acceptodds
Under review as a conference paper at ICLR 2027

L1ResPrune: General Pruning for Explicit Scene Representation Primitives

Abstract

In this paper, we present L1 Loss Residual Selective Pruning, termed L1ResPrune, a general pruning method for explicit scene representation primitives. L1ResPrune is applicable to both static and dynamic scenes and is compatible with representations based on Gaussians as well as other types of primitives. Building upon the standard alpha-blending formulation, we derive an importance score for each primitive based on the change in the L1 loss induced by its removal. This formulation provides a theoretically grounded criterion for pruning and does not rely on the specific characteristics of any particular representation primitive. Extensive experiments demonstrate that, compared with existing 3D Gaussian Splatting based pruning methods, L1ResPrune consistently preserves higher rendering quality while achieving superior SSIM and LPIPS performance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Explicit Scene Representations, Primitive Pruning, Neural Rendering, Scene Compression

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_44jgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

L1ResPrune derives an importance score from an analysis of the L1 reconstruction-loss change induced by removing alpha-blended primitives. The score combines an opacity factor alpha/(1 − alpha) with the magnitude of the L1 gradient and aggregates it across pixels and views. Non-gradient and integral approximations reduce computation. Evaluation covers static and dynamic scenes and several primitive types.

Critical strengths
  1. Broad empirical coverage. The paper evaluates static and dynamic Gaussian representations as well as alternative alpha-blended primitives.
  2. Useful perceptual improvements. At 90% pruning on Mip-NeRF360, Table 1 improves SSIM/LPIPS over Speedy-Splat from 0.776/0.298 to 0.785/0.281, although PSNR decreases from 26.74 to 26.62. Table 2 also reports improvements over SpeeDe3DGS on NeRF-DS.
  3. Practical approximations. The NGS and IAS variants offer lower-cost scoring, and the appendix studies aggregation choices and scoring factors.
Critical weaknesses
  1. The finite-deletion bound needs an additional assumption. Equation 8 ensures nonzero channel residuals at the expansion point, but does not prevent sign changes during complete deletion. Appendix A.7, Equations 19–29, establishes local mixed-derivative identities within a smooth L1 branch; these identities do not establish the general finite-deletion bound in Equation 9. A counterexample is one pixel and one primitive on a black background, with color c = (1, 1, 0), alpha = 0.6, and target y = (0.5, 0.7, 0.01). The mean RGB L1 loss is locally constant at 0.07, with all derivatives zero and all channel residuals nonzero. Deletion raises the loss to 121/300, an increase of 1/3, while Equation 9 has a zero right-hand side. Equation 13 likewise assigns zero importance. This challenges the stated bound, not the empirical usefulness of the score.
  2. The independent score does not inherit a joint-deletion guarantee. Even within a fixed-sign region, Equation 12 contains interactions among removed primitives. The transition to Equation 13 does not establish that the independent ranking controls their combined removal error. The opacity reweighting can be empirically useful, but the derivation does not justify a general non-heuristic error guarantee.
  3. Cross-primitive evidence does not directly test the main L1 score. The setup on page 6 specifies the gradient-free NGS variant for alternative primitives. Table 3 therefore supports applicability of the opacity-based simplification rather than generality of the central L1-gradient score. Matched-budget comparisons with other generic pruning criteria would strengthen this claim.
  4. Evaluation details require reconciliation. Table 5 reports 27.26/0.806/0.246 for Full, n = 2, Visibility, high-ratio pruning, whereas Table 7 reports 27.22/0.801/0.252 for the corresponding setting; the former matches the Max row in Table 7. The appendix also states that test views are sampled from training views. This may be imprecise wording, but camera splits and scoring-view usage should be clarified before concluding that evaluation is independent.
Questions for authors
  1. Can the authors reconcile Equation 9 with this counterexample? Does the intended claim assume that every channel residual retains its sign throughout deletion? Please state and justify any such restriction, account for sign changes, or explicitly frame the analysis as a local approximation.
  2. What formal relationship connects Equation 13 to simultaneous deletion error, and which claims should instead be framed as empirical motivation?
  3. Can the full gradient-based score be evaluated across primitive types? Please reconcile Tables 5 and 7 and specify whether optimization, scoring, and test cameras are disjoint where required.
2 comments
AuthorsOpenReview ↗

Appreciation for the Thorough Analysis and Constructive Suggestions

W1.

We sincerely thank the reviewer for pointing out this important issue. We have address this concern in our public response (Concern 1) and provided further clarification. We have also revised the manuscript accordingly.

We focused too much on the derivation itself and overlooked a subtle condition for the validity of the Taylor expansion, namely, the smoothness of the function between the expansion point and the evaluation point. As explained in our public response, this issue can be resolved with a minor modification by replacing LL with CC. The resulting importance scores are nearly identical to the original ones, and the experimental results also remain essentially unchanged.

W2.

We address this concern in our public response (Concern 2).

W3.

In our public response (Concern 3), we have provided additional comparisons using the Full formulation. We have further improved the comparison in the revised manuscript.

W4.

We sincerely apologize for these errors.We conducted extensive experiments with various parameter configurations. However, we inadvertently reported the results for n=1n=1 in the main paper, while those in the appendix correspond to n=2n=2. We have corrected this inconsistency in the revised manuscript.

Regarding the camera views, we sincerely apologize for the misleading wording. What we intended to convey is that regions with denser training-view coverage also tend to have more test views, rather than that test views are sampled from training views. The training and test sets are strictly separated, and test images are never used to compute importance scores or train. We will revise the Appendix to clarify this distinction.

Q1.

The Taylor expansion is based on rigorous mathematical analysis. The subsequent equations are more intuitive, empirically motivated formulations.

Q2.

see W2, and We address this issue in our public response (Concern 2).

Q3.

We address this issue in our public response (Concern 3).

Thank you for taking the time to address my questions. However, I believe that the pruning score used in Mini-Splatting could also be applied to other types of primitives; therefore, its applicability across different representations is not, in itself, sufficient to establish this method's advantage over that baseline. For Triangle Splatting and 3DCS, the improvements offered by L1ResPrune are quite limited—with PSNR gains of only 0.12 dB and 0.10 dB, respectively—and the differences in SSIM/LPIPS are also marginal. While I acknowledge the more significant improvements observed with 2DGS, the current results do not yet provide sufficient evidence to support a clear performance advantage across different representations. I would be interested to hear the authors' perspective on this; for now, I maintain my initial assessment. If the authors have any further points to add, please feel free to provide a brief response before the deadline.

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

To address primitive redundancy in explicit scene representations, the authors develop L1ResPrune, a pruning framework based on L1 reconstruction-loss sensitivity. Unlike existing approaches that primarily rely on heuristic or representation-specific pruning criteria, the proposed method estimates primitive importance by analyzing the impact of primitive removal on the L1 reconstruction loss under alpha blending. Based on this formulation, the authors develop a primitive importance score that accounts for both first-order sensitivity and higher-order interactions. Different aggregation strategies are explored to obtain global importance scores, and two simplified variants, NGS and IAS, are introduced to reduce computational overhead. The method is evaluated on static and dynamic scene reconstruction benchmarks, as well as different primitive representations. Experimental results demonstrate competitive rendering quality under various pruning ratios, with notable improvements in SSIM and LPIPS.

Critical strengths
  1. A meaningful and relevant problem. Reducing redundant primitives while preserving rendering quality is an important problem for efficient explicit scene representations. Deriving pruning criteria from rendering loss rather than relying exclusively on heuristic importance measures is a promising direction.

  2. Broad experimental evaluation. The paper evaluates the proposed method on multiple static and dynamic datasets and considers several primitive representations, including 2D Gaussians, convex splatting, and triangle splatting. This is a useful attempt to demonstrate applicability beyond conventional static 3D Gaussian Splatting.

  3. Consideration of computational efficiency. The NGS and IAS approximations offer practical alternatives to the full gradient-based importance formulation. The ablation studies provide useful evidence of the trade-off between pruning quality and importance-score computation time.

Critical weaknesses
  1. Validity of the theoretical analysis. In Eqs. (9)–(12), primitive removal is analyzed through a local Taylor expansion of the L1 loss. However, setting an opacity to zero is a finite change that may cross nondifferentiable points. For instance, L(α)=∣α−0.1∣+∣α−0.9∣L(\alpha)=|\alpha-0.1|+|\alpha-0.9| satisfies L′(0.5)=0L'(0.5)=0, although L(0)−L(0.5)=0.2L(0)-L(0.5)=0.2. The authors should clarify the assumptions under which the expansion characterizes the actual pruning-induced loss.

  2. Unclear relationship between the importance score and actual loss changes. Eq. (13) defines si=(αi1−αi∂L1∂αi)ns_i=\left(\frac{\alpha_i}{1-\alpha_i}\frac{\partial L_1}{\partial\alpha_i}\right)^n at the pixel level. It is unclear whether ranking primitives independently by this score captures the interactions involved in removing multiple primitives. A correlation analysis between estimated importance and measured loss changes, together with an ablation under the main n=2n=2 setting, would strengthen the justification.

  3. Inconsistencies in the experimental results. The per-scene PSNR values for L1ResPrune in Table 14 average to approximately 38.83 on D-NeRF and 23.79 on NeRF-DS, whereas Table 2 reports 38.74 and 23.74. Additionally, Table 15 lists identical per-scene SSIM results for Mini-Splatting and SpeeDe3DGS, despite different aggregate SSIM values in Table 2. These discrepancies need clarification.

  4. Limited evidence for representation generality. The non-Gaussian experiments use NGS, which removes the L1-gradient term central to the proposed formulation. Furthermore, Table 3 reports the same primitive count (1.032M) for 3DCS before and after applying L1ResPrune. Evaluating the full score on non-Gaussian representations and using matched primitive budgets would better support the generality claim.

  5. Incomplete efficiency evaluation. The paper mainly reports rendering quality and primitive counts. Beyond the score-computation timings in Table 6, comparisons of rendering FPS, GPU memory consumption, and end-to-end pruning overhead would help establish the practical efficiency benefits.

Questions for authors
  1. How well does the proposed importance score in Eq. (13) correlate with the actual reconstruction loss changes caused by primitive removal? Could the authors provide a quantitative analysis to validate this relationship?

  2. Could the authors provide additional comparisons of rendering speed, GPU memory consumption, and overall pruning overhead to better demonstrate the practical advantages of L1ResPrune?

5 comments
AuthorsOpenReview ↗

Appreciation for Identifying Potential Issues and Suggestions

W1.

We sincerely thank the reviewer for identifying this subtle and easily overlooked issue. This issue arises from the presence of a non-differentiable point between the expansion point and the evaluation point. We addresse this concern in our public response (Concern 1). Furthermore, we have corrected this issue in the revised manuscript.

W2.

We sincerely thank the reviewer for raising this valuable question

We addresse this concern in our public response (Concern 2).

We will revise the manuscript to clarify this point.

W3.

We sincerely apologize for these transcription errors. We conducted a large number of experiments, including more than ten configurations of L1ResPrune with different values of nn and aggregation strategies. The results of all methods were collected in a single spreadsheet, and some values were inadvertently copied from incorrect entries during manuscript preparation.

We will carefully verify and correct these inconsistencies in the revised manuscript. Importantly, we have provided our code to ensure that the experimental results can be independently reproduced.

W4.

We sincerely thank the reviewer for suggesting this rigorous evaluation protocol. As detailed in our public response (Concern 3), we have addressed this concern by fixing the number of pruned primitives to ensure a fair comparison.

W5.

GPU Memory Usage and Computational Efficiency.

We would like to clarify the GPU memory usage of different pruning methods. Existing pruning methods, particularly high-performing ones such as Speedy-Splat and Mini-Splatting, generally extract their importance scores directly from the backward pass. For methods that compute importance scores explicitly, the GPU memory consumption typically does not exceed that required during training. In practice, GPU memory usage is primarily determined by the peak number of Gaussian primitives during training. The results are presented below, where we also report FPS for comparison.

Rendering FPS and GPU memory usage on NVIDIA RTX PRO 6000.

MethodFPS ↑\uparrowGPU Memory (GB) ↓\downarrow
3DGS1959.9
Speedy-Splat7349.9
MaskGS3555.5
L1ResPrune7439.9

Notably, MaskGS achieves lower GPU memory usage because it progressively prunes a small number of Gaussians during the densification stage, preventing the number of Gaussians from reaching its usual peak. In fact, our method can also reduce GPU memory usage by periodically applying a small pruning ratio during the early training stages. This is primarily an engineering strategy rather than a fundamental limitation of the pruning method.

In practical applications, computational time is often a more significant concern than GPU memory usage. The main computational overhead arises from importance score calculations that require forward or backward propagation, especially when processing large numbers of high-resolution training images. This is precisely why we introduce our simplified formulations, which avoid the need for forward or backward propagation during importance score computation.

Regarding MaskGS runtime, a direct comparison is not straightforward, as the results depend heavily on the pruning strategy and the number of training views. Assuming an appropriate pruning strategy is applied during the early training stages, our method can achieve a comparable end-to-end runtime when the number of training views is relatively small. However, when the number of training views becomes very large, MaskGS have a clear efficiency advantage.

The main bottleneck of score-based method is that each pruning operation requires an additional pass over the training images to compute importance scores, whereas MaskGS(End-to-End) can obtain these scores as a byproduct of the training process.

Q1.

This does not directly reflect changes in the reconstruction loss, since our mathematical model is defined at the single-pixel level. Moreover, introducing the exponent nn assigns different weights to different pixels, meaning that the resulting importance scores cannot directly represent changes in the reconstruction loss.

We would also like to highlight an important observation: the rendering quality immediately after pruning does not necessarily reflect the final reconstruction quality. In fact, at a 90% pruning ratio, our method initially achieves lower rendering quality than Speedy-Splat immediately after pruning, but outperforms Speedy-Splat after fine-tuning.

Furthermore, we would like to clarify the role of nn. We introduce nn to adjust the relative weighting of different score contributions for each primitive. It serves as a hyperparameter to accommodate different types of representation primitives, which have different geometric distributions.

Q2.

see W5.

Thank you for the authors' response and the additional experiments. I appreciate their efforts to address the concerns, particularly their acknowledgment of the issue with the original Taylor expansion.

However, my main concern remains unresolved. The revised formulation shifts the analysis from L1 loss changes to rendered-color differences, but it still does not establish a rigorous connection between the theoretical derivation and the final importance score. The authors also acknowledge that the aggregated score does not directly reflect the actual reconstruction loss change, which further weakens the central theoretical motivation of the paper.

While some experimental concerns have been addressed, these clarifications do not fundamentally change my initial assessment. Given the remaining theoretical gap and other concerns raised in my review, I will maintain my original rating.

We sincerely apologize for not clarifying this distinction sufficiently. We provide a simple and intuitive example to illustrate the relationship between our theoretical derivation and the actual reconstruction loss change.

Consider two primitives, A and B, each contributing to two pixel-channel locations. Primitive A has pixel-level importance scores of 0.1 and 0.4, while primitive B has scores of 0.3 and 0.3. For simplicity, we assume that these scores are equal to the corresponding loss changes. Thus, the reconstruction loss changes caused by removing A and B are:

  • Primitive A: (0.1 + 0.4) / 2 = 0.25.
  • Primitive B: (0.3 + 0.3) / 2 = 0.30.

When n = 2, the global importance scores of A and B are 0.17 and 0.18, respectively. Therefore, A is pruned, resulting in a reconstruction loss change of 0.25.

However, when n = 3, their global importance scores become 0.065 and 0.054, respectively. Consequently, B is pruned instead, resulting in a larger reconstruction loss change of 0.30.

Theoretically, this discrepancy arises because our derivation is performed at the single-pixel, single-channel level, whereas the reconstruction loss is evaluated over all pixels and color channels. When constructing the global importance score from pixel-level scores, we introduce an exponent n as a heuristic hyperparameter rather than a theoretically derived term. Consequently, the global importance score does not necessarily correspond directly to the actual reconstruction loss change.

In other words, our theoretical derivation establishes the pixel-level importance score. Due to the introduction of the exponent nn, the two are not necessarily fully consistent.

Moreover, the immediate loss change caused by pruning does not necessarily correlate with the final reconstruction quality after fine-tuning. Therefore, we introduce the exponent nn to provide a more flexible criterion for ranking primitives. It is worth emphasizing that our theoretical derivation is established at the single-pixel, single-channel level, rather than over all pixels and channels globally.

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

This paper proposes L1ResPrune, a theoretically grounded pruning framework for explicit scene representation primitives. Motivated by the idea that removing less important Gaussians has a negligible impact on the rendering loss, the authors define the transmittance compensation term and the partial derivative of the loss w.r.t. opacity as the key factors influencing the rendering loss through an analysis of the Taylor expansion of loss perturbation caused by removing a Gaussian. Consequently, the authors propose a power of the product of these key factors as the importance score in their framework, with different choices of score aggregation methods among global, visibility, and maximum aggregation, depending on the base primitive representation type. The benchmarking experiments show that L1ResPrune achieves rendering performance comparable to that of a SoTA method, Speedy-Splat, under all tested pruning ratio settings. In the comparison of dynamic 3DGS pruning methods, L1ResPrune outperforms prior works only on a limited dataset (NeRF-DS). The authors also demonstrate L1ResPrune's applicability to non-Gaussian primitive representations.

Critical strengths
  1. The authors provide logical math work to derive effective importance score.
  2. The extensive experiments show the proposed method, L1ResPrune, can achieve comparable performance to SoTA pruning methods.
Critical weaknesses
  1. Most crucially, the proposed score function is very similar to that of SpeedySplat, while the quantitative results of SpeedySplat and L1ResPrune are also nearly indistinguishable. This raises a substantial concern about the novelty and practical significance of the proposed modification, as it appears to introduce only a minor change to SpeedySplat's Hessian-driven score function without yielding a meaningful improvement in performance. How do the authors differentiate the novelty of this work from Speedy-Splat? As stated in the manuscript, when n=2, the proposed L1-residual-based importance score is similar to SpeedySplat's formulation, with the main difference being the use of opacity instead of the alpha contribution term. Given that the two methods achieve almost identical performance even under extreme pruning ratios, it is difficult to see this modification as practically meaningful or to support the claim that it "... explain the superior robustness of L1ResPrune under extreme pruning." (line 384)
  2. The paper's writing flow does not follow a typical progression, and moreover, the Related Work section contains misleading or incorrect summaries.
  • The Introduction is overly summarized. It discusses high-level limitations of existing pruning methods without providing supporting evidence or examples. The text should properly explain the problems or research gaps in prior works and the motivation for the proposed method.
  • The authors state, "Existing pruning methods can be broadly classified into three categories" (in Line 48). However, technically, Adaptive Density Control is not a pruning method but rather a method for controlling Gaussian generation.
  • In line 62, the authors state that "early pruning may affect convergence and limit the achievable pruning ratio." This statement might be incorrect because LP-3DGS does not train pruning masks during the densification phase, unlike Compact3DGS and MaskGS. Moreover, why is LP-3DGS not mentioned in Section 2.2 even though it uses learnable pruning masks?
  • In line 67, it is stated that "The third category removes redundant primitives after optimization." However, FastGS and SpeedySplat do not fall into this "post-training pruning" category, since they perform pruning during training.
Questions for authors
  1. What is direct pruning? This term sounds weird, since methods such as Distilled-3DGS and LocoGS do not seem to employ any kind of indirect pruning method. Distilled-3DGS transfers scene structure from a large set of Gaussians to a small set of Gaussians. LocoGS employs a trainable mask to remove less-contributing Gaussians.
  2. Where does the "adaptive" property come into play in Adaptive pruning (line 248)? Rather than Adaptive pruning, "thresholded pruning" may more accurately reflect its predefined threshold characteristic.
  3. Section 6 needs to be moved to Section 5.4.
  4. The Figure 1 caption appears to give an incorrect description. The visual showcases compare static 3DGS pruning methods, not dynamic scenes.
  5. L1ResPrune and Speedy-Splat tie in rendering performance. The minor improvement of L1ResPrune over Speedy-Splat might be due to randomness. Averaging the results over multiple runs, possibly at least 3 times, would provide stronger evidence as to which method is superior.
  6. In the experiments, how do the authors determine the pruning operation timing at 15k iterations? Pruning at a later stage may produce different results because the fine-tuning steps update the Gaussians, such that they would have different importance distributions under which baseline pruning methods might perform better than they do at 15k.
  7. Is the reported remaining Gaussian count of 3DCS (w/ p) and 3DCS+L1ResPrune correct (Table 3)? Are they both 1.032M Gaussians by coincidence, or is this a mistake?
  8. What is the purpose of applying the method to non-Gaussian primitive representations? The experiment in this section simply demonstrates its applicability to non-Gaussian primitive representations. What the authors need to show is that the proposed method is generally better than other pruning methods, as the authors do in Section 5.3, if they truly want to claim L1ResPrune's generalizability to non-Gaussian primitive representations.
  9. Calling it a "general pruning framework" is an overclaim because the proposed method is confined to representations that rely on alpha-blending-based rendering. For example, the following explicit representation methods are not applicable to L1ResPrune because they do not use alpha blending.
  • Jorge Condor, et al., Don't Splat your Gaussians: Volumetric Ray-Traced Primitives for Modeling and Rendering Scattering and Emissive Media, ACM TOG 2025.
  • Chen Ziwen, et al., Softmax-GS: Generalized Gaussians Learning When to Blend or Bound, CVPR FINDINGS Track (CVPRF 2026).
  1. Please disclose which dataset is used to produce the results in Table 7 in the nearby text referring to Table 7.
1 comment
AuthorsOpenReview ↗

Appreciation for the Review and Clarifications

We sincerely thank the reviewer for the detailed guidance on improving the manuscript's writing and presentation. We will carefully incorporate these suggestions into our revisions.

W1.

There are substantial differences in terms of definition, theoretical analysis, applicability, and performance. We have provided a detailed explanation in our public response (Concern 4).

W2.

We sincerely thank the reviewer for these thoughtful and detailed suggestions. We will revise the description.

However, we may have a slightly different perspective on the distinction between training-based and post-processing methods.

We distinguish training-based and post-processing pruning methods primarily based on whether their importance scores require additional training. For example, mask-based methods require training to learn the mask values, whereas importance-score-based methods can directly compute the scores without additional training.

In fact, almost all pruning methods can be applied during training by simply using a lower pruning ratio and a higher pruning frequency. However, classifying methods such as Fast-GS as training-based rather than post-processing methods solely because pruning is introduced at an early training stage may not be appropriate. Under such a criterion, virtually all pruning methods could be classified as training-based, making the distinction between training-based and post-processing methods meaningless.

Q1, Q2, Q3, Q4

We sincerely thank the reviewer for pointing out these detailed issues. We have addressed some of them in the revised manuscript and will make further corrections to ensure that all issues are fully resolved.

Q5. See Public response (Concern 4) Compared with real-world applications, Mip-NeRF360 contains several relatively simple scenes. However, on more complex scenes such as Bicycle, Garden, and Treehill, our method outperforms Speedy-Splat by 0.020 in SSIM and 0.034 in LPIPS. Improvements are also observed across the remaining, relatively simpler scenes. However, averaging the results across all scenes reduces the overall performance gap, making these improvements appear less pronounced.

Furthermore, in dynamic scenes, our method outperforms SpeeDe3DGS, the dynamic extension of Speedy-Splat.

Q6.

Delaying pruning generally improves performance. Our focus is on comparing methods rather than maximizing absolute performance. Pruning at 15K iterations is sufficient for comparison and has also been adopted in prior work, such as GHAP and PUP3DGS.

Q7.

Record the primitive count and retain the corresponding number of top-ranked primitives by importance score to ensure fair comparisons.

More simply, we can directly set a fixed budget.

Q8, Q9

We provide a detailed clarification in our public response (Concern 3).

Q10.

Datasets(Mip-MeRF360). We will further refine and improve the Appendix accordingly.

Comments on the paper

4 comments
AuthorsOpenReview ↗

Response to Key Concerns Shared by the Reviewers (Concern 1)

We sincerely thank the reviewer for the careful review and constructive feedback. Following the reviewer's suggestions, we have made several revisions to the manuscript, with further improvements still underway. Below, we first address the four key concerns raised by the reviewer: (1) the validity of the Taylor expansion, (2) the generalizability of our method, (3) the motivation for introducing nn and (4) the novelty and magnitude of performance improvements

(1)Taylor Expansion Issue

We sincerely thank Reviewer 44jg and QKQf for identifying this important and subtle issue.

The issue arises because the L1 loss may contain a non-differentiable point between MαM\alpha and α\alpha, introduced by the ground-truth pixel value. In such cases, the Taylor expansion used in our original derivation is not generally valid.

Fortunately, this issue can be resolved with a minor modification to Eq.(7)-(13) without changing the underlying motivation of our method. We have corrected this error in the revised manuscript.

Our solution is presented below. Specifically, we reformulate the mathematical model in terms of the change in the rendered color of a single pixel and color channel before and after pruning, rather than the change in the L1 loss. Accordingly, Eq.(7) is modified to express the difference between C(Mα)C(M\alpha) and C(α)C(\alpha), i.e. ∣C(Mα)−C(α)∣|C(M\alpha)-C(\alpha)| with the Taylor expansion performed around C(α)C(\alpha). This formulation preserves essentially the same interpretation as our original model while avoiding the invalidity of the Taylor expansion, since the C(α)C(\alpha) is smooth with respect to the α\alpha under the considered rendering model. $|C(M\alpha)-C(\alpha)| $ retains a practical interpretation similar to that of the original model.

According to Eq. (22) in the Appendix (Revised Version), we only need to replace L1L_1 with CC in Eqs.(8)–(13) and Eq.(13) is reformulated to represent the importance score at the single-view, single-pixel, single-channel level. Since the absolute value of the gradient of L1L_1 with respect to CC is 1 wherever it is differentiable, the only difference from the original formulation is that we now raise the contribution of each color channel to the n-th power before summation, rather than summing the contributions across the three color channels before raising the result to the n-th power.

We have already conducted preliminary experiments using this slightly modified formulation, and the results are nearly identical to those reported in our paper(PSNR ≤\le 0.02, SSIM ≤\le 0.001, LPIPS ≤\le 0.001), indicating that the impact of this modification is negligible. Nevertheless, we will re-run the experiments using the corrected formulation to ensure the consistency and rigor of our experimental evaluation.

AuthorsOpenReview ↗

Response to Key Concerns Shared by the Reviewers (Concern 2)

(2). the motivation for introducing nn

Reason

From the perspective of single-pixel importance, regardless of the value of nn, the relative ranking of primitives is determined by ### (2). the motivation for introducing nn

Reason

From the perspective of single-pixel importance, regardless of the value of nn, the relative ranking of primitives is determined by α1−α∣∂C∂α∣\frac{\alpha}{1-\alpha}|\frac{\partial C}{\partial \alpha}|. Since exponentiation preserves this ranking, different values of nn are equivalent at the single-pixel level.

The primary role of nn is to adjust the weighting during aggregation. Consider a simple example: primitive A covers 1000 pixels, each with an importance score of 0.01, while primitive B covers only 20 pixels, each with a score of 0.4. Intuitively, B is more important, whereas A resembles low-opacity noise. However, when n=1n=1, we obtain SA>SBS_A>S_B, causing A to be retained while B may be pruned. In contrast, when n=2n=2, we have SA<SBS_A<S_B, allowing the more important primitive B to be preserved.

Therefore, nn serves as a hyperparameter that, particularly at high pruning ratios (e.g., 80%), allows us to favor primitives with different importance score distributions, such as those with broad spatial coverage or high importance scores at individual pixels and color channels. We will provide a more detailed explanation in future revisions.

In our experiments, for high-ratio pruning, we observe that n=2,3,4n=2,3,4 consistently outperform n=1n=1 for Gaussian primitives, whereas for triangle primitives, n=1n=1 performs significantly better than n=3,4n=3,4.

Comparison of different values of nn across Gaussian and triangle primitives. L1ResPrune adopts global aggregation.

MethodPSNR$\uparrow$SSIM$\uparrow$LPIPS$\downarrow$Ratio
3DGS (Mip-NeRF360)
L1ResPrune (n=1n=1)26.490.7810.28290%
L1ResPrune (n=2n=2)26.620.7850.28190%
Triangle Splatting (Mip-NeRF360)
L1ResPrune (n=1n=1)25.870.7680.30880%
L1ResPrune (n=2n=2)25.370.7490.33680%

Relationship with the Loss

There is no direct relationship between the aggregated importance score and the loss change, as our importance score is defined locally, while the global score is obtained through aggregation. The role of nn is primarily to adapt the aggregation to different primitive types. Therefore, the aggregated score does not directly reflect the loss change before and after pruning. It is primarily designed to account for the differences in post-pruning fine-tuning capabilities across different primitive types

In fact, under 90% pruning on 3DGS in scene bicycle, our method achieves worse rendering quality than Mini-Splatting immediately after pruning, but outperforms it after fine-tuning. One possible explanation is that a larger nn tends to favor primitives with higher importance magnitudes over those with larger spatial coverage, thereby allowing some large-coverage primitives to be pruned. The remaining smaller but high-importance primitives may facilitate more effective scene recovery during fine-tuning.

Comparison of training-set performance before and after fine-tuning on the Bicycle scene at a 90% pruning ratio. All methods are pruned at 15K iterations and fine-tuned for an additional 15K iterations. L1ResPrune adopts n=2n=2 with global aggregation.

MethodStagePSNR ↑SSIM ↑LPIPS ↓
Speedy-SplatBefore Fine-tuning17.6510.53160.4334
Speedy-SplatAfter Fine-tuning23.8820.70120.3541
L1ResPruneBefore Fine-tuning16.2910.51530.4413
L1ResPruneAfter Fine-tuning24.2420.74090.3100
AuthorsOpenReview ↗

Response to Key Concerns Shared by the Reviewers (Concern 3)

(3).Generalizability

All reviewers have raised this concern, primarily regarding the experimental evaluation. We have therefore conducted additional experiments to address it. We will also explain how our method can be extended to Softmax-GS.

For Speedy-Splat, we follow its original formulation in 3DGS by computing gradients with respect to the kernel function KK. We additionally evaluate gradients with respect to the opacity parameter and α\alpha on Triangle Splatting and 3DCS. GHAP is designed specifically for Gaussian primitives and therefore cannot be directly applied.

We have additionally evaluated Mini-Splatting, as its importance score is based on ray contributions and does not depend on primitive-specific parameters. We also attempted to apply Speedy-Splat to 2DGS. The results are presented below.

For a fair comparison, we strictly align the pruning budgets across all methods, using a fixed pruning ratio of 80%, with L1ResPrune adopting its Full formulation and pruning starting after densification.

Pruning results across different primitives at an 80% pruning ratio. The complete importance scores for different primitive types are presented below:

MethodPSNR$\uparrow$SSIM$\uparrow$LPIPS$\downarrow$
Triangle Splatting
Opacity24.930.7200.357
Naive Pruning Method25.760.7540.323
Mini-Splatting25.750.7600.312
Speedy-Splat (opacity)22.480.7080.369
Speedy-Splat (alpha)21.620.7020.369
L1ResPrune (n=1n=1)25.870.7680.308
3DCS
Opacity26.720.7820.307
Mini-Splatting26.800.7930.280
Speedy-Splat (opacity)25.720.7600.324
Speedy-Splat (alpha)26.040.7670.315
L1ResPrune (n=2n=2)26.900.7940.274
2DGS
Mini-Splatting26.030.7640.367
Speedy-Splat26.350.7730.360
L1ResPrune (n=2n=2)26.370.7800.347

Softmax-GS Although Softmax-GS does not use conventional alpha blending, its formulation is essentially analogous to alpha blending. Therefore, our method can be extended by simply replacing α\alpha with the corresponding Softmax-GS weight α˚\mathring{\alpha}. Since the method has not yet been open-sourced, we have not reproduced the experiments. We will evaluate this extension once the code becomes publicly available.

Ray-Tracing Although our method is presented based on alpha blending, it actually provides a general analytical framework. Alpha blending is only explicitly used in the final step, where we decompose the higher-order derivatives of C with respect to α\alpha. The rendered color C can also follow other formulations, provided that its higher-order derivatives can be appropriately decomposed or approximated. However, for ray-tracing-based rendering, we have not yet identified a suitable decomposition or approximation of the higher-order derivatives of C. We acknowledge this limitation and will include it in the paper.

We sincerely thank reviewer ZbRz for pointing out this issue. We commit to revising the title and relevant statements in the manuscript to avoid overclaiming, while emphasizing the broader applicability of our method compared with previous approaches.

AuthorsOpenReview ↗

Response to Key Concerns Shared by the Reviewers (Concern 4)

(4s).Novelty and magnitude of improvement

We thank the reviewer ZbRz for highlighting this concern. We will further clarify these distinctions in the revised paper.

Our method differs significantly from Speedy-Splat and PUP3DGS in five key aspects: Theoretical Rigor, Alignment with Pruning, Explainability, Formulation and Improvement

Theoretical Rigor. PUP3DGS(or Speedy-Splat) focus on second-order gradient, without analyzing first- and higher-order gradient. Moreover, the fisher approximation has a notable limitation. Specifically, the Hessian H\mathbf{H} in PUP3DGS can be decomposed as H=F+B\mathbf{H}=\mathbf{F}+\mathbf{B}, where F\mathbf{F} is the Fisher-related term and B=L1⋅C\mathbf{B}=L_1\cdot\mathbf{C} is a residual-dependent term. Neglecting B\mathbf{B} solely because L1L_1 is small is insufficient. The relative magnitudes of F\mathbf{F} and B\mathbf{B} must also be considered. Even if both terms are small, neither is necessarily negligible relative to the other. In contrast, our method systematically analyzes the contributions of all expansion terms, from first to infinite order.

Alignment with Pruning. PUP3DGS evaluates importance score through gradients with respect to positional parameters. However, position only reflects the spatial movement of a primitive, rather than whether it participates in rendering. In fact, gradients with respect to opacity or α\alpha should be used. Changes in opacity or α\alpha reflect the presence of a primitive in rendering. Therefore, gradients with respect to opacity or α\alpha should be used. Our mathematical model is consistent with the actual definition of pruning.

Explainability. Speedy-Splat replaces the position- and scale-related gradients in PUP3DGS with gradients with respect to GG, primarily to improve computational efficiency, as GG incorporates both position and scale information. However, it does not explain the reasons behind the resulting performance improvement.

We think that its performance improvement may arises from the following coincidence. Since α=oG\alpha = oG, taking the gradient with respect to GG and set n=2n=2 happen to satisfy condition $\frac{\partial C}{\partial G} =o\frac{\partial C}{\partial \alpha} $. This coincidence inadvertently leads to an improvement in pruning quality.

In contrast, our method provides a theoretical explanation, to some extent, for the performance improvements in Speedy-Splat, shedding light on a phenomenon that was not adequately explained in prior work. This theoretical insight represents another important distinction between our method and Speedy-Splat.

Formulation Our importance score is defined differently from that of Speedy-Splat. The similarity between these expressions arises from the fact that, in many primitive-based representations, the opacity α\alpha can be expressed as oKoK, where KK denotes the kernel function. This coincidental factorization unexpectedly leads to similar mathematical formulations.

Ours: ∑[α1−α∣∂C∂α∣]n\sum [\frac{\alpha}{1-\alpha} \left | \frac{\partial C}{\partial \alpha} \right |]^n

Speedy-Splat or PUP3DGS: ∑(∇uC)(∇uC)T\sum (\nabla_{u} C)(\nabla_{u} C)^T, where uu denotes a parameter of the primitive. In Speedy-Splat, uu is GG.

Improvement. At a 90% pruning ratio, the performance differences among pruning methods are only determined by the quality of the remaining 10% of the initial Gaussian primitives. More importantly, over half of the scenes in Mip-NeRF360 are relatively less complex than challenging real-world scenes. When the evaluation metrics are averaged across all scenes, these relatively simple scenes tend to reduce the overall performance differences between methods.

In fact, Bicycle, Garden, and Treehill are among the highly challenging outdoor scenes in Mip-NeRF360 in terms of geometric complexity and texture richness, featuring intricate structures, dense vegetation, and fine-grained details. On these three challenging scenes, L1ResPrune outperforms Speedy-Splat by 0.020 in SSIM and 0.034 in LPIPS on average. These improvements, particularly the reduction in LPIPS, are substantial for a pruning task and further demonstrate the effectiveness of our method in geometrically and texturally complex scenes.

For dynamic scenes, our method outperforms SpeeDe3DGS, the dynamic extension of Speedy-Splat.

For non-Gaussian primitives, In comparison with our method, Speedy-Splat achieves inferior performance on Triangle Splatting.

Furthermore, we would like to clarify that these improvements are not due to random fluctuations. We have also provided our code in the supplementary materials to ensure reproducibility.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.