acceptodds
Under review as a conference paper at ICLR 2027

EffGS: Efficient and High-Fidelity Gaussian Splatting

Abstract

3D Gaussian Splatting (3DGS) enables real-time novel view synthesis, but existing general-purpose acceleration methods suffer severe rendering quality degradation when extended to more complex, large-scale scenes. To address this issue, we propose EffGS, a more general acceleration framework that improves training and rendering efficiency while maintaining reconstruction quality comparable to or better than vanilla 3DGS across bounded and large-scale scenes. EffGS combines frequency-aware guidance, localized density control, and adaptive primitive scale modulation. First, an importance scoring mechanism combines pixel-wise reconstruction errors with a difference-of-Gaussians mask scheduled over training to provide stage-dependent spatial guidance. Second, localized densification and pruning restricts density modifications to Gaussians with valid projected footprints in the sampled views. Third, learnable per-Gaussian scale modulation adjusts effective primitive extent during optimization while retaining the Compact Box rasterization rule. Extensive experiments on bounded and large-scale scene datasets demonstrate a favorable balance between reconstruction quality, training time, and primitive count. Component ablations and matched-primitive-budget comparisons further support the effectiveness of the framework.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
3DGS

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_mvRKOpenReview ↗
Rating
3(+1)
Confidence
4(+1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

EffGS proposes three mechanisms: frequency-aware guidance, local density control, and adaptive elementary scale modulation, which can be extended to city-level scenarios. EffGS highlights the specific image structure through density control combined with spatial details at different stages of training, modifies the density control by distinguishing primitives beyond the sampling coverage and primitives with effective projected footprints, and changes the spatial support area by adjusting the Gaussian effective scale in the optimization process. EffGS achieves high reconstruction quality, efficient training and real-time rendering performance through these three mechanisms.

Critical strengths

EffGS proposes a universal acceleration framework, which not only has comparable rendering quality and training efficiency in small scenes, but also has good performance in large scenes.

Critical weaknesses
  1. The EffGS components are partially coupled, and the existing ablation experiments cannot effectively isolate the actual contributions of each component.
  2. The author should add the comparison of FastGS-big, and also add FastGS in the same budget experiment for comparison.
  3. The image on the right side of Figure 2 is ambiguous. According to the definition in the appendix, it is a high-frequency corresponding difference diagram. The content illustrated in the figure is that effgs has greater high-frequency error, which is contrary to the author's proposition.
  4. The author claims that the experimental configuration is consistent with 3DGS, but the data in Table 1 are basically from FastGS papers.
  5. The Gaussian number of EffGS in large scenes is significantly higher than that of FastGS, while it is close to that of FastGS in Mip-NeRF 360, Deep Blending and Tanks & Temples datasets. Does this mean that EffGS has been targeted optimized in large scenes?
  6. Both γ\gamma and SS are learnable parameters, but the regularization term $R_\gamma $ can change when S~\tilde{S} is constant.
Questions for authors

See Weaknesses.

Reviewer_g4C4OpenReview ↗
Rating
3
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a method named EffGS that improve the reconstruction quality and efficiency of 3D Gaussian Splatting (3DGS), particularly for large-scale scenes. This paper introduces three components: (1) frequency-aware importance scoring, which combines reconstruction errors with scheduled Difference-of-Gaussians (DoG) masks to guide Gaussian allocation; (2) Localized Densification and Pruning (LDP), which restricts density control to Gaussians covered by sampled training views; and (3) Adaptive Primitive Compactness, which introduces a learnable scale modulation factor for each Gaussian to adjust its effective spatial extent. Experiments on bounded and large-scale datasets demonstrate improved reconstruction quality compared to several existing acceleration methods while maintaining relatively efficient training and rendering.

Critical strengths

The proposed components address different aspects of Gaussian optimization, including importance scoring, density control, and primitive compactness. Compared with FastGS, EffGS enables more precise Gaussian selection for better control over densification and pruning during training. The method achieves better reconstruction quality than FastGS, particularly on large-scale scenes. The experiments cover multiple datasets and include component ablations and matched-Gaussian-budget comparisons.

Critical weaknesses
  1. The original Gaussian scales are already learnable, so introducing an additional scale factor does not increase representation capacity. Since the method also modifies regularization and pruning, it is unclear which part contributes most to the improvement.
  2. Although EffGS improves reconstruction quality, it requires longer training time and achieves lower rendering FPS than FastGS.
Questions for authors

Please refer to "Critical Weaknesses".

Reviewer_UntyOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses reconstruction quality degradation when Gaussian Splatting acceleration methods are applied to large-scale scenes. Building upon FastGS, EffGS combines a scheduled DoG mask, LDP, and learnable per-Gaussian scale modulation. Experiments on bounded and city-scale datasets evaluate reconstruction quality, training efficiency, and rendering speed, supplemented by component ablations and matched-Gaussian-budget comparisons.

Critical strengths
  1. The paper addresses a practical limitation and reports substantial quality improvements on large-scale scenes. Aggressive Gaussian reduction can compromise reconstruction quality in large environments. On Mill-19, EffGS improves PSNR from 22.32 to 23.68 and reduces LPIPS from 0.391 to 0.279 compared with standard FastGS. These results motivate further investigation of density control across different scene scales.

  2. The evaluation covers a broad range of scenes. Experiments on six datasets, together with component ablations, matched-Gaussian-budget comparisons, and frequency-specific error analysis, provide useful evidence for the effectiveness of the complete framework.

Critical weaknesses
  1. The methodological distinction from prior work is insufficiently established. Frequency-based guidance, scale reparameterization, and learnable scale-based masking have been explored in FreGS, SAD-GS, FDS-GS, and Compact 3D Gaussian Representation for Radiance Field, while EffGS retains FastGS’s Compact Box rule. A clearer explanation and validation are needed to establish how the proposed masking schedule, continuous scale modulation, and active-set restriction differ from existing approaches and what additional benefits they provide.

  2. The quality–efficiency advantage is not sufficiently established under comparable conditions. EffGS achieves higher reconstruction quality than standard FastGS but uses more Gaussians and longer training time. For example, on Mill-19, the Gaussian count increases from 1.60M to 3.76M and training time increases from 12 to 20 minutes, while rendering speed decreases from 251 to 145 FPS. These results alone do not distinguish whether the proposed method improves quality more efficiently or achieves higher quality by using additional representation capacity and computation.The matched-budget experiments in the supplementary material partially address this concern, but their limited baseline coverage does not directly establish an advantage over the underlying FastGS method. A controlled comparison is needed in which EffGS’s Gaussian count is reduced to the lower count produced by standard FastGS on each scene, while keeping the FastGS configuration unchanged. Reconstruction quality, training time, and rendering speed should be reported to determine whether EffGS retains its advantage after reducing its representation budget.

Questions for authors
  1. Density-control implementation: Does the implementation use Eq. (39) as the sole densification eligibility criterion, or additionally retain the original positional-gradient thresholds? Are the original scale-based pruning conditions also retained beyond the rules in Sec. 3.3? Controlled ablations with and without these inherited criteria are needed, reporting reconstruction quality, Gaussian counts, and training time. This would distinguish the contribution of the proposed scoring mechanism from that of the existing density-control rules.

  2. Scale modulation: Does a decrease in i produce smaller effective Gaussians during training, or is it compensated by increasing base scales? Joint statistics of i, base scales, and effective projected scales would clarify this interaction. Separate evaluation of modulation, regularization, and modulation-based pruning is needed, together with a comparison against direct effective-scale regularization. The spatial distribution and effective sizes of Gaussians removed by i<0.01 would also clarify what this criterion selects.

  3. LDP mechanism: Does LDP provide greater benefits for Gaussians observed by fewer training views? Spatial visualizations and coverage-conditioned statistics of added and removed Gaussians, with and without LDP, would directly test this motivation. Active-set restriction and valid-view normalization should be evaluated separately, with Gaussian counts controlled where possible.

  4. Hyperparameters: Are the proposed parameters fixed across scenes and datasets, or tuned separately? A complete configuration table and the parameter-selection procedure are needed. Sensitivity analysis should cover all introduced hyperparameters and their main interactions, evaluating effects on reconstruction quality, Gaussian counts, training time, and memory usage. This is necessary to assess robustness to parameter choices and the extent of scene-specific tuning.

  5. Baseline fairness: Why is FastGS-Big excluded? Comparisons with FastGS-Big and standard FastGS under comparable training-time or Gaussian-count budgets are needed to establish the quality–efficiency advantage. The existing matched-budget experiments do not resolve this comparison.

  6. Frequency guidance: What exactly is removed in the frequency-aware and error-aware mask ablations in Table 3? The composite mask affects both densification and pruning, while Mfreq also weights the reconstruction loss. The ablations should specify which pathways are disabled and distinguish their contributions.

  7. Computational scalability: How do sampled-view scoring time, active-set construction time, and peak memory change as Gaussian and training-view counts increase? The corresponding training costs and GPU configuration of EffGS-GPUs are also needed. Comparing EffGS and FastGS within the same spatial partitioning framework would help distinguish the proposed method’s benefits from those of partitioning and additional computational resources.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.