acceptodds
Under review as a conference paper at ICLR 2027

Compression-aware edge densification for compact 3d gaussian splatting

Abstract

Improving novel-view synthesis under a fixed storage budget requires deciding not only which Gaussians to retain, but also how to create them. While compression pipelines emphasize pruning, the representation built before pruning offers another opportunity to improve compact reconstruction. We introduce Compression-Aware Edge Densification (CAED), which controls early Gaussian growth for a fixed-capacity output. CAED adapts edge-guided selection and long- axis splitting from ImprovedGS, limits growth with a square-root budget, and uses the reconstruction-aware pruning and Difference-of-Gaussians refinement of PWRS to produce a compact model. With exactly the same final Gaussian count and PLY size as PWRS in every scene, CAED improves mean PSNR by 0.2384 dB and SSIM by 0.0045, and reduces LPIPS by 0.0049 across 13 scenes. LPIPS improves in 12 of the 13 scenes. Three-seed experiments on five scenes further support the overall quality gains, with an observed reduction of approximately 21% in training-command wall-clock time within the same framework. These results show that designing early densification for a compact output can improve reconstruction without increasing final model capacity.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Novel View Synthesis, Compact Scene Representation, Budget-Constrained Densification, Reconstruction-Aware Pruning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_g9ayOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies novel-view synthesis for 3D Gaussian Splatting under a fixed storage budget. The core claim is that models with the same final Gaussian count need not follow the same growth trajectory — how the representation grows before pruning also affects the quality of the compressed model. The authors propose CAED, which combines an ImprovedGS-style frontend (edge-guided selection, long-axis splitting, and a square-root growth budget) with the PWRS compression backend (reconstruction-aware RPS/SPS pruning and DoG refinement), configuring the training-time growth budget B(t) separately from the final target capacity N_tar. Three-seed experiments on five scenes support the overall conclusion; training wall-clock time decreases by roughly 16%–21%, while rendering FPS slightly drops and rendering memory slightly increases.

Critical strengths
  1. The matched-capacity experimental design is the strongest part of the paper. CAED and PWRS share identical Gaussian counts and PLY sizes in every scene (Table 1), cleanly isolating "early densification" as the single variable of interest and avoiding the apples-to-oranges capacity comparisons common in the compression literature. In Figure 1 the two methods share identical horizontal coordinates, making the fixed-capacity comparison explicit and faithful to the stated design intent.

  2. The evaluation is honest and systematic. Paired t-tests are reported with 95% confidence intervals and p-values; three-seed experiments are included (Tables 6/10); ablations cover frontend components (Tables 4/5/14), schedule shape and budget amplitude (Tables 18/12), DoG interaction (Tables 15/16), and multiple capacity levels (Table 17). The appendix candidly discloses the three timing conventions (W/I/C), run provenance, and the fact that historical main runs did not preserve complete optimization arguments — a level of transparency uncommon in conference submissions.

Critical weaknesses
  1. Limited novelty: the method is a combination of existing modules, and the headline claim is not supported by the ablations. All three frontend operators (edge score, long-axis split, square-root budget) come from ImprovedGS (acknowledged by the authors in the Figure 2 note), and the backend (RPS/SPS/DoG) comes entirely from PWRS. No new mechanism is proposed; the contribution is largely the empirical finding that "ImprovedGS-style densification in front of PWRS compression" helps. More importantly: (a) "compression-aware" in practice only means that the growth budget and the final target are configured separately — B(t) in Eq. (4) does not depend on N_tar at all, so the budget is not actually aware of the compression target; (b) the "Edge Densification" in the title is contradicted by Table 14: CAED vs. Gradient-LAS (which removes EAS) shows five-scene mean differences of ΔPSNR = −0.015 dB, ΔSSIM = +0.0001, ΔLPIPS = −0.0001 with a seed-mean std of 0.076 dB, and the paper itself concedes that EAS is "a sampling design choice rather than a consistently superior scoring signal." The component that names the paper contributes close to no measurable benefit, and this tension needs to be addressed head-on in the main text.

  2. Attribution is not clean, effect sizes are small, and results are scene-dependent. The overall +0.2384 dB gain is driven by Mip-NeRF 360 and Tanks & Temples; on Deep Blending CAED is slightly worse, and in Table 17 playroom degrades by −0.2865 dB PSNR at the larger capacity. Table 4 shows that the budgeted frontend alone (B-G-I) already contributes most of the LPIPS improvement over PW ADC, while the EAS/LAS increments are small and inconsistent in direction. In addition, the abstract highlights quality gains and reduced training time but omits the rendering-side costs: FPS drops by 22.8 and peak memory rises by ~96 MiB — an important caveat for a "compact model" narrative.

  3. Hyperparameter selection and statistical conventions raise soundness concerns. The appendix ("Randomness and coverage") admits that the historical per-scene selection of B_max "does not establish a tuning protocol entirely independent of test information." Since B_max varies per scene between 1.0M and 2.5M (Table 7), if these budgets were chosen with test data in the loop, the 13-scene gains may be subject to selection bias. Separately, the main text claims PSNR/SSIM/LPIPS improve in 11/13, 11/13, and 12/13 scenes, respectively, but per the appendix Table 8 values, PSNR decreases in four scenes (counter, bonsai, train, playroom) and SSIM decreases in three (bonsai, train, playroom), which does not reconcile with the stated counts — the authors should verify.

Questions for authors
  1. Please reconcile Table 8 with the claimed "11/13, 11/13, and 12/13 scenes improve": per the appendix values, PSNR decreases on counter, bonsai, train, and playroom, and SSIM decreases on bonsai, train, and playroom. How do the counts align?

  2. Table 14 shows that EAS yields essentially zero average gain over gradient-based sampling (with large variance): how much does EAS contribute to the 13-scene main results? Can you provide the 13-scene result for "budget + LAS without EAS" to support the "Edge" claim in the title?

  3. B_max is selected per scene with no evidence that test information was excluded. Could you re-run the main experiments with B_max set as a function of N_tar (e.g., α·N_tar) and report sensitivity? If the growth budget is fully decoupled from the compression target, is "compression-aware" too strong a term?

  4. In the DoG interaction ablation (Table 16), the interaction term is +0.64 dB on playroom but −0.15 dB on bicycle. How do you explain this scene-level reversal, and how much do the main results depend on DoG?

Reviewer_r15eOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Compression-Aware Edge Densification (CAED), a framework that improves compact 3D Gaussian Splatting by controlling the early densification process under a fixed final Gaussian budget. Building upon ImprovedGS and PWRS, the method combines edge-guided Gaussian selection, long-axis splitting, and a square-root growth budget with reconstruction-aware pruning and Difference-of-Gaussians refinement. By controlling early Gaussian growth while maintaining the same final model capacity, CAED aims to improve reconstruction quality without increasing storage requirements. Experiments on 13 scenes from Mip-NeRF 360, Tanks & Temples, and Deep Blending demonstrate improved overall reconstruction quality compared with PWRS at identical final Gaussian counts and model sizes.

Critical strengths
  1. The paper adapts the edge-guided sampling strategy of ImprovedGS by modifying the edge-response computation and candidate selection mechanism. These modifications provide a practical way to guide Gaussian allocation during early densification under a fixed final-capacity constraint.
Critical weaknesses
  1. The proposed framework primarily integrates existing techniques from ImprovedGS and PWRS, including edge-guided sampling, long-axis splitting, growth budgeting, and reconstruction-aware pruning. Although Section 3.2 introduces modifications to the edge-response computation and candidate selection, the overall method remains largely built upon existing components, with limited algorithmic innovation beyond their adaptation and integration.

  2. The ablation results in Table 4 do not clearly establish the benefit of the proposed edge-aware sampling strategy. In particular, the Budgeted-Gradient-LAS variant achieves slightly better PSNR and SSIM than the full CAED model, while obtaining the same LPIPS. Moreover, the three-seed experiments show only marginal differences between CAED and Gradient-LAS, suggesting that edge-aware sampling does not provide a consistent advantage. These results raise questions about the necessity and effectiveness of the proposed edge-guided component within the overall framework.

Questions for authors
  1. The main methodological novelty of this paper appears to lie in the edge-guided parent sampling introduced in Section 3.2. However, this component itself is closely related to the edge-aware selection strategy of ImprovedGS, while the remaining components are largely inherited from ImprovedGS and PWRS. Could the authors further clarify the methodological novelty of CAED beyond the adaptation and integration of these existing techniques?
  2. Please refer to the second point in the Weaknesses. The current ablation results do not clearly demonstrate the benefit of the proposed edge-aware sampling strategy. Could the authors provide further analysis showing when and why this component improves over the simpler Gradient-LAS variant?
Reviewer_6VYoOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies early densification for compact 3D Gaussian Splatting. CAED combines edge-guided sampling, long-axis splitting and a square-root growth budget adapted from ImprovedGS with PWRS pruning, recovery and DoG refinement. At matched per-scene final Gaussian counts and actual PLY sizes, it reports a mean PSNR improvement of 0.2384 dB over PWRS across 13 scenes. This controlled quality comparison is useful, and component, budget, backend and multi-seed experiments provide additional evidence. The paper explicitly attributes inherited operators and acknowledges scene-dependent gains and rendering trade-offs.

Critical strengths
  1. Matched-capacity protocol is a genuine control.
  2. "Compression-aware" is operationalized, not just claimed.
Critical weaknesses
  1. The independent methodological contribution is insufficiently established. CAED adapts ImprovedGS sampling, long-axis splitting and square-root growth control, and combines them with PWRS pruning, recovery and DoG refinement. The matched-capacity improvement is useful, but the frontend is not directly conditioned on the final target or post-pruning utility; the authors define “compression-aware” through the training configuration. Please explain the new insight beyond the adapted combination and isolate the local scoring/gating changes from the original operators. Table 4 gives Gradient-LAS slightly higher PSNR/SSIM than CAED, with equal LPIPS at the displayed precision. Table 14 already provides paired three-seed results, but the mean PSNR gain of 0.00897 dB is small relative to the reported seed-mean standard deviation of 0.076195 dB. These results support treating EAS as a design choice, as the authors do, rather than establishing a reliable extra benefit within the full LAS frontend.
  2. Comparison with relevant allocation and simplification strategies is incomplete. GDAGS (Zhou et al., 2026) and GaussianSpa (Zhang et al., 2025) are cited but absent from the quantitative comparisons; CDC-GS (Dong et al., 2025) is not discussed or evaluated. CDC-GS is relevant because it uses visual complexity and primitive density to guide densification and pruning. A focused comparison with representative alternatives at a common output capacity or storage budget, and a shared-pruning-backend comparison of frontends, would help establish the allocation contribution. This need not be an exhaustive new benchmark.
  3. A timing comparison needs a concrete presentation correction. Table 2 marks CAED’s wall-clock training times as best and ImprovedGS’s internal times as second-best, although its footnote acknowledges different timing boundaries. Please remove the cross-protocol ranking or use a common measurement boundary. The small labels in Figures 1-2 could also be enlarged. These are limited presentation issues and do not determine my recommendation.
Questions for authors
  1. What is the mechanism by which early allocation survives pruning? The claim is that how Gaussians are created affects what survives PWRS. Mechanistically, does edge-guided splitting place primitives where SPS is more likely to retain them, or does it produce children whose long-axis geometry is simply better fitted before pruning? Is there any evidence (e.g., retention rate of CAED-born vs. PWRS-born Gaussians) that distinguishes these?

  2. Why should edge response be the right allocation signal? Eq. (2) scores Gaussians by squared response to an edge-weighted RGB objective. Is the mechanism claim that reconstruction error concentrates at edges, or that edges are where splitting geometry matters most? These imply different expected failure modes (textureless regions vs. thin structures), and the Deep Blending result may be informative here.

  3. Does the budget schedule interact with the scoring signal? Eq. (5) truncates sampling when B(t)−N_t is small. Early in training the budget is tight, so only the highest e_i candidates are split. Does this mean edge scoring matters mainly early, and long-axis geometry mainly late? Can the two contributions be separated by time?

  4. What is the mechanism-level difference from GDAGS? GDAGS also uses gradient-direction/consistency signals for density control. Is CAED's contribution the edge-weighted response q_{vi} (Eq. 2), the budget-coupled sampling (Eq. 5), or the combination with PWRS? A short mechanistic contrast would clarify what is new versus what is a re-parameterization of existing density-control signals.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.