acceptodds
Under review as a conference paper at ICLR 2027

Prune by Proxy: A Unified Framework for 3D Gaussian Splatting Pruning

Abstract

Post-hoc pruning makes 3D Gaussian Splatting (3DGS) compact by removing primitives ranked by an importance score. Existing scores rank Gaussians by their attributes, rendering contributions, or parameter sensitivity, yet what pruning actually trades off is the photometric cost of deleting a Gaussian. Evaluating this deletion cost by leave-one-out rendering requires one re-rendering per Gaussian, and no common target or protocol exists to tell which of the tractable scores actually estimate it. We present Prune by Proxy, a framework that fixes the deletion cost as the common target, specifies every score by its scoring information, scored attributes, and aggregation rule, and thereby transfers the zero-cost proxies of NAS-Bench-Suite-Zero from network pruning and architecture search to 3DGS. Within this framework we show that a multiplicative gate on activated opacity is the coordinate of deletion: alpha blending is affine in a single gate, so the deletion-induced image change equals minus the gate derivative, and the gap between a first-order estimate and the exact deletion cost lies entirely in the photometric loss, where it has a closed form. The resulting Corrected and scores equal the exact single-Gaussian deletion cost in the ideal blending model and cost one forward and one backward pass per view. At 90% pruning on Mip-NeRF 360, Tanks & Temples, and Deep Blending, Corrected ranks first in eight of nine dataset–metric comparisons and exceeds PUP on all thirteen scenes by 0.58 dB on average at one-third of its scoring time. The closed-form correction alone adds up to 1.6 dB over its first-order control, and seven transferred proxies also surpass all published 3DGS scores, with deletion alignment explaining which proxies transfer.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, pruning, zero-cost proxies, neural architecture search, 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_KMSDOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The paper studies post-hoc pruning for 3D Gaussian Splatting (3DGS). It identifies the fundamental objective of post-hoc pruning as minimizing the photometric reconstruction error increase resulting from primitive removal, termed the single-Gaussian deletion cost. Because computing this target via brute-force leave-one-out re-rendering across millions of primitives is computationally intractable, existing heuristic and sensitivity-based methods are framed as zero-cost proxy estimators. The paper introduces "Prune by Proxy", a unifying framework that maps proxies from neural architecture search (NAS-Bench-Suite-Zero) into 3DGS via slot substitution (ranked unit: Gaussian, query: camera view, forward model: alpha blending). Under an idealized front-to-back alpha blending formulation, the authors show that image formation is strictly affine in a multiplicative opacity gate attached to each Gaussian, meaning the deletion-induced image change equals minus the gate derivative. Consequently, the non-linear gap between first-order Taylor estimates and the true finite deletion cost is confined entirely within the photometric loss function (L1/L2L_1 / L_2), where closed-form remainder corrections are derived. The resulting Corrected L1L_1 and L2L_2 scores require only a single forward and backward pass per view. Evaluated on Mip-NeRF 360, Tanks & Temples, and Deep Blending under a two-round pruning protocol (90% and 95% total sparsity) with refinement, Corrected L1L_1 outperforms baselines such as PUP and LightGaussian in reconstruction fidelity while maintaining faster scoring times than second-order criteria.

Critical strengths

S1: Principled theoretical formulation of opacity-gate deletion. The mathematical observation that front-to-back alpha blending is elementwise affine with respect to an individual opacity gate (C(mi)=C(0)+miBiC(m_i) = C(0) + m_i B_i) is insightful and clean. Establishing the identity Dv,i=−Bv,i=−∂Rv/∂miD_{v,i} = -B_{v,i} = -\partial R_v / \partial m_i effectively isolates the rendering non-linearity from the loss non-linearity. Deriving analytical closed-form remainder corrections for L1L_1 and L2L_2 photometric objectives provides a rigorous, principled bridge between infinitesimal gradient perturbations and finite single-primitive deletion costs, avoiding heuristic parameter scalings.

S2: Unification of disparate scoring heuristics under a shared target. By defining the exact single-Gaussian deletion cost as a reference target, the paper provides a taxonomy covering attribute-based, contribution-based, and sensitivity-based pruning methods. Mapping thirteen zero-cost proxies from NAS literature to Gaussian scenes offers a systematic benchmarking testbed and clarifies which proxy configurations correlate with deletion impact.

S3: Computational efficiency relative to second-order sensitivity baselines. By bypassing explicit Hessian-vector products or log-determinant Jacobian Gram evaluations required by curvature-based methods like PUP, the proposed Corrected L1L_1 score computes finite deletion approximations in a single forward-backward pass per training camera view. This achieves roughly a 3x speedup in scoring wall-clock time over PUP while achieving consistently higher PSNR across multiple benchmarks after fine-tuning.

Critical weaknesses

W1: Confounding effect of heavy post-refinement and complete absence of zero-shot pruning evaluation. The central premise of the work is that Corrected L1L_1 computes a quantitatively superior estimation of the ground-truth deletion cost. However, the evaluation relies on a multi-stage pruning pipeline with 10,000 steps of post-pruning Adam refinement (5k steps after Round 1 at 80% sparsity, and 5k steps after Round 2 at 50% sparsity). Extensive gradient descent enables survivor Gaussians to adapt spatial positions, opacities, and spherical harmonics to compensate for deleted primitives. In Table 8, completely random pruning at 90% sparsity recovers to 24.28 dB on Bicycle (only 0.91 dB below unpruned 3DGS, and within 0.91 dB of Corrected L1L_1). Because no unrefined (zero-shot) reconstruction metrics are reported directly after pruning, it remains unverified whether the final 0.58 dB average margin over PUP originates from accurate deletion cost scoring or from incidental structural suitability for Adam optimization. Without zero-shot pruning curves (PSNR/SSIM degradation versus sparsity before refinement), the core claim regarding deletion cost accuracy lacks empirical closure.

W2: Disconnect between idealized affine blending assumptions and production CUDA rasterizers. Proposition 4.1 explicitly hinges on the condition that rendering involves "no opacity clipping, contribution thresholds, or early termination." However, standard high-performance 3DGS implementations (e.g., Inria CUDA rasterizer, gsplat) rely strictly on discrete algorithmic cutoffs, notably transmittance saturation (T<10−4T < 10^{-4} early stopping), pixel alpha contribution thresholds (α⋅footprint<1/255\alpha \cdot \text{footprint} < 1/255), and tile-based bounding-box binning. These discrete non-linearities break the strict affine relationship along the continuous deletion path mi∈[0,1]m_i \in [0, 1]. The paper provides no numerical correlation analysis (e.g., Spearman rank correlation or mean absolute relative error) comparing the proposed closed-form score against brute-force leave-one-out re-rendering on a sampled subset of Gaussians under realistic rasterization engines. Consequently, the degree to which Proposition 4.1 holds in actual software execution remains unquantified.

W3: Attribution ambiguity: performance gains appear dominated by coordinate selection rather than loss correction. The ablation results in Table 7 demonstrate that replacing spatial parameters with opacity gates in the baseline PUP method (PUP-α\alpha) causes PSNR to jump from 26.15 dB to 26.83 dB on Mip-NeRF 360 (+0.68 dB). Similarly, SynFlow-GS, which operates on opacity gates without accessing ground-truth reference images (Jv=1⊤RvJ_v = \mathbf{1}^\top R_v), achieves 26.78 dB. Meanwhile, adding the closed-form loss correction on top of first-order SNIP-GS (26.73 dB to 27.06 dB) provides a comparatively modest gain of ~0.33 dB. These numbers suggest that the vast majority of empirical improvement arises simply from confining derivative evaluations to opacity coordinates rather than the theoretical proxy framework or the analytical loss remainder. The narrative appears to over-package coordinate selection as a broader architectural and loss-correction breakthrough.

W4: Inadequacy of single-Gaussian deletion theory for massive joint pruning. Pruning removes 80% to 90% of the scene primitives simultaneously. While removing an isolated Gaussian is affine in its individual gate, joint deletion introduces multi-linear coupling terms across overlapping primitives (∏j<k(1−mjaj)\prod_{j < k} (1 - m_j a_j)). In high-density clusters, multiple semi-transparent primitives often contribute redundantly to the same ray; each primitive evaluated in isolation exhibits a negligible individual deletion cost BiB_i, leading a single-Gaussian score to mark all of them for concurrent deletion. Once removed jointly, severe background transmittance leakage and catastrophic visual tearing occur. The paper provides no theoretical bounding or empirical analysis regarding how single-Gaussian proxies degrade under varying batch pruning ratios (e.g., one-shot 90% pruning vs. iterative schedules).

W5: Incomplete computational profiling regarding peak VRAM and operational overhead. Table 10 reports scoring wall-clock time, but omits peak GPU memory usage during scoring. Evaluating backpropagation and analytical residuals across high-resolution training views requires caching multi-view rendering buffers and gradient tensors. In complex outdoor scenes (e.g., Mip-NeRF 360), this overhead could induce out-of-memory errors on commodity GPUs (16GB-24GB VRAM), whereas feed-forward heuristic methods like LightGaussian require no backward pass graphs. Additionally, the paper does not report final inference rendering FPS or uncompressed storage footprints to demonstrate downstream practical utility.

Questions for authors

Q1: Can you provide zero-shot pruning results without any post-refinement (0 refinement steps) across 50%, 80%, and 90% pruning ratios on Mip-NeRF 360 and Tanks & Temples? Evaluating reconstruction fidelity immediately after pruning is essential to verify whether Corrected L1L_1 preserves scene appearance more accurately than PUP, LightGaussian, and SNIP-GS before gradient optimization masks scoring errors.

Q2: What is the empirical correlation between Corrected L1L_1 and ground-truth numerical deletion costs under a standard 3DGS rasterizer? On a representative scene, if 10410^4 Gaussians are randomly sampled and evaluated via exact leave-one-out re-rendering using the standard CUDA pipeline (with transmittance early stopping T<10−4T < 10^{-4} and contribution threshold α<1/255\alpha < 1/255), what are the Spearman rank correlation and relative error metrics for Corrected L1L_1, SNIP-GS, and PUP against the true deletion cost?

Q3: How do you decouple the empirical contribution of the opacity coordinate from the theoretical loss correction? Given that PUP-α\alpha (+0.68 dB) and unreferenced SynFlow-GS (+0.63 dB) achieve the vast majority of performance gains simply by evaluating derivatives with respect to opacity, can you provide a formal variance decomposition quantifying the isolated marginal benefits of (i) opacity gating, (ii) ground-truth image supervision, and (iii) the analytical loss remainder?

Q4: What is the peak GPU memory (VRAM) consumption during the forward-backward scoring pass across the benchmark datasets, and what are the actual post-pruning inference rendering framerates (FPS) achieved on standard hardware?

Reviewer_ekf9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies post-hoc pruning of trained 3D Gaussian scenes. It treats the photometric cost of deleting one Gaussian, with the remaining scene fixed, as a common target for existing and transferred scores. An activated-opacity gate makes the ideal rendered image affine along a single deletion path, yielding closed-form L1 and squared-L2 loss corrections. The method performs two rounds of scoring, pruning, and fine-tuning. On thirteen scenes, Corrected L1 improves over PUP by 0.58 dB on average under the common pipeline. Scoring time falls from 25.22 to 8.58 seconds.

Critical strengths
  1. The local analysis is concrete and checkable. Equations (4)–(9) identify a deletion-aligned coordinate and separate renderer affinity from the finite-step loss remainder. Table 4 holds the gate, loss and normalization fixed while removing the correction, providing a useful control for the construction of these individual scores.

  2. The experiments give credible evidence within the class of static scoring rules. Methods share starting checkpoints and pruning/refinement schedules, and the corrected score improves over PUP.

Critical weaknesses
  1. My main concern is that the single-Gaussian deletion cost in Equation (1) is not sufficiently justified as a target for compaction or compression. A leave-one-out score measures a Gaussian’s marginal contribution with all other Gaussians present. Two overlapping, mutually redundant Gaussians can each have a small individual deletion cost, while their simultaneous removal causes a large error. Thus, even exact singleton scores do not ensure that ranking them yields a good retained subset. Appendix B and the conclusion acknowledge multilinearity and higher-order interactions. However, the paper does not provide an interaction-aware subset-selection procedure or empirical evidence connecting singleton scores to the immediate loss caused by large-batch deletion.

  2. The evaluation also lacks comparisons with set-level alternatives such as GaussianSpa (Zhang et al., 2025) and GHAP (Wang et al., 2025). GaussianSpa optimizes reconstruction under an explicit sparsity constraint, whereas GHAP formulates compaction as global Gaussian mixture reduction. Comparisons at matched primitive budgets and comparable total computational costs would help establish whether and when the proposed approach is preferable. Restricting Table 3 to ranking rules leaves the broader justification of singleton deletion cost as a common pruning target unresolved.

References:

Zhang, Y., Jia, W., Niu, W., and Yin, M. (2025). GaussianSpa: An “Optimizing-Sparsifying” Simplification Framework for Compact and High-Quality 3D Gaussian Splatting. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 26673–26682.

Wang, T., Li, M., Zeng, G., Meng, C., and Zhang, Q. (2025). Gaussian Herding across Pens: An Optimal Transport Perspective on Global Gaussian Reduction for 3DGS. Advances in Neural Information Processing Systems (NeurIPS).

Questions for authors
  1. How accurately does the sum of individual deletion costs predict the actual loss increase when the selected Gaussians are deleted together?

  2. Does recomputing scores after smaller pruning steps improve rendering quality under comparable computational budgets? How much of the quality loss immediately after pruning is recovered by fine-tuning?

  3. How does the proposed method compare with joint compaction approaches such as GaussianSpa and GHAP? Please match retained counts, use shared initial checkpoints where applicable, and report total optimization costs.

I would consider raising my rating if these concerns are adequately addressed.

Reviewer_9zRKOpenReview ↗
Rating
2
Confidence
3(−1)
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies Gaussian importance from the perspective of reconstruction loss incurred by removing individual primitives. A central observation is that, under ideal alpha blending, the rendered image is affine with respect to an opacity gate, allowing the effect of deleting a Gaussian to be recovered from its image derivative. This leads to corrected L1 and L2 scores that account for the discrepancy between first-order sensitivity and finite deletion cost. The authors also revisit a range of existing 3DGS pruning criteria and adapt zero-cost proxies from network pruning and NAS. Experiments on 13 scenes show consistent improvements over PUP and competitive performance with relatively low scoring overhead.

Critical strengths

The opacity-gate analysis is the most compelling part of the paper. The distinction between parameter sensitivity and the actual effect of removing a Gaussian is well motivated, and the derivation in Section 4 gives a clear explanation of why first-order gradients can be misleading, particularly when the reconstruction residual is small. The resulting loss correction is simple and mathematically well justified under the stated assumptions.

I also appreciate the effort to evaluate different importance criteria within a common pruning pipeline. This makes the comparison more meaningful than combining independently reported compression results. In particular, the ablations in Tables 4 and 5 are useful: they separately examine the effect of correcting the first-order approximation and choosing a coordinate that directly represents deletion. The performance improvements across the 13 scenes are fairly consistent, and the scoring overhead is modest compared with PUP.

Critical weaknesses

1. Limited Novelty. The theoretical contribution appears somewhat overstated. In particular, the proposed loss-corrected scores can be algebraically reduced to the direct difference between the photometric losses before and after deleting a Gaussian. The L1 correction follows from the absolute-value loss, while the L2 correction is simply the exact quadratic expansion of squared error. Therefore, the correction itself does not seem to introduce a fundamentally new mathematical principle. The potentially nontrivial contribution lies instead in efficiently obtaining the per-Gaussian image changes required to evaluate these loss differences. I would encourage the authors to clarify what is algorithmically new in this computation, beyond exploiting the standard alpha-compositing equation and reorganizing the loss difference into first-order and remainder terms.


2. Limited empirical validation of the exact deletion-cost claim.

The theoretical guarantee relies on idealized alpha blending, whereas the actual rasterizer involves opacity clipping, contribution thresholding, and early termination. Although these limitations are acknowledged in Appendix B, their practical impact is not quantified. In particular, the paper does not compare the proposed scores with actual single-Gaussian deletion costs obtained through leave-one-out rendering. Such an evaluation would be important to establish how accurately the theoretical formulation reflects the behavior of the deployed renderer.


3. The relevance of individual deletion costs to aggressive joint pruning remains unclear.

Even if the deletion cost of each Gaussian could be computed exactly, ranking Gaussians independently does not account for interactions among simultaneously removed primitives. This is particularly relevant to the proposed pipeline, which removes 80% of the Gaussians in the first pruning round. Overlapping or redundant Gaussians may have small individual deletion costs but cause substantial reconstruction degradation when removed together. While Appendix B briefly discusses higher-order interactions, their practical significance is not investigated. An evaluation under more gradual pruning schedules would help clarify how well the proposed criterion extends to large-scale joint pruning.


4. The practical gain from exact loss correction over simpler gate-based criteria is relatively modest. The comparison with the signed first-order estimate is convincing: adding the correction substantially improves performance. However, the gap becomes much smaller when comparing against stronger proxies that already use the opacity gate. In Table 3, Corrected L1 exceeds SynFlow-GS by only 0.12 dB on Mip-NeRF 360, 0.04 dB.

Questions for authors

Please refer to the Weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.