acceptodds
Under review as a conference paper at ICLR 2027

Gradient-Driven Natural Selection for Compact 3D Gaussian Splatting

Abstract

Compact 3D Gaussian Splatting must reduce the number of primitives while preserving reconstruction quality. In a dense representation, overlapping Gaussians jointly support local details, so a member's current contribution may not reflect its value after other members leave and the scene is reoptimized. We introduce Gradient-Driven Natural Selection (GNS), which organizes competition through shrinkage and reconstruction-driven adaptation of native opacity, allowing necessary local support to concentrate on fewer members before pruning. GNS applies a shared linear penalty to pre-activation opacity parameters to reduce selection bias introduced by the shrinkage term itself. It also increases the opacity learning rate during selection so that members respond to reconstruction feedback as the representation contracts. The surviving subset emerges during reconstruction optimization without extra learnable selection variables or inference modules. With common dense inputs, training iterations, and target Gaussian budgets, these two simple changes achieve better overall reconstruction quality than the compared learned pruning baselines across multiple datasets.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Gaussian Pruning, Model Compression, Novel View Synthesis, Differentiable Rendering

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

The paper proposes a new pruning method for 3DGS called Gradient-Driven Natural Selection (GNS). It compresses a scene by encouraging redundant Gaussians to compete and adapt before removal. By applying a shared shrinkage force to opacity logits rather than opacity values. GNS aims to reduce regularisation bias and achieve better quality–size trade-offs than several existing pruning methods.

Critical strengths

The pruning idea is quite interesting, and the results look promising, although the method is quite close to MaskGS.

Critical weaknesses
  • Section 3.2, the derivation seems to correctly shows that the proposed regulariser contributes an identical additive gradient to all opacity logits. However, the paper should clarify whether this directly implies equal opacity shrinkage or not, and also it is unbiased selection dynamics after the sigmoid activation and optimization process. The current interpretation appears stronger than what is formally proven.
  • Section 3.3 shows that a higher learning rate produces larger updates to the internal opacity parameters, but larger updates do not necessarily mean faster adaptation. The sigmoid function means that actual opacity changes may not increase proportionally. Larger steps could help, but could also overshoot or cause oscillation. Equations (6) and (7) alone do not show which happens. Please explain what "faster adaptation" means and discuss these possible effects.
  • The ablations show better average final image quality, and the two-scene diagnostic suggests that Gaussians redistribute their contributions. However, these results do not directly show that useful adaptation happens sooner across scenes. A stronger test would measure how many training steps are needed to recover a specified image quality after the same pruning events, using a wider range of scenes.
  • Table 1 fixes the number of Gaussians, But Gaussian count is not a complete measure of model size or efficiency, suggesting also report storage, rendering speed and training memory.
  • The comparison would benefit from additional relevant baselines or clearer justification for their exclusion. Where feasible, comparisons under matched Gaussian budgets would help establish the contribution beyond existing methods, particularly recent approaches such as ConeGS and Prune Wisely. If direct comparisons are unsuitable due to differences in primitive addition or modification, please clarify these differences and consider controlled component-level comparisons.
Questions for authors

Please see weakneses, particularly for Section 3.3 and its ablation study.

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

This paper studies post-training pruning of 3D Gaussian Splatting (3DGS) under a fixed Gaussian budget and proposes Gradient-Driven Natural Selection (GNS). Without introducing additional learnable masks, the method directly applies a linear shrinkage regularization term to the pre-activation parameters of the native opacity, increases the opacity learning rate during the pruning stage, and dynamically adjusts the regularization strength according to the opacity quantile corresponding to the target retention number. The authors interpret the above design as “competitive neutrality” and “competitive sufficiency,” hoping that reconstruction gradients will drive locally overlapping Gaussians to gradually differentiate their contributions, followed by discrete deletion.

Critical strengths
  1. Simple implementation. The method directly acts on the native opacity parameters of 3DGS and does not require additional learnable selection variables, making it easy to integrate with existing training pipelines in engineering practice.
  2. Relatively carefully controlled internal evaluation. Comparisons with the same dense starting point, the same final iteration number, and the same retention budget, as well as ID replay and module ablations, help reduce some experimental confounding factors.
Critical weaknesses
  1. Insufficient novelty and independent insight of the method.

1.1 The core shrinkage mechanism lacks substantial distinction from existing work. Eq. (4) in this paper directly applies a linear penalty to the opacity pre-activation parameters, and Eq. (5) states that it adds an identical first-order gradient to each parameter. However, ConeGS (3DV 2026) has already adopted a pre-activation opacity penalty, and its public implementation includes an operation that takes the mean of the original opacity parameters and adds it to the reconstruction loss, which directly corresponds to the core regularizer emphasized in this paper. The overall application stages of the two works differ (ConeGS targets continuous densification, whereas GNS targets a fixed candidate set after densification), so it cannot be said that the complete algorithms are identical. However, this paper should still limit the truly new contribution to the learning rate adjustment and budget feedback in the pruning stage, rather than treating the shared shrinkage itself as a new selection principle.

1.2 The relationship with existing “reconstruction-feedback-driven selection” is also unclear. ControlGS has used opacity sparsity regularization to produce continuous shrinkage, and MaskGaussian dynamically evaluates the survival probability of Gaussians through gradient updates. Although these works differ in parameterization and algorithm implementation, they all embody the idea of “gradually deciding Gaussian retention during reconstruction optimization.” This paper only describes the differences in related work and lacks strict same-condition analysis against the same-origin regularizer of ConeGS, the ControlGS-style shrinkage mechanism, and dynamic selection strategies. The “Pre only” ablation in the paper cannot replace a direct comparison with prior mechanisms and contribution boundaries.

1.3 The “competition” concept has not yet formed sufficiently distinctive mechanistic insight. Eq. (5) is mainly a direct consequence of adding the gradients of linear terms, and Eq. (7) mainly shows that increasing the learning rate amplifies relative parameter updates; this is not enough to explain why a new competition dynamics beyond general sparsification and accelerated adaptation would arise. Although Section 4.5 shows statistics of opposite changes in the contributions of adjacent Gaussians, this experiment only involves two scenes, bicycle and counter. In the appendix, the two branches also separately adjust the regularization strength and align the number of retained Gaussians through additional opacity-sorted deletion, making it difficult to rule out alternative explanations based on update magnitude, budget trajectory, or selection path.

  1. Experimental gains are limited, and robustness and efficiency evidence is insufficient.

2.1 The improvements over strong baselines are small and inconsistent. Table 2 shows that on ordinary 3DGS input, GNS improves PSNR over MaskGaussian only from 27.770 dB to 27.784 dB (+0.014 dB), and SSIM from 0.8193 to 0.8207; on MiniGS-D input, the PSNR of GNS (27.766 dB) is even slightly lower than that of Naive L1 (27.774 dB). In Table 1, GNS achieves a PSNR of 30.248 dB on Deep Blending, which is lower than the 30.343 dB of MaskGaussian. These results indicate that the method does not demonstrate a stable and significant overall leading advantage.

2.2 The marginal contribution of the innovative modules still needs further justification. In Table 3, under automatic feedback, from Pre only to Full GNS, the average PSNR on Mip-NeRF 360 changes from 27.901 dB to 27.946 dB (only +0.045 dB). Since Pre only already includes the main mechanism related to existing pre-activation shrinkage, the gain brought by the added learning rate adjustment is not sufficient to independently support a strong novelty claim. The SSIM and LPIPS in this table do improve, but no multi-random-seed means, standard deviations, or significance tests are provided, making it difficult to judge the stability of small differences.

2.3 There are gaps in evaluation dimensions and mechanism validation. The paper covers multiple datasets and dense inputs; the problem is not simply a “small number of scenes,” but rather the lack of training/pruning time, rendering throughput, peak memory, and end-to-end efficiency comparisons directly corresponding to the “compact and efficient” claim. After different methods are ported to a unified code framework, clearer explanations of implementation fairness and tuning cost are also needed. On the other hand, the curves at different retention rates are mainly limited to Mip-NeRF 360, and the mechanism diagnosis involves only two scenes, which is insufficient to prove that the described gradient competition phenomenon can generalize stably.

Questions for authors
  1. Novelty boundary: What is the specific relationship between the pre-activation opacity mean penalty in the public implementation of ConeGS and Eq. (4) in this paper? If the same regularization form is applied to the same post-densification starting point, the same pruning duration, and the same budget, how much stable gain can the added learning rate adjustment and quantile feedback each bring?

  2. Result robustness and practical value: Can the authors provide means and variances from at least multiple independent runs under the key 20%/10% retention budgets, and report pruning time, total training time, peak memory, and final rendering FPS on uniform hardware? Are these gains significantly beyond random fluctuations and additional tuning costs?

  3. Falsifiable evidence for the competition mechanism: Can the authors control the same regularization strength trajectory, retention-number trajectory, and budget on more scenes, and on this basis directly compare the reconstruction gradients, contribution transfer, and final local reconstruction errors of overlapping Gaussians, thereby ruling out the explanation of mere learning rate amplification?

1 comment

Thank you for asking us to clarify what GNS adds beyond the shared opacity penalty and to strengthen the repeatability, efficiency, and mechanism evidence.

Components and prior mechanisms. The revision explicitly credits ConeGS for using the same pre-activation mean penalty. In the fixed post-densification candidate set, GNS adds a selection-phase opacity learning-rate change and target-count feedback. With Pre and the same initial Auto coefficient, the higher rate adds 0.045 dB PSNR and 0.00335 SSIM and lowers LPIPS by 0.00678 across nine scenes. With common-source ID replay, the gains are 0.041 dB, 0.00276 SSIM, and a 0.00571 LPIPS reduction (Tables 3 and 16). At the higher rate, Auto lowers LPIPS by 0.00191 relative to per-scene calibrated Fixed. The Pre arms share the 5K selection limit and 30K endpoint. Appendix D.1 reports when each arm reaches the budget, which varies under Auto and Fixed. Section 2 distinguishes ControlGS's activated-opacity sum and continued densification from the fixed-candidate Naive L1 comparison. Appendix B describes the common inputs and endpoints, the method-specific schedules, and the implementation checks for the transferred baselines.

Repeatability and scope. At 20% retention with ImprovedGS inputs, Table 10 compares four methods across five outdoor scenes and three seeds. For GNS minus MaskGaussian, the paired differences in the five-scene means are +0.02785 ± 0.00323 dB PSNR, +0.004052 ± 0.000042 SSIM, and −0.006501 ± 0.000026 LPIPS (mean ± sample SD across seeds). With original 3DGS inputs, GNS lowers LPIPS in all 15 pairs across Counter, Bicycle, and Stump, while the mean PSNR differences are negative in each scene. At 10% retention on Bicycle, GNS improves on MaskGaussian in all three metrics in all five pairs, with a PSNR gain of 0.07472 ± 0.00695 dB (Tables 11–12). These repeats compare complete pruning methods. Table 3 reports the component differences under the stated configuration. Appendix Figure 5 extends the retention curves to Tanks and Temples and Deep Blending.

Efficiency and tuning. Table 6 compares four methods on one RTX 4090 with the same dense inputs and 20% budgets. GNS averages 45.74 s for selection and 108.47 s for the 15K–30K training process, or 114.56 s through independent evaluation. The corresponding training-process times are 109.19 s for Naive L1, 129.04 s for GaussianSpa, and 215.95 s for MaskGaussian. Including the common dense stage gives an estimated 278 s in total for GNS. The final models average 113.53 MiB for each method. GNS records the highest mean FPS (1354.8) and highest peak training memory (5.769 GiB allocated, 7.563 GiB process). Appendix D.1 documents 4,766 s of calibration for the three additional Fixed variants. Section 4.1 discloses the use of feedback from the reported test views during method development and baseline tuning.

Mechanism control. At matched counts, the increase in opposite compositing-weight changes is larger for close pairs than for mismatched pairs across nine scenes at the 0.1% threshold (Table 4). Appendix Table 15 reports the threshold dependence. On Bicycle, Table 5 and Appendix D.4 also hold the realized coefficient and count trajectories fixed. Under the LR5 reference trajectory, the replayed LR5 subset gains 0.134 dB over LR1 and yields lower local L1 error in 14 of 16 regions. Quota-based deletions account for 99.97% of deletions at LR1 and 0.40% at LR5. The association between contribution changes and local error varies across stages. The jointly matched control and local-gradient measurements are specific to Bicycle.

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

The paper prunes a fixed Gaussian candidate set using a linear penalty on pre-activation opacity, a higher opacity learning rate, and budget feedback. Experiments match dense initialization, training endpoints, and final counts. ID replay examines selected-subset quality, while multiple datasets and initialization methods test the quality–count trade-off.

Critical strengths
  1. Common initial models and final counts reduce important confounders. ID replay further separates subset selection from parameter optimization during pruning.

  2. The method adds no inference module or selection variables, and is evaluated with several dense initializations and a 2DGS setting. This transferability gives the simple intervention practical value.

Critical weaknesses
  1. The advance over the existing penalty needs sharper definition. Section 2 acknowledges that ConeGS implements the same parameter-mean penalty. In Table 3, Pre only + Auto achieves 27.901 dB versus 27.946 dB for the full method, with additional SSIM/LPIPS gains. The contribution should be grounded in these incremental improvements.

  2. Some close comparisons lack repeatability evidence. Table 2 reports a 0.014 dB PSNR advantage over MaskGaussian. The perceptual gains matter, but without repeat-run information the stability of the ranking remains uncertain; this does not establish that the gains are invalid.

Questions for authors
  1. Since the paper identifies the pre-activation penalty as shared with ConeGS, using Table 3, which metrics, budgets, or scenes benefit from the learning-rate change and budget feedback beyond the shared penalty? A comparison with the entire ConeGS pipeline is unnecessary.

  2. Can you provide repeatability information for a limited representative set of the closest comparisons, reporting all three metrics? Please also clarify whether hyperparameters were chosen on held-out validation data or on the reported test results.

  3. Can the authors distinguish more explicitly what is proved by the pre-optimizer gradient identity from what is inferred from the two-scene diagnostic? Does the existing data connect the responsibility-change statistic to the quality of the selected subset, or should the causal interpretation remain tentative?

P.S. If authors could answer my questions, I'm willing to increase my score.

3 comments

Thank you for asking us to quantify the gains beyond the shared penalty and clarify the evidence for the mechanism. The revision explicitly credits the use of the same pre-activation mean penalty in ConeGS and reports the added effects of learning rate and budget feedback separately.

Component gains. With Pre and the same initial Auto coefficient, raising the opacity learning rate improves mean PSNR by 0.045 dB and SSIM by 0.00335, and reduces LPIPS by 0.00678 (Table 3). With common-source ID replay, the changes are +0.041 dB, +0.00276 SSIM, and −0.00571 LPIPS. SSIM and LPIPS improve in all nine scenes under both protocols. Auto gives quality close to per-scene calibrated Fixed with Pre at the base rate, and lowers LPIPS by 0.00191 with the higher rate. Appendix D.1 and Table 16 give the calibration protocol and per-scene results.

Repeatability and tuning. At 20% retention, Tables 11–12 report all three metrics for five paired seeds on Counter, Bicycle, and Stump with original 3DGS inputs. GNS lowers LPIPS in all 15 pairs against MaskGaussian. The mean PSNR differences are negative in all three scenes, which qualifies the small positive PSNR difference in the nine-scene comparison in Table 2. Section 4.1 and Appendix D.1 state that method development and baseline tuning used feedback from the reported test views. The additional Fixed calibration used the count trajectory as its acceptance criterion.

Gradient identity and mechanism. Equation (5) identifies the equal additive logit gradient before Adam. Actual opacity updates also depend on the sigmoid and optimizer state. The nine-scene diagnostic matches counts, with each arm adjusting its own regularization coefficient. On Bicycle, where both coefficient and count trajectories are matched, common-source replay gives the higher-rate subset a 0.134 dB PSNR gain and lower local L1 error in 14 of 16 regions (Table 5 and Appendix D.4). We report the association between contribution changes and local error by stage and limit the interpretation to these controls.

Thank you for the detailed rebuttal and additional experiments. The responses address several of my concerns, particularly regarding repeatability and mechanism interpretation. However, I still have reservations about the incremental novelty and test-set-informed tuning. And I consider the paper borderline and will maintain my rating.

Thank you for taking the time to consider our rebuttal and the additional experiments. We appreciated your initial openness to raising the rating if we could address your questions. We are glad that the added evidence helped address your concerns about repeatability and the interpretation of the mechanism.

We hope the revision also clarifies what GNS adds beyond the shared ConeGS penalty. At the same retention budget, increasing the opacity learning rate improves SSIM and LPIPS in all nine scenes. Both gains persist when the selected subsets are replayed from a common checkpoint (Tables 3 and 16). Budget feedback reduces the need to calibrate a fixed coefficient for each scene. In a mature area with strong baselines, even modest gains at matched budgets are difficult to achieve. We see value in showing which simple changes to native opacity improve the choice of retained Gaussians and make it easier to reach a specified budget.

We take your concern about test-set-informed tuning seriously. Section 4.1 and Appendix D.1 disclose that feedback from the reported test views informed method development and baseline tuning. This limits how strongly we can generalize small margins to independent evaluation. We interpret the ablations, replay, and repeatability results within the reported protocol.

We respect your assessment and appreciate the time you have put into this discussion. If these points change how you weigh the contribution as the discussion continues, we would be grateful if you reconsidered your assessment.

Comments on the paper

1 comment

We thank the reviewers and Area Chair for the discussion. The GNS revision keeps the original fixed-candidate pruning method and adds experiments and clarifications on component gains, mechanism, repeatability, and cost. The main changes are marked in blue.

  1. Contribution beyond the shared penalty (pp. 1, 3–5, and 8). The revision explicitly credits ConeGS's use of the pre-activation mean penalty. Table 3 separates the effects of the higher selection-phase opacity learning rate and the target-count feedback. Under Pre + Auto, the learning-rate change adds 0.045 dB PSNR and 0.00335 SSIM and lowers LPIPS by 0.00678 across nine scenes. Similar improvements persist after common-source ID replay. Appendix D.1 documents the role of Auto in reducing per-scene calibration and the tuning protocol.
  2. Mechanism and recovery (pp. 9–10 and Appendix D.4). Table 4 extends the pre-deletion contribution diagnostic to nine scenes at matched counts. Table 5 adds a Bicycle control that also matches coefficient trajectories. Its common-source replay favors the higher-rate subset by 0.134 dB PSNR. The separate shared-deletion test finds that the higher rate leads early on Bicycle, while the base rate sustains recovery of 90% of the initial PSNR loss at 150 updates, compared with 750 for the higher rate. Table 14 reports the non-monotonic longer-window results on Bicycle. The revision distinguishes evidence on selected-subset quality from evidence on recovery speed.
  3. Repeatability and cost (p. 10 and Appendix D.2). Tables 10–12 give repeated three-metric comparisons at 20% retention and five paired runs on Bicycle at 10%. At 20% with original 3DGS inputs, the three representative scenes have lower mean PSNR than MaskGaussian and lower LPIPS in all 15 paired runs. Table 6 reports same-GPU pruning time, estimated end-to-end training time, memory, rendering speed, and file size. GNS is close to Naive L1 in training time, with the highest mean rendering FPS and peak training memory among the four methods.
  4. Additional comparisons (p. 6 and Appendix D.3). We added Prune Wisely's schedule and ranking components under matched ordinary-Gaussian budgets and retention curves on two additional datasets. Section 4.1 and Appendix D.1 describe the use of test-view feedback for tuning and the calibration protocol.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.