acceptodds
Under review as a conference paper at ICLR 2027

SplatQ: Scalable Interaction-Aware Pruning of 3D Gaussian Splatting via Quadratic Selection

Abstract

Existing 3D Gaussian Splatting (3DGS) pruning methods primarily rely on global Top-K selection based on independent importance scores, lacking regional budget control and ignoring joint deletion effects. Consequently, they cannot ensure balanced reconstruction quality across a scene. We propose SplatQ, a combinatorial pruning algorithm that coordinates regional budget allocation and joint Gaussian selection under a fixed budget. SplatQ partitions Gaussians by spatial proximity and appearance similarity into size-limited blocks, decomposing global selection into subproblems of a specific size. It then smooths importance-based block retention quotas to prevent excessive budget concentration. Finally, it constructs a quadratic approximation of joint deletion effects using the responses of representative pixel colors to changes in Gaussian opacity. Individual responses, pairwise interactions, and quota penalties are incorporated into a quadratic unconstrained binary optimization (QUBO) objective, solved using simulated annealing (SA) or a coherent Ising machine (CIM) to jointly select Gaussians. Experiments on 5 benchmarks demonstrate that SplatQ outperforms state-of-the-art pruning methods in aggregate reconstruction quality, spatial quality balance, and rendering speed. Ablations confirm that budget smoothing improves local detail reconstruction, while interaction modeling improves Gaussian selection quality under a fixed budget.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Post-training Pruning, Quadratic Unconstrained Binary Optimization (QUBO)

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_uUGzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses post-training compression of 3D Gaussian Splatting (3DGS) under extreme pruning budgets (e.g., 90% primitive removal). The authors argue that conventional global Top-K ranking based on independent importance scores leads to spatial quality imbalances (over-allocating primitives to complex regions while degrading others) and ignores joint deletion effects (overlapping volumetric rendering contributions among primitives). To resolve this, the authors introduce SplatQ, a combinatorial pruning framework consisting of three main stages: (1) similarity-guided spatial-appearance clustering that partitions Gaussians into size-bounded sub-blocks; (2) budget smoothing across blocks to enforce regional retention balance; and (3) opacity response modeling that captures first-order rendering responses and pairwise interaction terms to formulate a Quadratic Unconstrained Binary Optimization (QUBO) problem within each block. The resulting QUBO is solved via Simulated Annealing (SA) or evaluated on an optical Coherent Ising Machine (CIM) backend, followed by post-pruning fine-tuning. Experiments across 27 scenes on 5 benchmarks report superior PSNR, spatial quality consistency, and rendering throughput compared to existing pruning baselines.

Critical strengths

S1: Formulating 3DGS pruning as a combinatorial optimization problem rather than independent greedy ranking is theoretically principled. The paper correctly observes that alpha compositing in volumetric rendering induces strong inter-primitive dependencies, where the photometric impact of removing a cluster of overlapping Gaussians cannot be decomposed into a linear sum of individual deletion penalties. Approximating joint deletion effects via pairwise opacity response interactions and mapping the selection into a QUBO formulation represents an elegant mathematical departure from traditional heuristic scoring.

S2: The empirical benchmark evaluation is extensive in scope, evaluating 27 scenes across five standard datasets spanning indoor, outdoor, and object-centric environments under severe compression regimes (90% pruning). The paper also thoroughly explores varied settings, including iterative pruning schedules, varying fine-tuning budgets, and a challenging picture-free setup (supervising fine-tuning solely with pseudo-views from the unpruned model), where SplatQ demonstrates robust rate-distortion behavior against eight competing baselines.

Critical weaknesses

W1: Theoretical contradiction with Rate-Distortion Theory and absence of quantitative local error metrics. The paper's foundational premise posits that global Top-K ranking is inherently flawed because it produces "spatial quality imbalance" by over-allocating Gaussians to high-scoring regions. However, in classic rate-distortion theory, optimal bit/capacity allocation dictates allocating higher representational density to high-frequency, geometrically intricate regions (e.g., fine foliage, text) while allocating minimal capacity to smooth, low-frequency regions (e.g., flat walls, sky), in order to minimize global L1L_1/L2L_2 reconstruction loss. Artificially smoothing block quotas (Sec 3.3) inherently redistributes budget from complex foregrounds to simple backgrounds. Crucially, the authors only demonstrate this "spatial quality balance" qualitatively via zoomed-in patches in Figure 1. The paper lacks any quantitative metric (e.g., standard deviation of patch-wise PSNR, patch LPIPS variance, or local error entropy across views) to substantiate that regional quality is truly more balanced, nor does it prove that quota smoothing does not penalize fidelity in high-frequency regions-of-interest (ROIs) compared to global rate-distortion optimal allocations.

W2: Physical disconnect between 3D Euclidean spatial partitioning and ray-marching volumetric occlusion. In 3DGS, the primary non-linear volumetric interaction is ray-depth occlusion governed by the alpha-compositing product C=∑i=1Nciαi∏j<i(1−αj)C = \sum_{i=1}^N c_i \alpha_i \prod_{j<i} (1 - \alpha_j). The strongest joint deletion coupling occurs between foreground and background primitives along the exact same line of sight. However, in Sec 3.1, SplatQ partitions Gaussians into localized blocks using 3D Euclidean spatial proximity (μi\mu_i) and appearance similarity (cic_i). This design guarantees that foreground and background Gaussians along the same ray are segregated into entirely separate optimization blocks. Because the response matrix RbR_b and the resulting QUBO interaction matrix (Sec 3.4) are strictly confined within each individual block, all cross-block interactions along the ray are discarded. The method effectively models only local spatial clustering overlaps while remaining blind to the most critical line-of-sight occlusion effects, contradicting the paper's core claim of being "interaction-aware."

W3: Confounding of combinatorial selection benefits by post-pruning fine-tuning. A major empirical gap is the absence of zero-shot pruning evaluation (evaluating novel-view rendering quality immediately after Gaussian deletion, before any gradient descent fine-tuning is performed). In 3DGS pruning pipelines, post-pruning fine-tuning allows the remaining 10% of Gaussians to adaptively expand their spatial scale (sis_i), shift centers (μi\mu_i), and increase opacities (oio_i) to compensate for missing geometry. Without isolating the pre-fine-tuning rendering metrics, it is impossible to determine whether the reported PSNR/SSIM improvements in Table 1 stem from the QUBO quadratic selection yielding a superior discrete subset, or simply because the spatially smoothed Gaussian distribution provides a more uniform gradient initialization landscape for subsequent Adam optimization.

W4: Omission of offline wall-clock pruning overhead and algorithmic scalability analysis. Traditional Top-K pruning baselines (e.g., LightGaussian, PUP 3D-GS) execute in seconds because they compute closed-form 1D importance vectors. In contrast, SplatQ requires: (1) calculating opacity Jacobians across multiple training camera views to form the response matrix RbR_b (Appendix A.2); (2) accumulating outer products to construct dense pairwise interaction matrices Q∈RB×BQ \in \mathbb{R}^{B \times B} per block; and (3) executing stochastic search algorithms (Simulated Annealing across thousands of Markov chains) to solve NP-hard binary quadratic programs for every block in a scene containing millions of primitives. The paper provides no end-to-end wall-clock time comparisons against LightGaussian or PUP 3D-GS for the pruning phase. In practical 3D asset pipelines, an algorithm requiring hours of offline quadratic optimization for marginal PSNR gains presents a severe usability trade-off that the paper fails to measure or discuss.

W5: Unsubstantiated claims regarding the Coherent Ising Machine (CIM) hardware backend. The paper repeatedly highlights the Coherent Ising Machine (CIM) as an alternative hardware solver backend to alleviate the computational cost of classical QUBO solving. However, Section 4.3 and Appendix A.7 do not clarify whether experiments were performed on actual, physical optical CIM hardware or on a classical numerical simulator running on GPUs (such as SimCIM). If evaluated purely on an emulator, claims regarding hardware acceleration are unsubstantiated. Furthermore, if Simulated Annealing on commodity GPUs already achieves comparable PSNR without requiring specialized optical/quantum hardware, framing CIM as a core contribution adds gratuitous novelty-padding without demonstrated practical necessity.

Questions for authors

Q1:Could the authors provide a rigorous quantitative evaluation of spatial balance across the 27 benchmark scenes? Specifically, what is the variance or standard deviation of patch-level PSNR (e.g., computed on 64×6464 \times 64 pixel grids across test views) for SplatQ versus PUP 3D-GS and LightGaussian? Additionally, when budget smoothing shifts quotas from high-frequency regions to smooth background blocks, does the PSNR on cropped, high-detail foreground ROIs suffer a measurable drop compared to unconstrained global Top-K selection?

Q2: What are the zero-shot quantitative results (PSNR, SSIM, and LPIPS measured immediately after 90% Gaussian pruning, with 0 fine-tuning iterations) for SplatQ vs. PUP 3D-GS, LightGaussian, and an internal ablation using only diagonal response scores? Providing this comparison is essential to prove that the QUBO selection yields an intrinsically superior primitive subset independent of gradient optimization.

Q3: Since alpha compositing couples primitives along the camera ray r(t)=o+tdr(t) = o + t d, partitioning Gaussians by 3D Euclidean distance μi\mu_i (Sec 3.1) ensures that foreground and background primitives along the same ray are placed in different blocks, zeroing out their mutual entry in the interaction matrix. Have the authors measured the fraction of intersecting ray interactions that are lost due to this spatial blocking? Could the authors ablate an alternative ray-frustum or screen-space tile partitioning strategy that preserves line-of-sight coupling?

Q4: What is the exact end-to-end wall-clock runtime for the entire pruning process of SplatQ (broken down into: multi-view response sampling, Gram matrix accumulation, and SA solving) across representative scenes (e.g., Mip-NeRF 360 Bicycle and Garden) on a standard single GPU (e.g., NVIDIA RTX 3090/4090)? How does this total time compare directly to the scoring times of LightGaussian and PUP 3D-GS?

Q5: Was the CIM experiment in Sec 4.3 conducted on a physical optical Coherent Ising Machine, or simulated via software (e.g., GPU/CPU-based SimCIM)? If physical hardware was used, what was the data transfer and compilation overhead between the GPU and the optical setup? If simulated, what is the wall-clock latency of the simulation compared to GPU Simulated Annealing, and what practical advantages does it offer?

5 comments

Thank you for raising specific questions about budget allocation, occlusion modeling, reconstruction quality before fine-tuning, and computational cost. The revised manuscript adds spatial quality statistics, alternative partitioning experiments, comparisons without fine-tuning, and a timing breakdown for the physical CIM backend.

W1 & Q1: Quantitative evaluation of spatial balance

We agree that high-frequency textures and complex geometry often require more representational capacity. Our concern is that independent-score Top-K can retain redundant Gaussians in highly scored, detailed regions while leaving smoother regions underrepresented; it need not provide the optimal allocation when deletion effects interact. Our logarithmic smoothing starts from Top-K quotas and offers adjustable redistribution while preserving unequal block budgets and the exact global count (Section 3.3). With the same diagonal scores and total budget, smoothing improves macro-average PSNR from 27.978 to 28.060 dB (Table 2). The frequency and residual maps in Appendix B.3, Figures 11–19, also show high-detail regions with lower errors than competing methods, alongside low-frequency regions where SplatQ is worse: for example, the sky patch above the railcar in the right-hand Train view of Figure 17 has a higher residual than RAP. We therefore do not interpret smoothing as necessarily improving smooth backgrounds at the expense of high-frequency detail. The visual evidence shows a more varied regional trade-off.

Regarding quantitative spatial balance, SplatQ achieves more consistent regional quality alongside strong overall reconstruction quality, as evidenced by lower patch-PSNR dispersion. Appendix B.2.11 and Table 17 report statistics across all 27 scenes. Each test image is divided into a 16×1616\times16 grid to compute within-image patch-PSNR variance and range, which are averaged over test views and then equally over scenes within each dataset. SplatQ achieves the lowest patch-PSNR variance and range among pruned methods on all five benchmarks and the highest full-image PSNR on four. For NeRF Synthetic and Ref-NeRF, boundary-connected white GT pixels are excluded. Table 17 additionally reports local SSIM variance and range. These results support improved regional consistency alongside overall quality.

W2 & Q3: Ablate an alternative ray-frustum or screen-space tile partitioning

Thank you for pointing out this limitation. We have revised the title to use Local Interaction-Aware, clarifying that joint deletion terms are modeled within blocks.

Our diagnostic measurements indicate that spatial blocking excludes approximately 40% of ray interactions from the within-block objective in the evaluated setting, with this fraction decreasing as block capacity increases. Larger blocks can place previously separated Gaussians along the same ray into a common subproblem. For NN Gaussians and maximum block capacity MM, matrix storage scales with ∑bnb2≤MN\sum_b n_b^2\le MN. The Mip-NeRF 360 capacity study shows that increasing capacity from 256 to 2,048 raises PSNR from 27.02 to 27.09 dB and reduces scoring time (Appendix B.2.6, Table 14). Thus, block capacity provides a trade-off among interaction coverage, reconstruction quality, and storage requirements.

Our default partitioning achieves slightly better reconstruction quality at lower computational cost than shared-ray and screen-tile grouping, potentially because spatially compact blocks require fewer unique sampled pixels for response computation. We evaluated the suggested shared ray and screen tile partitions in Table 9 and Table 10. Shared ray connects Gaussians contributing to the same sampled pixel, whereas Screen tile aggregates contributions within screen-space tiles. On Mip-NeRF 360, graph construction and partitioning take 0.86, 149.18, and 258.64 s for the default, Shared ray, and Screen tile methods, respectively; first-round RbR_b computation takes 16.74, 21.44, and 23.18 min. Both alternatives yield slightly lower macro-average PSNR across the five benchmarks (28.39/28.40 versus 28.43 dB). One possible explanation for the longer response computation is that projection-based groups can contain spatially distant Gaussians, leaving more unique sampled pixels after deduplication.

W3 & Q2: Zero-shot quantitative results

Appendices B.2.4–B.2.5, Tables 12–13, and add quantitative and qualitative evaluations after single-round 80% pruning without fine-tuning. Table 12 fixes pretrained models, partitions and quotas, comparing exact diagonal Top-K selection with SA on the quadratic objective. Table 13 compares SplatQ, LightGS and PUP. Figures 6–7 shows qualitative results.

Our main pruning protocol reaches 90% removal in two stages. For the zero-fine-tuning evaluation, we render the test views immediately after the first stage, with 80% of the original Gaussians removed and zero fine-tuning steps. Pairwise terms improve PSNR by 0.34, 0.30, and 0.28 dB on Mip-NeRF 360, Deep Blending, and Tanks & Temples, with improved SSIM and LPIPS. SplatQ also outperforms both baselines on all three metrics across five datasets. It improves PSNR by 1.29/2.03 dB and SSIM by 0.0144/0.0193, while reducing LPIPS by 0.0142/0.0139 relative to LightGS/PUP. With fixed quotas and no parameter updates, these controlled results show that our interaction-aware joint selection outperforms the compared methods.

W4 & Q4: Exact end-to-end wall-clock runtime for pruning process

To address the missing end-to-end comparison, we measured single-pass 80% pruning on Bicycle and Garden with one RTX 5090, with no fine-tuning or evaluation. SplatQ used SA version. End-to-end pruning times were 3,638.7/3,454.5 s for SplatQ, 68.6/67.1 s for LightGaussian, and 81.0/80.7 s for PUP. LightGaussian’s scoring alone took 6.2/6.0 s, and PUP’s took 20.5/18.8 s.

For Bicycle/Garden, SplatQ’s partitioning took 3.5/3.4 s, response-matrix construction 2,001.8/1,866.1 s, parallel SA 1,468.6/1,400.2 s, 1-swap 156.2/144.9 s and count repair 7.9/9.0 s. Additional overhead data preparation and saving. Please refer to Table 6 for more scoring and selection times across the compared methods.

In practical 3D asset pipelines, SplatQ incurs an offline preparation cost per asset, and reconstruction quality and rendering efficiency affect every use of the compressed model. For assets rendered repeatedly, we consider additional pruning time an acceptable trade-off for improved quality and rendering efficiency. And greedy provides a faster option when turnaround time is a priority.

W5 & Q5: CIM experiments

The CIM experiments in Section 4.3 use a cluster of 15 physical optical coherent Ising machines. To avoid resource contention, we reserved the public-cloud cluster for exclusive use and suspended other compute jobs during the measurements. The available reservation window limited the public-cloud benchmark to the complete batch of 2,396 matrices from the second pruning round of Bicycle.The service schedules the matrices after the entire batch is uploaded. We measured wall-clock time from the start of the upload until all matrices had been solved. Submitting the same batch from the cluster site took 496 s, and submitting it through a Hong Kong VPN endpoint took 1,17 s. See Table 11 in Appendix B.2.3.

In the non-public direct-access mode provided for academic testing, 15 machines solved 15 matrices in parallel. The end-to-end wall-clock time was 1.33 s, including matrix transfer, waiting for automatic optical tuning, and on-device processing. On-device processing took 0.7 s, including matrix retrieval from the scheduler, compilation, and optical evolution. Optical evolution for a single matrix took 0.075 s. Table 11 reports timings at the finest granularity permitted by the provider’s disclosure policy. Individual timings for steps such as transfer and compilation are subject to disclosure restrictions because a finer breakdown could reveal commercially confidential hardware choices and internal system configurations.

The Mip-NeRF 360 evaluation in Table 6 covers all nine scenes, with one pass of 80% pruning per scene and 59,112 QUBO matrices in total. The table reports per-scene measurements averaged equally across the nine scenes. The first two rows of Table 11 use this batch size to report greedy and SA solver runtimes on a single RTX 5090.

In Table 6, the average selection times for SA and direct-access CIM are 845.9 s and 581.4 s, respectively. The main SA configuration achieves a macro-average PSNR of 28.426 dB, and the CIM configuration achieves 28.488 dB (Table 2). These measurements show a shorter selection stage and a modest gain in reconstruction quality with CIM. The time savings for the full pruning pipeline remain limited at present.

There is room to improve both matrix transfer from the client to the CIMs and matrix retrieval from the scheduler. Faster transfer can reduce end-to-end wall-clock time, and faster retrieval can reduce both on-device processing time and end-to-end latency. As these engineering improvements reduce overhead, we expect the selection-stage speed advantage to translate more fully into time savings for the complete pruning pipeline.

Thank you again for your careful review. We hope the additional experiments and clarifications address the concerns you raised. If the revised evidence changes your assessment of the work, we would appreciate your consideration of an update to your overall rating.

Reviewer_J8zVOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The proposed method compresses 3DGS scenes by grouping nearby Gaussians and selecting which to keep based on both their individual importance and their interactions, helping preserve details more evenly across the scene. It formulates this selection as a QUBO optimisation problem and refines the remaining Gaussians without adding new ones. Experiments show improved visual quality and rendering speed when removing 90% of the Gaussians, but the method requires substantially more computation during compression.

Critical strengths
  • The motivation is clear, and the proposed methodology follows logically from it.
  • Comprehensive results and comparisons with the baselines are provided.
Critical weaknesses
  • The partition is a plausible engineering approach, but its geometric justification and benefit over simpler partitioning are not fully established.
  • Section 3.1, the spatial similarity used to partition Gaussians into blocks considers their centres and scales but ignores orientation. For elongated or direction-dependent structures, this may misrepresent Gaussian overlap and place strongly interacting Gaussians in different blocks. Please justify this approximation and compare it with a similarity measure that uses the full covariance matrices.
  • Figure 3 provides only indirect support for the proposed explanation of rendering speed. Smaller 3D Gaussian volumes do not necessarily imply smaller screen-space footprints or lower rendering cost.
  • Several qualitative comparisons are provided in the supplementary material, but no model size information is given, making it difficult to judge.
Questions for authors

Please see weaknesses above.

8 comments

Thank you for carefully reviewing our work and recognizing its motivation, methodological design, and breadth of evaluation. Your suggestions have helped us clarify the role of partitioning and supplement the paper with comparisons of partitioning strategies, screen-space rendering workload measurements, and actual model sizes. We address your concerns below.

W1: Geometric justification and benefit over simpler partitioning

The geometric rationale for our partitioner is that spatially nearby Gaussians with overlapping extents are more likely to affect the same local image regions and should therefore be considered jointly. We limit blocks to 512 members to control matrix storage and solving costs. The affinity measures center separation relative to Gaussian scales and incorporates color similarity, favoring nearby candidates with related appearance (Section 3.1).

Our partitioner improves on simpler alternatives by grouping Gaussians using spatial proximity, scale, and appearance, producing more suitable candidates for joint selection in rendering. Appendix B.2.2 and Tables 9–10 compare seven partitioners under two-round pruning to 10% retention with 10k fine-tuning steps per round. The default achieves the best macro-average quality (28.43 dB PSNR, 0.8839 SSIM, and 0.1750 LPIPS), versus 28.17 and 28.38 dB for random and Morton partitioning. This supports preserving spatial locality and further refining groups by scale and appearance for joint selection. On Mip-NeRF 360, default partitioning takes 0.86 s versus 0.22/0.31 s for random/Morton, providing better quality at a small additional partitioning cost. Random grouping also incurs greater subsequent response-computation cost. Thus, the benefit concerns both candidate organization and downstream computation.

W2: Justify this approximation and compare it with a similarity measure that uses the full covariance

As explained above, our partitioner uses scale-aware spatial proximity and color similarity to group Gaussians likely to affect the same local image regions. This is a low-cost approximation for organizing joint-selection candidates.

Our default approximation provides slightly better aggregate quality at substantially lower preprocessing cost because more detailed geometric overlap does not necessarily improve interaction-aware grouping, while full covariance requires expensive matrix-based distance calculations. To evaluate this approximation, Appendix B.2.2 replaces the default geometric distance with a full-covariance Mahalanobis distance while retaining spatial candidates and color similarity. Table 9 reports macro-average PSNRs of 28.43 dB for the default and 28.41 dB for full covariance, indicating no aggregate quality gain from the richer geometry in these experiments. Meanwhile, full covariance requires matrix-based distance calculations: graph construction and partitioning on Mip-NeRF 360 take 155.36 s versus 0.86 s for the default (Table 10). These results support the default approximation as providing slightly better aggregate quality at substantially lower preprocessing cost.

W3: Explanation of rendering speed

Appendix B.2.10 and Table 16 add projected coverage, candidate checks, Gaussian–tile pairs, and blending counts. All metrics are measured across all test views, first averaged equally over views within each scene and then equally over scenes within each dataset.

Our lower projected coverage, fewer candidate checks and Gaussian–tile pairs, and fewer blending operations jointly help explain SplatQ’s FPS advantage. These metrics directly measure screen-space workload. Relative to PUP on Mip-NeRF 360, SplatQ reduces coverage, checks, tile pairs, and blending by 19.2%, 16.6%, 23.4%, and 10.1%. It has the lowest coverage on all five datasets and the fewest checks, tile pairs, and blends on four. These observations support reduced rendering workload as an explanation for the FPS gains in Table 1.

W4: Model size information

Main comparisons share pretrained models, retained counts, and uncompressed PLY representations, so pruned 3DGS sizes coincide between every method at the reported precision. Table 15 adds original, pruned, and attribute-compressed model sizes. The main pruning comparison uses FP32 attributes with the same attribute configuration and retained counts. At 90% pruning, average sizes are 83.39, 73.80, 44.24, 7.16, and 4.13 MB across the five datasets, matching quantitative storage budgets.

Conclusion

Thank you for your careful review. Your feedback helped us clarify the role of partitioning, compare alternative partitioning strategies, and report screen-space rendering workloads and actual model sizes. We have addressed each of your concerns, as detailed below. In light of the paper’s contributions and the additional experiments and analyses in the revision, we would be grateful if you could reconsider your overall score.

As the rebuttal period draws to a close, I encourage reviewers to engage with the authors’ responses and revised manuscript, clearly identifying which concerns have been addressed and which remain unresolved.

Given the limited time remaining, please prioritize evaluating the evidence and clarifications already available in the current revision. If you nevertheless consider additional experiments essential to support the paper’s claims, please specify what is needed and why. Even if the authors cannot complete these experiments within the remaining time, a discussion of the underlying concerns and the available evidence would help inform our final assessment.

Thank you all for your time and constructive engagement

Thank you for your reply.

I have to say that, in Table 9, all six partitioners that preserve spatial locality are within 0.05 dB of each other in macro PSNR (28.38–28.43 dB). Our results are not significantly different from the others, as the performance gaps are small relative to this variation. Similarly, for W2, the difference between 28.43 dB and 28.41 dB can be considered insignificant.

Is the time difference you mentioned for the partitioning process only, or does it refer to the end-to-end runtime?

I also think the point about "For elongated or direction-dependent structures, this may misrepresent Gaussian overlap and place strongly interacting Gaussians in different blocks" has not yet been addressed.

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

SplatQ proposes a post-training pruning framework for 3D Gaussian Splatting. It partitions Gaussians using spatial and appearance similarity, assigns smoothed retention quotas, and performs joint selection using a quadratic objective derived from opacity-response Jacobians. The authors evaluate greedy selection, simulated annealing (SA), and a coherent Ising machine (CIM) backend. Experiments across five benchmarks show improvements in reconstruction quality and rendering speed, with ablations supporting the benefit of pairwise interactions.

Critical strengths
  1. Useful interaction modeling. Comparisons against the same diagonal scores, including a fixed-quota experiment without fine-tuning, support the benefit of joint selection.
  2. Broad evaluation and solver exploration. The experiments use shared checkpoints and refinement settings across five benchmarks. The CIM backend is an interesting effort to improve the selection process.
Critical weaknesses
  1. Additional quantitative evaluation. The pruning-ratio curves provide a useful comparison. Reporting model sizes and a summary metric such as BD-rate over a common quality range would further clarify the storage–quality trade-off across methods.
  2. Relationship to GaussianPOP. The diagonal score D_i = 〖‖J_i ‖〗^2 appears equivalent to GaussianPOP’s squared single-deletion error under matching assumptions and sampling. This connection should be explained alongside the derivation; a direct comparison is also missing.
  3. Substantial preparation cost. I appreciate the CIM implementation, but scoring alone takes about 1,005–3,456 seconds. SA adds only 0.052 dB over greedy while taking 7.6–8.4× longer to solve. The full pipeline’s quality–time trade-off remains unclear.
  4. Rotation omitted from the affinity. Equation (1) applies local-axis scales to world-coordinate distances without accounting for rotation. This may misgroup anisotropic Gaussians and exclude relevant interactions; justification or an ablation is needed.
  5. Limited discussion of compression integration. The paper focuses on pruning, with little discussion of how it could be combined with existing attribute compression methods. Clarifying this compatibility would help readers understand its role in a complete compression pipeline.
Questions for authors
  1. Compression and rate–distortion. Could the authors discuss how SplatQ could be combined with existing attribute compression methods? Reporting model sizes and BD-rate would also help assess its compression performance relative to the baselines.
  2. Relationship to GaussianPOP. Could the authors clarify the mathematical relationship between their diagonal score and GaussianPOP’s single-Gaussian deletion error alongside the derivation?
  3. Computational cost. Greedy achieves results close to SA at substantially lower cost. Could the authors discuss the trade-off between the two solvers and whether greedy could serve as a practical alternative?
  4. Partitioning metric. Why is rotation omitted from the affinity? Does using a rotation-aware covariance distance change partitioning or reconstruction quality?

Additional Minor Comments

  1. FPS interpretation. The volume CDFs on Tanks & Temples and Ref-NeRF appear similar to those of some competing methods despite the FPS advantage. Are there additional insights concerning projected coverage, visibility, or blending workload?
  2. Storage precision. Please specify the attribute storage precision and format across methods.
  3. Patch PSNR. Since pure background patches are included, please clarify how zero or near-zero MSE is handled when computing patch PSNR and its variance and range.
5 comments

Thank you for carefully reviewing our work and for recognizing the value of interaction modeling, controlled ablations, and evaluation across datasets. Your suggestions have helped us clarify the relationship between SplatQ and prior work and strengthen our analysis of the storage–quality trade-off, partitioning design, and computational cost. We have revised the main text and appendix accordingly. Below, we address your comments individually.

W1 & Q1: Additional quantitative evaluation

Table 15 reports original, pruned, and attribute-compressed model sizes. Appendix B.2.7 and Figure 8 add BD-rate over nine pruning ratios (10%–90%).

At 90% pruning, average sizes are 83.39, 73.80, 44.24, 7.16, and 4.13 MB across the five datasets, matching quantitative storage budgets. Uncompressed 3DGS sizes coincide between every method at the reported precision. Relative to SplatQ, PUP requires +40%, +43%, +92%, +74%, and +44% Gaussian budgets across the five datasets. MesonGS and C3DGS instead require 30% and 25% less on Tanks & Temples. LightGaussian on Deep Blending is N/A because the quality ranges do not overlap. Table 15 separately reports compressed sizes. The BD-rate results indicate that SplatQ requires fewer Gaussians to achieve comparable reconstruction quality in most scenes.

W2 & Q2: Relationship to GaussianPOP

At the time of our check, GaussianPOP’s official implementation is not released. We therefore could not conduct a full pipeline comparison using its official implementation.

First, GaussianPOP ranks single-deletion errors with iterative rescoring, whereas SplatQ combines regional quotas with pairwise interactions for joint selection. GaussianPOP derives its score from the color difference before and after deleting one Gaussian [GaussianPOP, Eq. (4)]. SplatQ instead differentiates rendered colors with respect to per-Gaussian opacity multipliers to obtain a response Jacobian JJ [Eq. (4); Appendix A.3, Eq. (11)]. Its Gram matrix J⊤JJ^\top J defines a quadratic approximation to joint deletion error, including reinforcement or cancellation between deletion responses.

Second, the connection between their single-Gaussian scores holds under specific conditions. With fixed compositing paths, rendered color is affine in a single opacity multiplier. Changing that multiplier from 1 to 0 deletes the Gaussian, so its derivative equals the color difference between retaining and deleting it. With matched sampled pixels, background treatment, and aggregation, SplatQ’s diagonal term therefore equals the squared single-deletion error. In practice, SplatQ samples representative pixels within each block, so the accumulated scores need not match entry by entry.

Third, this diagonal equivalence does not make the complete selection objectives identical. SplatQ combines regional budget allocation with off-diagonal interactions: diagonal terms measure individual deletion effects, while cross terms capture how these effects reinforce or cancel. Two deletion sets with the same diagonal cost can therefore have different joint quadratic costs. The distinction lies in selecting Gaussians jointly, not merely in deriving their individual scores differently.

Finally, Table 12 and Figure 6 isolate the benefit of interaction-aware selection. With pretrained models, partitions, and per-block quotas fixed, we compare diagonal Top-K with the full quadratic objective under single-round 80% pruning without fine-tuning. PSNR improves by 0.34, 0.30, and 0.28 dB on Mip-NeRF 360, Deep Blending, and Tanks & Temples, respectively, with improved SSIM and LPIPS. These results support joint selection beyond the same individual scores; they are a controlled ablation, not a full pipeline comparison with GaussianPOP.

W3 & Q3: Substantial preparation cost

On one hand, we chose SA for its reconstruction quality: Table 2 reports a macro-average PSNR of 28.426 dB for SA versus 28.374 dB for greedy. Greedy is also a practically valuable solver option and can serve as a fast variant of SplatQ when preparation time is a priority.

On the other hand, our main contribution is the interaction-aware pruning formulation, which combines regional budget allocation with joint Gaussian selection, rather than a particular QUBO solver. Both greedy and SA optimize this formulation with the same postprocessing, allowing users to choose between preparation cost and reconstruction quality.

Although greedy accelerates selection by 7.6–8.4×, scoring remains the dominant cost, so the overall speedup is only 1.67× on Mip-NeRF 360: combined scoring and selection time decreases from 1,850.9 to 1,106.1 s, a 40.2% reduction (Appendix B.2.3, Table 6). Greedy is therefore suitable for time-sensitive preparation, while SA offers a modest quality gain at additional offline cost. We prioritize reconstruction quality and rendering efficiency in pruning and therefore use SA in our main experiments.

W4 & Q4: Rotation omitted from the affinity

Appendix B.2.2 compares seven partitioners under the same pruning and fine-tuning settings: default, random, Morton, rotation-aware, full-covariance, shared-ray, and screen-tile partitioning. Table 9 reports reconstruction quality; Table 10 reports partitioning time and response-matrix computation timed separately.

Table 9 shows that default, rotation-aware, and full-covariance partitioning achieve macro-average PSNRs of 28.43, 28.39, and 28.41 dB, respectively, with the default also achieving the best macro-average SSIM and LPIPS. These results suggest that richer geometric information does not necessarily produce better groups for joint selection. Similarly oriented Gaussians may affect different pixels, whereas differently oriented Gaussians may have correlated deletion responses.

Table 10 shows the corresponding cost trade-off. On Mip-NeRF 360, graph construction and partitioning take 0.86 s for the default, 0.98 s for rotation-aware, and 155.36 s for full covariance. Full covariance requires matrix-based distance calculations and incurs approximately 181× the default runtime. Together with Table 9, these results support the default design as offering both the best aggregate reconstruction quality and the lowest preprocessing cost among these three variants.

An illustrative counterexample explains why angular similarity need not better reflect Gaussian geometry. Consider two isotropic Gaussians with identical centers, scales, colors, and opacities, but rotations differing by 90°. They are consecutive contributors along a ray, and their deletion responses are also identical under fixed compositing paths, making them strongly coupled candidates for joint selection. Nevertheless, their quaternion angular distance is nonzero. Thus, angular distance can distinguish Gaussians whose rendering responses are identical, illustrating why orientation similarity need not improve contribution-aware grouping.

W5 & Q1: Limited discussion of compression integration

Section 4.4 and Tables 3–4 apply LightGaussian, MesonGS, EAGLES, and C3DGS attribute compression to 90%-pruned inputs from SplatQ, PUP, RAP, REFINE, and each compressor’s own pruning method. We evaluate reloaded compressed models; Table 15 adds full storage and runtime results.

SplatQ inputs achieve the highest PSNR in 18/20 compressor–dataset combinations and the highest decoded-model FPS in all 20. On Mip-NeRF 360, their 83.39 MB pruned inputs become 16.28, 5.78, 26.30, and 7.48 MB under the four compressors. With MesonGS, SplatQ and PUP both occupy 5.78 MB, while SplatQ achieves 26.59 versus 26.39 dB and 987.4 versus 857.6 FPS. These results support combining our selection with attribute compression. Further details are provided in the revised manuscript.

Additional Minor Comments

1. FPS interpretation

Appendix B.2.10 and Table 16 add direct rendering-workload measurements. All metrics are measured across all test views, first averaged equally over views within each scene and then equally over scenes within each dataset.

Smaller 3D volumes need not imply less screen-space workload. Relative to PUP, SplatQ reduces coverage from 78.11 to 68.02 and tile pairs from 1.664 M to 1.423 M on Tanks & Temples. On Ref-NeRF, these decrease from 38.60 to 30.99 and from 0.734 M to 0.614 M. Candidate checks and blending operations also decrease, supporting the FPS explanation beyond volume CDFs. Further details are provided in the revised manuscript.

2. Storage precision

Appendix B.2.9 and Table 15 report the shared PLY format and pruned sizes.

The main pruning comparison uses FP32 attributes with the same attribute configuration and retained counts. Consequently, pruned and uncompressed 3DGS sizes coincide between every method at the reported precision.

3. Patch PSNR

Figure 4 is updated using the protocol in Appendix B.2.11, with numerical results in Table 17. Each image has a 16×16 grid. NeRF Synthetic and Ref-NeRF exclude boundary-connected white GT pixels (four-connectivity; all RGB channels ≥250) and other datasets retain all pixels. Empty cells are omitted.

MSE uses native 8-bit RGB and zero MSE maps to 100 dB. Positive MSE uses the standard PSNR formula without capping finite values. Population variance and range are computed within each image, then averaged over views and scenes.

Thank you again for your careful and detailed review. We have added experiments and strengthened the analysis to address your main concerns, while further clarifying SplatQ’s relationship to prior work. In light of the updated evidence, we respectfully ask you to reassess the algorithmic contributions of this work, particularly the originality and research value of formulating 3D Gaussian pruning as a combinatorial optimization problem and exploring coherent Ising machine solving with a view toward quantum computing applications. If these additions satisfactorily address your concerns, we would be grateful if you would consider raising your score. We will also carefully address any further comments or suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.