{"id":"60dcb5db-a456-4eb9-90ba-999beb98ed00","arxiv_id":"2508.12313","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"A set of densification and training modifications for 3D Gaussian Splatting that improves rendering quality and reduces the number of Gaussians needed.","lead":"This paper proposes several improvements to the densification stage of 3D Gaussian Splatting: an edge-aware score for choosing which Gaussians to split, a long-axis split strategy, and three training tricks to reduce overfitting. The authors report better rendering quality with fewer Gaussians than 3DGS and several recent variants on standard benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unablated hyperparameter changes listed in Implementation Details (position_lr_init/final, opacity_reset_threshold, disabled large-Gaussian pruning) may explain a large share of the reported gains, so the densification contribution is not yet isolated from these confounds.","rationale":"The reader's weakest assumption was that cloning is unnecessary and that the w/o LAS ablation does not isolate the removal of clone from the introduction of Long-Axis Split. That is a valid experimental-design concern. However, the most load-bearing issue is the unablated hyperparameter and pruning changes listed in the Implementation Details. These changes affect the baseline comparison and the same-budget comparison, and they are not controlled in any ablation. If the modified hyperparameters alone reproduce most of the improvement, then the paper's central claim that its densification strategy enhances fidelity would be undermined, even though the full pipeline may still perform well. The overhead contradiction noted by the reader is also real but is a wording issue rather than a threat to the technical contribution. Therefore I partially agree with the reader: the comparison-protocol concern overlaps, but the weakest assumption should be the hyperparameter confound rather than the clone ablation. The verdict remains CONDITIONAL: the paper should be accepted only if the authors add an ablation of the Implementation Details modifications or show that these changes are necessary parts of the proposed method and are isolated from the densification contributions.","tokens_in":16674,"tokens_out":11015,"duration_ms":120644,"concrete_test":"Run the original 3DGS (or TamingGS) baseline with only the Implementation Details modifications applied--position_lr_init=0.00004, position_lr_final=0.000002, opacity_reset_threshold=0.05, and no large-Gaussian pruning--while keeping the original clone/split densification and default Gaussian budget. Compare PSNR/SSIM/LPIPS on Mip-NeRF 360 against both the default baseline and Ours Full. If this modified baseline closes most of the gap, the proposed densification techniques are not the primary source of the reported gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that its densification pipeline improves rendering fidelity and achieves state-of-the-art results with fewer Gaussians. For that claim to hold, the proposed densification techniques (EAS, LAS, RAP, MU, GC) must be the source of the improvements over baselines. The Implementation Details appendix, however, lists several non-densification modifications applied to the authors' method but not to the baselines: position_lr_init is set to 0.00004, position_lr_final to 0.000002, opacity_reset_threshold to 0.05, and pruning of large Gaussians is removed. These changes are never ablated. Since Table 2 shows individual proposed components contribute only about 0.1--0.3 dB PSNR on the combined dataset, a 25x lower position learning rate and a 5x higher opacity reset threshold could plausibly account for a substantial part of the 0.71 dB PSNR gap over 3DGS on Mip-NeRF 360. The same-budget comparison in Appendix A.2 also uses these modified settings for Ours while TamingGS and 3DGS-MCMC use their defaults. Therefore the empirical evidence does not yet establish that the densification pipeline, rather than auxiliary hyperparameters, is responsible for the reported state-of-the-art quality.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a set of improvements to the densification pipeline of 3D Gaussian Splatting, organized around three questions: when to densify (Edge-Aware Score, EAS), how to densify (Long-Axis Split, LAS), and how to mitigate overfitting (Recovery-Aware Pruning, Multi-step Update, Growth Control). The method is evaluated on Mip-NeRF 360, Tanks and Temples, and Deep Blending, with quantitative comparisons against eight recent baselines, per-component ablations, and a same-budget comparison against TamingGS and 3DGS-MCMC. The paper reports consistent gains in PSNR and SSIM with markedly fewer Gaussians, and claims in the Abstract and Conclusion that these gains come 'without introducing additional training or inference overhead.' The main quantitative results are credible at face value, but the no-overhead claim is directly contradicted by the paper's own ablation table, and several non-densification hyperparameter changes in the Implementation Details are not ablated, so the attribution of the reported gains to the densification components is not yet established.","tokens_in":16976,"tokens_out":5761,"duration_ms":61781,"significance":"If the central claims withstand scrutiny, this is a useful contribution: a drop-in densification improvement for 3DGS that improves fidelity and compactness, with code released and component-level ablations. The empirical evidence does show that each proposed component helps on the combined benchmark, and the same-budget comparison in Appendix A.2 is a commendable attempt to control for Gaussian count. However, the paper's headline 'no overhead' claim is false under its own measurements, and the unablated hyperparameter changes noted in the Implementation Details weaken the attribution of the state-of-the-art numbers to the proposed densification techniques. The core idea is defensible and the experiments are extensive, but the manuscript needs substantive revision before the claims can be accepted as stated.","major_comments":[{"comment":"The central claim that the method works 'without introducing additional training or inference overhead' is contradicted by the ablation data in Table 2: Ours Full runs at 289 FPS while the w/o EAS row runs at 346 FPS, a 16% reduction in rendering speed, and the ablation text in 'Effect of EAS' explicitly acknowledges that 'EAS increases rendering overhead.' The same table also shows training time rising from 6.4 to 6.7 minutes. Please remove the no-overhead claim from the Abstract and Conclusion, or rephrase it as an explicit quality-versus-speed trade-off with the overhead quantified.","section":"Abstract / Conclusion; Table 2, Ablation Experiments"},{"comment":"The implementation applied to Ours but not to the baselines includes position_lr_init=0.00004, position_lr_final=0.000002, opacity_reset_threshold=0.05, and removal of large-Gaussian pruning. These changes are catalogued in the Implementation Details but are never ablated. Since Table 2 shows that each proposed component contributes roughly 0.1-0.3 dB PSNR on the combined datasets, the 0.71 dB gain over 3DGS on Mip-NeRF 360 could plausibly be driven in part by these unablated hyperparameters. Please ablate the proposed components while keeping the original 3DGS hyperparameters, or apply the modified settings to all baselines, including the same-budget comparison in Appendix A.2.","section":"Appendix: Implementation Details; Tables 1 and 2"},{"comment":"The decision to abandon the clone operation is justified in the text by the premise that 'during optimization, the size of Gaussians tends to converge to a value that balances under-reconstruction and over-reconstruction,' but that premise is not proven and the w/o LAS ablation removes both LAS and the clone operation simultaneously. The reader therefore cannot tell whether the measured improvement comes from the long-axis split geometry, from dropping clone, or from both. Please add ablations that (i) keep clone with LAS and (ii) use the original split operator without clone, so the contribution of each design choice is isolated.","section":"Methods, Long-Axis Split; Table 2, w/o LAS row"},{"comment":"The claimed proof that Rs = R0 * sqrt(1 - d^2/L0^2) minimizes the geometric difference after splitting is presented as an 'intuitive geometric comparison,' but the written argument does not define the overlapping and non-overlapping area expressions that are being compared, nor does it derive the optimum of their difference. Since Eq. (9) is used to motivate a load-bearing design choice, please either provide a rigorous derivation or explicitly label the formula as a heuristic that is supported only by the sensitivity study in Figure 12.","section":"Appendix A.4, Eq. (9)"}],"minor_comments":[{"comment":"The number Ns of sampled views used to compute the average edge-aware score is never specified; please state its value and the exact threshold used for the absolute gradient criterion in Eq. (8).","section":"Methods, Edge-Aware Score, Eq. (7)"},{"comment":"The statement that 'all results represent the best performance obtained over three runs' is statistically weak; please report mean and standard deviation across runs, or justify why best-of-three is the appropriate summary for this benchmark.","section":"Experiments, Quantitative Analysis"},{"comment":"The choice of pruning at iterations 3300 and 6300 is motivated by the opacity reset schedule, but the exact relationship between these iterations and the 3DGS resets at 3K, 6K, 9K, and 12K is not stated; please make this connection explicit.","section":"Methods, Recovery-Aware Pruning"},{"comment":"The section heading 'Muti-view Update' contains a typo and should read 'Multi-view Update.'","section":"Methods, heading"}],"recommendation":"major_revision","confidential_remarks":"This is a promising empirical paper, and the component ablations are a definite strength. The main risks are the unablated hyperparameter changes in the Implementation Details and the contradiction between the no-overhead claim and Table 2. If the authors add the requested ablations and correct the claim, I would view the paper favorably; at present the evidence supports a useful method but not the headline claims in their current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Xiaobin and colleagues have written a useful, incremental paper on 3DGS densification. The core idea is to replace the standard gradient-threshold plus clone/split procedure with an edge-aware selection score (EAS), a deterministic long-axis split (LAS), and three training-stage tricks (RAP, MU, GC). On Mip-NeRF 360, Deep Blending, and Tanks and Temples, they report consistent PSNR/SSIM gains over eight baselines, about 0.7 dB over 3DGS on the big benchmark, with noticeably fewer Gaussians. That is a solid result, and the ablation table suggests each component earns its place.\n\nWhat is genuinely new: the combination of absolute gradients with per-pixel Laplacian edge weights is a sensible way to target blur at edges; the LAS geometry (children placed on the longest axis with computed radii and reduced opacity) is a clear improvement over random sampling; and the idea of pruning early, then switching to multi-view batching after densification, is a nice practical trick. Full ablations, per-scene numbers, and an appendix with same-budget comparisons for TamingGS and 3DGS-MCMC—this is more thorough than many papers at this level.\n\nThe soft spots are real but mostly fixable. First, the abstract claim of 'no additional inference overhead' is contradicted by the authors' own Table 2: adding EAS drops FPS from 346 to 289. The text later says EAS increases rendering overhead—that is a direct inconsistency. Second, the Implementation Details list non-densification hyperparameter changes (position lr init 0.00004 vs 3DGS default 0.00016, opacity reset threshold 0.05 vs 0.01, and removal of large-Gaussian pruning) that are never ablated. The stress-test note says '25x lower position learning rate,' which is wrong—it is 4x lower—but the core concern is valid: these settings plausibly contribute to the gap over baselines, and the paper does not isolate the densification contribution. Third, the decision to drop clone is bundled with LAS in the ablation, so the premise that clone is unnecessary is not directly tested. Fourth, all results are reported as best of three runs with no variance; that makes 0.1–0.3 dB component differences hard to evaluate.\n\nOverall the work deserves a serious referee. The right path is peer review with a request for ablation of the changed hyperparameters, a corrected overhead statement, and ideally error bars. The paper is aimed at 3DGS practitioners who want a drop-in densification improvement; for them it is worth reading even in this form.","headline":"Solid incremental 3DGS densification improvements, but unablated hyperparameter changes and a contradicted 'no overhead' claim need fixing before the headline claims are credible.","tokens_in":17512,"tokens_out":2493,"would_cite":false,"duration_ms":24342,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A reworked densification pipeline renders sharper 3D scenes from fewer Gaussians.","keywords":["3D Gaussian Splatting","densification","novel view synthesis","edge-aware score","long-axis split","adaptive density control","overfitting","real-time rendering"],"falsifier":"Run the same method on the Mip-NeRF 360 scenes with the clone operation restored alongside Long-Axis Split while keeping Edge-Aware Score and all overfitting controls fixed; if restoring clones improves PSNR or SSIM on any scene, the clone-free premise is wrong. Alternatively, construct a scene with large uniformly textured under-reconstructed regions, where no Gaussian has high edge-aware score, and check whether splitting alone can densify them.","tokens_in":16454,"feed_emoji":"🎨","tokens_out":4590,"duration_ms":43725,"temperature":0.7,"pith_summary":"The paper argues that the bottleneck in 3D Gaussian Splatting is not the rendering primitive but the densification schedule that grows the Gaussian set. It proposes replacing the standard clone-and-split procedure with an edge-aware candidate score, a deterministic long-axis split, and a set of overfitting controls, and claims this yields better reconstruction quality while using fewer Gaussians and adding no training or inference overhead. A sympathetic reader would care because the fix is local to the density-control stage, so it can be dropped into existing 3DGS-based systems without changing their CUDA kernels, and it makes high-fidelity real-time rendering cheaper.","feed_headline":"Edge-aware splitting sharpens 3D scenes, halves Gaussian count","feed_subtitle":"A redesigned densification pipeline improves image quality while cutting Gaussian count, with no added training or render cost.","key_machinery":"The load-bearing device is Long-Axis Split: instead of probabilistically sampling child positions, each parent is replaced by two children centered on its longest axis at distance $d=0.45L_0$ from the parent center, with the child's longest semi-axis set to $L_0-d$ and its two minor semi-axes set by $R_s=R_0\\sqrt{1-d^2/L_0^2}$, so the children are tangent to the parent surface and the before/after shape difference is minimized. Around it, the Edge-Aware Score $S_{i,j}=\\sum_{p\\in P}\\omega^p_{i,j}\\alpha^p_{i,j}$ selects which Gaussians split by weighting each Gaussian's per-pixel rendering weight by a Laplacian edge map, and Growth Control imposes a convex budget curve $N=N_{\\max}\\sqrt{(I-I_{\\text{start}})/(I_{\\text{end}}-I_{\\text{start}})}$ so the Gaussian count peaks only at the end of densification. Together these determine when to densify, which Gaussians to split, and how to split them with minimal disruption.","core_discovery":"On the paper's own terms, the discovery is that the geometric disturbance introduced when a Gaussian is cloned or split is itself a major source of reconstruction error, and that this disturbance can be largely removed. The paper claims that scoring split candidates by their summed edge-weighted rendering contribution (Edge-Aware Score), splitting deterministically along the parent's longest axis with child size $R_s=R_0\\sqrt{1-d^2/L_0^2}$, opacity scaled to 60%, and pruning/update/growth controls to curb overfitting, together outperform prior densification methods on Mip-NeRF 360, Tanks and Temples, and Deep Blending while using the smallest Gaussian budget.","pith_inferences":["The paper does not isolate the contribution of dropping clone: its w/o LAS ablation removes both Long-Axis Split and the clone operation at once, so restoring cloning alongside LAS would be a direct test of whether the clone-free premise is necessary.","The Laplacian edge weighting is only one choice of edge operator, so the Edge-Aware Score could plausibly be combined with perceptual losses or saliency maps, though the paper does not test this.","The growth-control curve resembles a progressive-training schedule, so the same budget law might transfer to other resource-limited Gaussian refinement settings such as SLAM or dynamic scenes, an extension the paper does not claim.","Because the method sets children's opacity to 60% of the parent, its success suggests that opacity re-balancing after split is a free gain that other densification schemes could adopt independently of the geometric placement."],"forward_implications":["Because the changes are confined to adaptive density control, any 3DGS-based renderer that currently uses the original clone/split schedule can adopt this pipeline as a drop-in replacement.","Fewer Gaussians at equal or better quality directly reduces memory footprint and per-frame render cost, which matters for mobile and embedded deployment.","The clone-free densification claim, if correct, implies that under-reconstruction can always be repaired by splitting alone, simplifying the densification operator to a single deterministic rule.","The overfitting controls (pruning, multi-step updates, growth cap) are decoupled from the split mechanism, so each can be tuned independently when porting to new scene types.","The reported training-time reduction (roughly half that of the fastest baseline under the same CUDA kernel) follows from both the lower peak Gaussian count and the multi-step update schedule."],"supporting_citations":[{"why":"Supplies the 3DGS baseline, the adaptive density control being improved, and the rendering formulation.","marker":"(Kerbl et al. 2023)"},{"why":"Identifies gradient conflicts in view-averaged positional gradients and motivates the absolute-gradient criterion used inside the Edge-Aware Score.","marker":"(Ye et al. 2024)"},{"why":"Provides the TamingGS codebase and budget-based training setup that the paper builds on and compares against.","marker":"(Mallick et al. 2024)"},{"why":"Supplies the Mip-NeRF 360 dataset that anchors the main evaluation.","marker":"(Barron et al. 2022)"},{"why":"Defines the LPIPS perceptual metric used to report the quality gains.","marker":"(Zhang et al. 2018)"},{"why":"Provides the Structure-from-Motion initialization that determines the initial Gaussian set in the pipeline.","marker":"(Schonberger and Frahm 2016)"}],"fun_headline_variants":["3D rendering fidelity up, Gaussian count down","Edge-aware densification cuts Gaussians, boosts quality","Long-axis split strategy improves 3D Gaussian rendering","Overfitting-free 3DGS with recovery-aware pruning","No-cost fidelity boost for 3D Gaussian Splatting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's strongest unproven premise is that cloning can be dropped entirely because, during optimization, Gaussian sizes converge to a value that already balances under- and over-reconstruction, so every needed densification is a split; if that balance does not hold in some scene, the method has no clone operation to fall back on.","fun_headline_variants_meta":{"raw":{"variants":["3D rendering fidelity up, Gaussian count down","Edge-aware densification cuts Gaussians, boosts quality","Long-axis split strategy improves 3D Gaussian rendering","Overfitting-free 3DGS with recovery-aware pruning","No-cost fidelity boost for 3D Gaussian Splatting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000342,"raw_usage":{"total_tokens":1821,"prompt_tokens":820,"completion_tokens":1001,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":436,"completion_tokens_details":{"reasoning_tokens":923}},"tokens_in":436,"tokens_out":1001,"duration_ms":9475,"temperature":1.0,"reasoning_tokens":923,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:23:18.677054+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same method on the Mip-NeRF 360 scenes with the clone operation restored alongside Long-Axis Split while keeping Edge-Aware Score and all overfitting controls fixed; if restoring clones improves PSNR or SSIM on any scene, the clone-free premise is wrong. Alternatively, construct a scene with large uniformly textured under-reconstructed regions, where no Gaussian has high edge-aware score, and check whether splitting alone can densify them.","supporting_citations":[{"cited_title":"S.; Goel, R.; Kerbl, B.; Steinberger, M.; Carrasco, F","cited_arxiv_id":null,"evidence_quote":"Provides the TamingGS codebase and budget-based training setup that the paper builds on and compares against."}],"review_version":2}