RAGP maps these contrasts into covariance-tangent space, whitens and contribution-normalizes them, aggregates reliable observations across interventions, and applies a scene-relative empirical CDF to obtain RigeoR_i^{geo}.

Both quantify compensation under localized information removal at different stages and granularities: CTC_T is an intervention-level retrospective measure of realized multi-step gap closure, whereas RigeoR_i^{geo} is a pre-pruning, per-Gaussian response from paired gradients. They are complementary, not numerically equivalent. The first-order result links the response direction to instantaneous gap-closure tendency; it does not predict multi-step CTC_T.

Empirical evidence. Recovery-schedule consistency assesses the stability of intervention-level recoverability rankings:

HorizonSchedule pairSpearman ρ\rho (95% CI)
100Seed 0 vs. 10.831 [0.524, 0.991]
100Seed 0 vs. 20.480 [−0.069, 0.948]
100Seed 1 vs. 20.432 [−0.115, 0.933]
500Seed 0 vs. 10.818 [0.491, 0.985]
500Seed 0 vs. 20.755 [0.415, 0.930]
500Seed 1 vs. 20.637 [0.178, 0.925]

The freezing intervention evaluates the functional relevance of RigeoR_i^{geo} by comparing local recovery after freezing high-response Gaussians with random and low-response controls:

HorizonComparisonMean difference (95% CI)Negative scenes
100High − Random−0.031 [−0.042, −0.020]12/13
100High − Low−0.032 [−0.043, −0.020]12/13
500High − Random−0.011 [−0.016, −0.007]12/13
500High − Low−0.011 [−0.017, −0.007]12/13

All four freezing comparisons yield one-sided Wilcoxon p<0.001p<0.001. Negative differences indicate weaker subsequent local recovery when high-response Gaussians are frozen (Appendix F.1–F.2, Tables 15–17). Together, the first-order result and freezing intervention support RigeoR_i^{geo} as pre-pruning evidence of compensatory participation, complementary to instantaneous importance.

Revision. We clarified the distinct roles of CTC_T and RigeoR_i^{geo} in Section 3, added the first-order derivation in Appendix A.2 (Eq. 16), and reported schedule consistency and freezing analyses in Appendix F.1–F.2.

AuthorsOpenReview ↗

Response part 2

[W3] Motivation and independent contribution of HRNB

Response. HRNB calibrates the integration of counterfactual geometric response with heterogeneous host importance criteria. Raw rendering contribution and projected-area-normalized contribution capture different aspects of Gaussian importance, while native host criteria already encode these effects to varying degrees.

Different host criteria exhibit different sensitivity to projected-area normalization, so directly adding an auxiliary response can produce host-dependent ranking shifts. HRNB estimates how strongly the native host ranking already reflects a projected-area-normalized contribution trend and applies a reliability-gated, bounded correction along this normalization direction. The counterfactual response is incorporated afterward, preserving the native host ranking as the prior while adapting the integration of compensation evidence.

Additional experiments. We conducted two complementary comparisons.

1. Comparison with simpler calibration methods.

ConfigurationPSNR ↑SSIM ↑LPIPS ↓
Host+Fine27.07890.83140.2334
Simple area normalization27.06320.83250.2313
Simple residual correction27.07710.83240.2314
Full HRNB27.10240.83210.2321

Full HRNB improves mean PSNR by +0.0235 dB over Host+Fine and achieves higher PSNR than both simpler alternatives. It also achieves higher PSNR than Host+Fine in 11/13 scenes (Appendix G.8, Table 27).

In the final-pipeline component ablation (Table 19), Whitened Fine alone achieves +0.10592 dB over Official OMG, while the full method reaches +0.12083 dB. These comparisons use separate matched experimental protocols.

2. Independent HRNB contribution across seven host criteria.

We compare Host+Fine against Host+Fine+HRNB for each criterion:

Host criterionHRNB ΔPSNR ↑
Opacity+0.19625
FoV-style−0.00310
MesonGS+0.10642
GoDe grad-all+0.00058
PUP Fisher+0.00585
PWR+0.04135
FastGS-style+0.77519

HRNB provides positive mean PSNR increments for 6/7 host criteria, with different magnitudes depending on the native importance formulation (Appendix H.3, Table 30).

The Indoor/Outdoor host-prior comparisons in Table 29 further demonstrate the dependence on host construction. Together, these experiments establish the complementary contribution of HRNB beyond the Fine response.

Revision. We strengthened the HRNB motivation in Section 4.3 and summarized its contribution in Section 5.3. The simple-calibration comparison is reported in Appendix G.8 (Table 27), the host-prior interaction in Appendix H.1 (Table 29), and the seven-host incremental effects in Appendix H.3 (Table 30). The component ablation is reported in Appendix G.1 (Table 19).

AuthorsOpenReview ↗

Response part 3

[W4] Reconstruction quality under comparable total training time

Response. To assess whether RAGP's quality gains justify its additional probing and calibration costs, we extend the native OMG optimization from 30K to 35K and 40K, allowing the baseline to use the additional time for further optimization. We then compare these results with RAGP at 30K across the same 13 scenes.

Additional experiment. The results are:

MethodPSNR ↑SSIM ↑LPIPS ↓Time (s/scene) ↓
OMG 30K26.99460.83080.2337903
OMG 35K27.01990.83120.23341052
OMG 40K27.02760.83140.23331200
RAGP 30K27.09070.83170.23261193

At nearly identical cumulative runtime, RAGP 30K achieves +0.0631 dB higher PSNR, +0.0003 SSIM, and 0.0007 lower LPIPS than OMG 40K. Across all 13 scenes, their cumulative runtimes are 15,515.34 s and 15,604.50 s, respectively; OMG uses only 0.57% more time but still does not match RAGP's reconstruction quality.

Compared with OMG 30K, RAGP improves mean PSNR by +0.0961 dB. Under this matched 30K protocol, both methods share the same source state, Gaussian budget, and post-pruning recovery procedure, differing only in retained Gaussian identities. Thus, within the evaluated range up to 40K, the extended OMG baseline does not recover the quality gap at nearly matched runtime.

Runtime overhead. Across 13 scenes, RAGP 30K incurs 32.21% additional cumulative runtime relative to OMG 30K, including its probing and calibration costs. A separate three-scene stage-wise profiling experiment reports 19.66% overhead under its corresponding profiling protocol (Appendix J, Table 31). RAGP introduces no additional inference-time parameters or rendering operations.

Revision. Section 5.4 and Table 3 present the runtime–quality comparison. Appendix J.1 (Table 32) reports the 13-scene equal-time evaluation, while Appendix J (Table 31) provides the separate stage-wise runtime and sampled peak-VRAM measurements.

Reviewer_2RpzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a fixed-budget refinement framework based on the Gaussian importance criterion: Recoverable Sensing Gaussian Pruning (RAGP). RAGP combines intervention derived compensation evidence with the native host standards under a fixed Gaussian budget.

Critical strengths

The paper considers post-pruning recoverability. The use of counterfactual gradient responses provides an interesting perspective on Gaussian pruning and is supported by the experiments. RAGP improves reconstruction quality across five pruning budgets and seven importance criteria. The comparisons use matched Gaussian counts and recovery protocols, making the results more convincing. The method also introduces no additional parameters or rendering operations at inference time.

Critical weaknesses
  1. Unclear connection between gradient response and recoverability

It is unclear whether the proposed gradient response truly reflects post-pruning recoverability. The method seems more like an importance refinement approach based on counterfactual gradient responses, rather than directly measuring recoverability.

Although the freezing experiments show that high-response Gaussians contribute to recovery, it remains unclear whether this response provides information beyond the original importance scores. For example, how do the methods compare immediately after pruning, before further optimization? How much does each method improve during recovery?

In addition, RAGP requires continuous gradient probing during training. Can the response be computed independently at a fixed checkpoint or through offline probing?

  1. Insufficient justification for several design choices

The method only uses Scale and Rotation gradients to measure compensation. While the covariance tangent space provides a geometric explanation, Position, Opacity, and SH parameters also play important roles in post-pruning recovery. Ablations with different parameter groups would help justify this choice.

In Section 4.4, the motivation for the rank-based priority and exponential fusion is not fully explained.

Similarly, HRNB is motivated by differences in how host importance criteria handle projected-area normalization. However, its individual contribution is mainly evaluated with OMG. The experiments with seven importance criteria only evaluate the full RAGP method. A direct comparison of (Host+Fine) and (Host+Fine+HRNB) across these criteria would better show whether HRNB consistently provides additional benefits.

Questions for authors
  1. Could the authors further explain why counterfactual gradient responses can reflect post-pruning recoverability? Could you compare the reconstruction quality immediately after pruning and after recovery optimization to show whether the improvement comes from better initial selection or stronger recovery?

  2. RAGP collects gradient responses over several thousand training iterations. Is it possible to compute these responses at a fixed checkpoint without continuous training? How would reducing the number of probes affect reconstruction quality and computational cost?

  3. Why are only Scale and Rotation gradients used? Have the authors considered Position, Opacity, or SH gradients? An ablation across different parameter groups would help justify this choice.

  4. What motivates the rank-based priority and the exponential fusion in Eqs. (10)–(11)? In particular, why should lower-ranked Gaussians receive stronger adjustments from the response score? Could the authors compare this design with simpler fusion methods?

  5. Table 2 evaluates the full RAGP method across seven importance criteria, but does not isolate the effect of HRNB. Could the authors provide a comparison between Host+Fine and Host+Fine+HRNB for these criteria to show whether HRNB consistently improves performance?

  6. Comparison with learning-based pruning methods The paper discusses several learning-based and optimization-coupled pruning methods, such as LP-3DGS, MaskGaussian, GaussianSpa, and Prune Wisely, but does not include them in the quantitative comparisons. Could the authors provide comparisons with these methods under similar Gaussian budgets and recovery settings? In addition, the paper lacks comparisons with other pruning methods across different pruning ratios. It would be helpful to include PSNR, SSIM, and LPIPS curves at different retention ratios, along with qualitative comparisons at different compression levels.

4 comments
AuthorsOpenReview ↗

Response part 1

We thank the reviewer for the constructive questions regarding post-pruning recoverability and probing efficiency.

[Q1] Relationship between counterfactual response and post-pruning recoverability

Response. We distinguish the retrospective recoverability diagnostic CTC_T from the pre-pruning geometric response RigeoR_i^{\mathrm{geo}}. The former measures relative signed loss-gap closure after an adaptation trajectory, whereas the latter provides a local intervention signal of compensatory participation before pruning. Thus, RigeoR_i^{\mathrm{geo}} is not intended to estimate CTC_T directly.

At a shared pre-update state and camera, paired Scale/Rotation gradient contrasts are derivatives of the signed Masked–Normal loss gap. The Scale contribution to the first-order directional derivative is −∥Δgi,ms∥22νms-\frac{\lVert\Delta g_{i,m}^{s}\rVert_2^2}{\nu_m^s}, and the Rotation contribution is −(Δgi,mq)⊤PiΔgi,mqνmq-\frac{(\Delta g_{i,m}^{q})^\top P_i\Delta g_{i,m}^{q}}{\nu_m^q}, where Pi=I−qiqi⊤P_i=I-q_iq_i^\top is positive semidefinite and both normalizers are positive. Consequently, the directional derivative at α=0\alpha=0 satisfies dDmdα∣α=0≤0\frac{dD_m}{d\alpha}\big|_{\alpha=0}\leq 0. This establishes a local mathematical connection between the paired contrasts and the geometric response, without equating a one-step direction with multi-step recovery (Appendix A.2, Eq. 16).

Empirical evidence. We test the functional relevance of RigeoR_i^{\mathrm{geo}} by freezing the top 10% high-response retained Gaussians and comparing subsequent local recovery against equal-count random and low-response controls.

HorizonComparisonMean difference (95% CI)Negative scenes
100 stepsHigh − Random−0.031 [−0.042, −0.020]12/13
100 stepsHigh − Low−0.032 [−0.043, −0.020]12/13
500 stepsHigh − Random−0.011 [−0.016, −0.007]12/13
500 stepsHigh − Low−0.011 [−0.017, −0.007]12/13

Negative differences indicate weaker subsequent local recovery when high-response Gaussians are frozen. The consistent direction across 12/13 scenes supports interpreting RigeoR_i^{\mathrm{geo}} as pre-pruning evidence of compensatory participation, rather than a numerical estimator of multi-step CTC_T (Appendix F.2, Table 16).

Immediate and final reconstruction quality. Appendix F.3, Table 18 compares quality immediately after exact-K pruning with final quality after the shared native recovery and compression pipeline.

MethodImmediate PSNRFinal PSNR
Official OMG20.042227.0367
Full RAGP20.521627.1024

Full RAGP improves Immediate PSNR over Official OMG by +0.4794 dB and retains a +0.0657 dB advantage after the shared recovery and compression pipeline. Since both methods use the same 20K source state, exact Gaussian budget, and recovery protocol, the comparison attributes the quality difference to the selected Gaussian identities: the gain is already present immediately after pruning and remains after subsequent optimization.

Revision. We clarified the distinction between recoverability and geometric response in Section 3 and added the first-order derivation in Appendix A.2.

[Q2] Online/offline probing and probe-budget requirements

Response. We compare online measurements collected along the 15.5K–20K training trajectory with read-only measurements at a frozen 20K checkpoint. The frozen-checkpoint variant uses the same Fine partition and probe schedule as the online variant. Both use the same pruning state, Gaussian budget, and recovery protocol.

VariantFinal PSNRSSIMLPIPSCollection time (s/scene)
Online27.09650.83140.2328151.86
Frozen-20K offline27.07360.83070.2337253.80

Online probing improves mean PSNR by 0.0229 dB and reduces additional response-collection time by 40.16% relative to frozen-checkpoint probing (Appendix G.9, Table 28).

We further evaluate Fine, Medium, and Coarse spatial probing configurations under the same minimum-coverage criterion:

ConfigurationMean ΔPSNR (dB)Scheduled probes
Fine+0.2420526,194
Medium+0.2267511,568
Coarse+0.204628,317

Medium and Coarse reduce scheduled probes by 55.8% and 68.2%, respectively, relative to Fine, with progressively smaller PSNR gains. Each configuration uses its own coverage-feasible schedule while maintaining the minimum-observation requirement (Appendix G.5, Tables 23–24).

Revision. We added the online/frozen-checkpoint comparison in Appendix G.9 and the spatial-granularity and probe-count comparisons in Appendix G.5.

AuthorsOpenReview ↗

Response part 2

[Q3] Selection of Scale and Rotation parameters

Response. Scale and Rotation jointly determine a Gaussian's covariance, capturing both its spatial extent and orientation. Their paired gradient contrasts can therefore be mapped into a common six-dimensional covariance-tangent representation.

We additionally conducted an early parameter-group exploration across 13 scenes, before introducing the final covariance-tangent whitening and HRNB modules. The measured mean PSNR changes are:

Parameter groupMean ΔPSNR (dB)
Scale+0.043530
Rotation+0.055088
Position+0.038223
Opacity+0.037314
Appearance−0.005463
Scale + Rotation+0.052970
Scale + Rotation + Position+0.050650
Scale + Rotation + Opacity+0.043150
Scale + Rotation + Appearance+0.042303

Rotation-only gives a slightly higher PSNR gain than Scale+Rotation in this exploratory study. We therefore do not claim that the joint parameter group maximizes the exploratory PSNR metric. We retain Scale and Rotation because their combination captures both scale and orientation changes of the Gaussian covariance. Appearance uses the native OMG static_feature and view_feature parameters.

These results provide parameter-selection evidence, while the final covariance-tangent whitening and HRNB are evaluated separately in the final pipeline.

[Q4] Motivation for rank-based priority and exponential fusion

Response. Host criteria have different score scales and distributions. After HRNB, we convert calibrated log-host scores into a common descending rank, where rank 1 has the highest priority. We define:

πiH=(N−riH+1)/N,qi=(riH−12)/N,h(qi)=qi2(3−2qi).\pi_i^H=(N-r_i^H+1)/N,\quad q_i=(r_i^H-\frac{1}{2})/N,\quad h(q_i)=q_i^2(3-2q_i).

The final score is:

Sifinal=πiHexp⁡(h(qi)Rigeo).S_i^{\mathrm{final}}=\pi_i^H\exp(h(q_i)R_i^{\mathrm{geo}}).

The midpoint coordinate and smoothstep progressively increase response influence toward lower-ranked candidates. Since Rigeo∈[−1,1]R_i^{\mathrm{geo}}\in[-1,1] and h(qi)∈[0,1]h(q_i)\in[0,1], the exponential correction is positive and bounded by e−1e^{-1} and ee. Thus, host priority remains the ranking basis, while geometric evidence modulates it without replacing it.

Rank-based priority (Appendix G.3, Table 21). This independent matched diagnostic compares response-only ranking, uniform calibration, and host-rank calibration.

VariantΔPSNR ↑ΔSSIM ↑ΔLPIPS ↓
Response-only ranking-0.098-0.0189+0.0276
Uniform calibration+0.222-0.0005+0.0013
Host-rank calibration+0.220+0.0015-0.0021

Host-rank calibration achieves comparable PSNR to uniform calibration, while improving SSIM and LPIPS. Response-only ranking degrades all three metrics, supporting the host prior as the ranking basis.

Final fusion (Appendix G.1, Table 19). This final-pipeline ablation at budget X compares the selected configuration with tested alternatives.

ConfigurationΔPSNR ↑ΔSSIM ↑ΔLPIPS ↓
Whitened Fine + Uniform+0.09715-0.000931+0.002017
Whitened Fine + Odds-5/View+0.10478+0.001226-0.001612
Full RAGP+0.12083+0.001289-0.001609

Full RAGP achieves the highest PSNR and SSIM, with LPIPS comparable to the best alternative. This supports the selected bounded multiplicative fusion over the evaluated configurations. The two tables belong to separate matched suites and are interpreted only within their respective comparisons.

Revision. Section 4.4 clarifies rank-based priority, smoothstep modulation, and exponential fusion. Supporting results are in Appendix G.1 (Table 19) and G.3 (Table 21).

[Q5] Independent contribution of HRNB across host criteria

Response. We evaluate HRNB incrementally over Host+Fine under matched 13-scene conditions:

Host criterionIncremental HRNB ΔPSNR (dB)
Opacity+0.19625
FoV-style−0.00310
MesonGS+0.10642
GoDe grad-all+0.00058
PUP Fisher+0.00585
PWR+0.04135
FastGS-style+0.77519

HRNB improves mean PSNR for six of the seven criteria. The FoV-style result is slightly negative, while the GoDe and PUP gains are small (Appendix H.3, Table 30).

We also compare Full HRNB with simple projected-area normalization and simple residual correction. Full HRNB achieves the highest mean PSNR (27.1024 dB), 0.0235 dB above Host+Fine. The simpler controls obtain slightly higher mean SSIM and lower LPIPS in this separate comparison (Appendix G.8, Table 27).

Revision. Section 4.2 and Appendix A.2 detail the covariance-tangent construction; Section 4.4 explains rank-based fusion; Sections 5.3 and Appendices G–H report the component, calibration, and seven-host experiments.

AuthorsOpenReview ↗

Response part 3

[Q6] Extended cross-method and multi-budget evaluation

Response. We sincerely thank the reviewer for suggesting broader comparisons with existing pruning methods and evaluations across compression levels. We have extended the experiments from three complementary perspectives.

1. Cross-method benchmarking. We compare RAGP with representative Gaussian pruning and compression methods, including LP-3DGS, MaskGaussian, GaussianSpa, Prune Wisely, and other baselines, across three benchmark datasets. Figures 5, 8, and 9 report PSNR, LPIPS, and SSIM against model size, respectively, demonstrating the quality–compression trade-offs across methods.

These methods employ different native pruning and recovery pipelines, including multi-stage pruning and method-specific optimization schedules. We preserve their original pipelines to evaluate each method under its intended procedure, rather than altering them to enforce a common recovery schedule. Cross-method results therefore use their available model-size operating points, while strictly controlled exact-K comparisons are conducted separately.

2. Matched multi-budget evaluation. To isolate recoverability-aware Gaussian selection, we compare RAGP against Official OMG at five budgets (XS, X, M, L, and XL). Each pair shares the same source state, native host prior, exact Gaussian count, and recovery protocol; only the retained Gaussian identities differ.

BudgetΔPSNR ↑ΔSSIM ↑ΔLPIPS ↓PSNR-positive scenes
XS+0.14762+0.001722−0.00183913/13
X+0.12083+0.001289−0.00160913/13
M+0.12909+0.001164−0.00132613/13
L+0.11162+0.000761−0.00074213/13
XL+0.06078+0.000451−0.00047912/13

RAGP improves mean PSNR, SSIM, and LPIPS at every budget, achieving consistent benefits across different pruning strengths. The mean PSNR improvement remains positive in all five settings, with 13/13 positive scenes at XS–L and 12/13 at XL (Appendices C–D, Tables 8–13).

3. Qualitative comparisons across compression levels. We further extend the existing budget-X visual comparisons (Figures 7 and 10–12) with budgets M and XL across Tanks & Temples, Deep Blending, and Mip-NeRF 360 (Appendix I.2, Figures 13–15).

For each scene, the newly added M and XL comparisons use identical test viewpoints and crop regions. Within each budget, Official OMG and RAGP retain exactly the same number of Gaussians and follow the same recovery protocol. These visual comparisons illustrate reconstruction differences in local structures and appearance across representative compression strengths, complementing the five-budget quantitative results.

Together, the cross-method quality–size benchmarks establish broader competitiveness, while the controlled multi-budget experiments and corresponding visual comparisons demonstrate the benefits of recoverability-aware Gaussian selection across compression levels. The seven-host experiments further support its applicability to heterogeneous importance formulations under a unified pruning-and-recovery pipeline.

Revision. We extended the cross-method comparisons in Figures 5, 8, and 9, added five-budget quantitative evaluations in Appendices C–D, and expanded the qualitative comparisons in Appendix I.2 (Figures 13–15), covering representative compression levels across all three datasets.

AuthorsOpenReview ↗

Follow-up on Submission 16618

Dear Area Chair, 2Rpz

I hope you are doing well.

We would like to kindly follow up on our responses to the reviewers' comments for Submission 16618. We have provided point-by-point responses to the questions raised and made additional clarifications and experimental analyses where needed.

We would appreciate your guidance on whether our responses have adequately addressed the concerns, or if there are any remaining issues that we could further clarify or address.

Thank you very much for your time and consideration. We appreciate your efforts in coordinating the review process.

Best regards,

The Authors

Comments on the paper

1 comment

Thank you for your positive reassessment and constructive feedback. We appreciate your recognition of the additional experiments and the improvements under fair and equal-runtime comparisons.

We agree that the presentation of RigeoR_i^{\mathrm{geo}} can be made clearer. In the revised manuscript, we will consistently describe it as a complementary geometric importance signal that improves Gaussian selection during pruning, rather than as a direct measure or predictor of post-pruning recoverability. We will also revise the relevant discussions and conclusions to ensure that the claims accurately reflect the evidence presented.

Thank you again for your time and helpful suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.