- 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 / 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 . 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 () and appearance similarity (). This design guarantees that foreground and background Gaussians along the same ray are segregated into entirely separate optimization blocks. Because the response matrix 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 (), shift centers (), and increase opacities () 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 (Appendix A.2); (2) accumulating outer products to construct dense pairwise interaction matrices 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 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 , partitioning Gaussians by 3D Euclidean distance (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
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.
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 16×16 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 N Gaussians and maximum block capacity M, matrix storage scales with ∑bnb2≤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 Rb 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.