acceptodds
Withdrawn from ICLR 2027

EchoGS: Compact 3D Gaussian Splatting via Conditional Response Credit

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Compact 3D Gaussian Splatting (3DGS) methods commonly score each Gaussian independently or merge primitives according to parameter similarity. Neither strategy answers the basic compression question: which rendered effects remain irreplaceable after nearby substitutes are considered? We introduce EchoGS, a three-stage compression pipeline centered on conditional response credit. Stage I collects demand to identify where reconstruction capacity is needed. The core Stage II encodes each Gaussian's multi-view rendering behavior in a compact, view-tagged signature and jointly compares signatures within spatial neighborhoods. Local ridge leverage measures each primitive's contribution to response coverage given its neighbors: repeated directions share credit, while sufficiently supported complementary directions retain value. Blending it with marginal demand produces the allocation score passed to Stage III, which distributes the target budget and fits replacement Gaussians by matching assigned geometric moments. In the reported comparison with vanilla 3DGS, the default configuration achieves 9.01×–11.58× count compression on three datasets with competitive rendering quality.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Scene Compression, Novel View Synthesis, Conditional Response Credit, Ridge Leverage Scores

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_8okaOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes a three-stage pipeline to reduce the number of Gaussians in a 3DGS scene with limited quality loss. Stage I establishes residual demand, Stage II builds per-Gaussian, per-view features and compresses them using projection to give credit to Gaussians whose rendering is not already covered by their neighbours, Stage III finally merges the Gaussians per block into fewer Gaussians by matching means and covariances. The paper reports roughly 9 to 12 times fewer Gaussians than vanilla 3DGS with short training times and comparable image quality on Mip-NeRF 360, Tanks & Temples, Deep Blending and Waymo.

Critical strengths
  1. According to the reported results, the method reaches roughly 9 to 12 times fewer Gaussians than 3DGS with roughly comparable image quality.
  2. The insight to only keep Gaussians whose rendering is not covered by neighbors is a good improvements over the common per-Gaussian importance scores.
  3. The choice of datasets is good and the reported metrics are relevant: per-scene results, peak memory, rendering speed, several hyperparameter sweeps.
Critical weaknesses
  • The method is very difficult to follow. Even for a reader familiar with the area, the methodology, and Stage II in particular, is very hard to follow. Steps are introduced without explaining what they do or why they are needed, the same notation are used for multiple meanings, and many symbols and terms are never defined. The authors should fix this and report the intuition and reasoning behind each design decision rather than only listing the steps and the math to improve readability. Specific examples:

    • Introduction example. The example with responses (1,0),(1,0) versus (1,0),(0,1) relies on "responses" and "rows" before either is defined. It only becomes understandable after Section 3.4.
    • h(t) = min{8, log(1 + t/a_t)}. This appears to be a log compression with a cap, presumably to stop large values from dominating, but no reason is given. The scale a_t is never defined, and it determines where each feature saturates. Is it a constant, a per-scene statistic or per-view? What is its value for each of the three features it is applied to?
    • Per-camera projections (Eq. 4). D_v and N_RMS are not defined. It is unclear whether N_RMS normalizes each vector or each coordinate across Gaussians. Proposition 2 holds only in expectation and before the final normalization, and the error at d = 24 is not discussed. It is also not argued why the 100-dimensional per-view descriptor needs to be compressed, since it seems affordable, and no comparison against using it directly is shown.
    • Other undefined symbols and terms. Ω_v and u_i^v in Eq. 2, the "flagged rasterization", how c_i is accumulated, "demand interface", "response coverage" (used before Eq. 9 defines it), "candidate set", "renderer support", "span novelty", "caps" and "seeded subset" in Appendix B.4, and "the archived demand/reduction base" in Section 4.2.2.
    • Overloaded notation. Many symbols carry several meanings, for example:
      • s_i is both the sketch in Eq. 4 and the allocation score in Eq. 12.
      • a is the alpha support a_i^v, the scale a_t, the block median mass a_B, and a row label in Figure 1. It is also easily confused with opacity α_i.
      • x is the per-view feature vector and, inside that same vector, the image x-coordinate.
      • e_i is the demand count in Eq. 2 and a unit vector in Proposition 1.
      • ℓ is the leverage score in Eq. 8 and the luminance in Eq. 21.
      • N / 𝒩 denotes the number of Gaussians, min-max normalization, the normal distribution 𝒩(0, 1/d) on the same page, RMS normalization and a Gaussian density.
    • Figure 1. The figure is unclear and uses terminology that is not used or defined in the text: "distinct response", "response energy", "Redundancy Mixture" and the High/Low priority scale. The meaning of the a/b/c rows and the check mark in Stage II-B is not explained, nor is "absorbed sources". The figure shows at least 24 views ("view 24") while the method samples 10 cameras.
  • The central contribution is not shown to matter. The paper spends most of its explanation on Stage II, but its own ablations show almost no effect on quality: Demand-only allocation (β = 0, Table 8): PSNR: 28.87; #GS: 0.29M Full method (β = 0.25): PSNR: 28.88; #GS: 0.26M The full method reaches the same quality with about 10% fewer Gaussians, but because quality and count change together and no variance is reported, it is not possible to tell whether this difference is real. Can you compare scoring rules at the same final Gaussian count? Given how close the results are, we would also recommend reporting results using multiple seeds with standard deviation. A quality-versus-count curve over several budgets would be even more informative.

  • Why should residual demand guide compression? Stage I counts pixels with high reconstruction error around each Gaussian. This is a known heuristic for densification, where it makes sense: high error indicates that more capacity is needed. It is much less clear to us why it should decide how much capacity each region keeps during compression. A residual measures how well the current Gaussians fit the training images, not whether they are redundant. A region with low error and one thin Gaussian per structure can be irreplaceable yet receive a small budget, while a high-error region crowded with badly placed Gaussians is protected. Table 8 shows that demand alone (β = 0) performs as well as the full method, so this unexplained heuristic carries most of the allocation. Please justify this choice.

  • The Stage II features are not ablated. The ten-dimensional per-view feature in Eq. 3 combines visibility, alpha support, demand, footprint, image position, depth and colour, but the paper never shows what each contributes. Table 5 only replaces the whole feature set with Gaussian parameters. Please report leave-one-group-out results or results where you only use the specific feature.

  • Evaluation of the work.

    • Scene subset: Mip-NeRF 360 is evaluated on 7 of its 9 scenes. Most baselines report all nine in their own papers, so the numbers cannot be compared with the literature. Please report the full benchmark.
    • Variance.: No result reports seeds or variance. Several of the differences discussed in the text are 0.02 to 0.08 dB, which is within typical run-to-run variation for 3DGS.
  • Minor issues

    • Typo: "creadit" in the contributions list.
    • The method samples only 10 training cameras (Table 7), and this choice is not justified or tested. Mip-NeRF 360 scenes have 150 to 300 images, and Waymo sequences are much more spread out. It might be that we misunderstand the contents of this table though, so additional explanation would be nice.

Using leverage scores to account for redundancy among neighbouring Gaussians is a sensible idea, and the reported reduction in Gaussian count is substantial. However, the paper's own ablations do not show that the claimed central contribution, Stage II, has a significant effect. Most of the allocation is carried by an error heuristic whose role in compression is not explained. The evaluation uses a non-standard scene subset. The methodology is very difficult to follow. A fixed-count comparison with seeds, a complete description of the pipeline, the full benchmark and a substantially clearer presentation would be needed for acceptance.

Questions for authors

The quality of the splat reduction could be better assessed if the authors provided the full scenes as supplementary material. I don't have specific additional questions for the authors. I have detailed the clarifications that are needed above to improve the quality of the work.

Reviewer_Y8EdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces EchoGS, a compact framework for 3D Gaussian Splatting (3DGS) that adaptively allocates Gaussians during optimization based on whether each Gaussian can be replaced by its neighbors. Specifically, it proposes Conditional Response Credit (CRC), which scores each Gaussian based on how much of its multi-view rendering response is not covered by neighboring Gaussians. To this end, it first measures how much each Gaussian contributes to regions with high reconstruction error (Stage 1), and then summarizes each Gaussian's responses across multiple views into a compact signature (Stage 2-A). Within local spatial blocks, it computes a ridge leverage score from these signatures, which becomes low when nearby Gaussians already capture similar rendering behavior (Stage 2-B), and this score is then combined with the demand from Stage 1 to obtain the final allocation score (Stage 2-C). Finally, the resulting score is used to allocate the target number of Gaussians to each block, where nearby Gaussians are merged by moment matching (Stage 3). Experimental results demonstrate that the variants of EchoGS achieve comparable (even higher) rendering quality with fewer Gaussians than the compared existing compact 3DGS methods.

Critical strengths
  • This paper addresses a new aspect that has not been explored in existing 3DGS compaction methods, and proposes a novel solution to address it.

    • While existing methods focus on whether removing a Gaussian has little impact on the rendering quality, the proposed method goes further and defines a score that measures whether each Gaussian can be covered by its neighboring Gaussians.
    • Beyond proposing this perspective, it suggests a novel framework that reduces the number of Gaussians by defining CRC and using it to determine the allocation budget, achieving higher efficiency than existing 3DGS compaction methods.
  • The experimental validation is extensive and comprehensive.

    • The paper conducts experiments on various datasets for NVS, and includes various 3DGS pruning and merging methods as baselines.
    • The ablation studies are conducted across most of the proposed components, sufficiently verifying the effect of each component.
Critical weaknesses
  • Although the idea of CRC is impressive, the validation shows that it does not lead to a meaningful compression performance gain. The main contribution of this paper is to reduce the number of Gaussians according to whether each Gaussian can be replaced by its neighbors. However, the ablation results show that its impact on rendering quality is incremental.

    • Tab. 8 directly compares between demand-only allocation (β=0\beta = 0) with the full CRC-based allocation (β=0.25\beta = 0.25). Simply using the demand instead of CRC achieves a similar level of efficiency (28.87 dB with 0.29M vs. 28.88 dB with 0.26M). This indicates that the effect of CRC, which is the core claim of this paper, seems marginal.
    • Moreover, Tab. 5 indicates that using conditional-only allocation (β=1\beta = 1) even degrades the performance. Therefore, it remains unclear whether CRC provides a significant contribution, as most of the performance gains appear to come from the blended demand.
  • Also, CRC is not employed in the merging step.

    • CRC is used only to determine the number of Gaussians allocated to each block of the tree, and the merging is performed solely based on geometric similarity without considering it. Therefore, it is unclear whether Gaussians identified as replaceable by CRC are actually merged.
Questions for authors
  • The CRC is not used in the merging process. Is there any validation that Gaussians with low CRC scores are actually merged? Also, is there a reason why CRC is not employed in the merging step?

  • As discussed in the weaknesses, the proposed framework already achieves high efficiency without CRC, and using CRC does not seem to provide a meaningful compression performance gain. Is this result general, or is it specific to the current experimental setting? A comparison on other datasets or other variants (Turbo and Quality) would be helpful.

  • Clarifying notations would improve the readability of the paper. Several symbols are used for different quantities, which makes the method difficult to follow.

    • For example, sis_i is used in both Eq. 4 and Eq. 12, and cic_i is used in both Eq. 3 and Eqs. 6 & 13.
    • In addition, Eq. 3 lacks a sufficient description of each element. It is unclear how the visibility, the accumulated alpha support, and the projected footprint are obtained.
Reviewer_tu35OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper targets compact 3DGS: rather than scoring each Gaussian in isolation or merging by parameter similarity, it asks whether a primitive's rendered response remains necessary given its neighbors. EchoGS is a three-stage pipeline. Stage I flags pixels where photometric and structural residuals coincide and accumulates per-view demand counts. Stage II encodes visibility, alpha support, footprint, demand, position, depth, and color into 10-D per-view features, sketches them with camera-specific random projections (d=24) into view-tagged signatures, and computes within importance-balanced KD-tree blocks a ridge leverage score whose derivative equals the log-det response coverage; the score equals the exact single-row removal loss (Prop. 1), decays under added substitutes (Prop. 3), and is blended with marginal demand. Stage III allocates the target budget across blocks and fits replacement Gaussians by matching first and second geometric moments (shown to be a forward-KL Gaussian projection), followed by continued training. Across Mip-NeRF 360, Tanks & Temples, Deep Blending, and Waymo, the Balanced preset reports 9.01x-11.58x Gaussian-count reduction versus vanilla 3DGS with +0.46/+0.33/-0.18 dB PSNR changes, and is competitive with strong recent baselines including FastGS, Taming 3DGS, Mini-Splatting, Speedy-Splat, and GHAP. Ablations cover every component plus hyperparameter sensitivity, runtime, memory, and per-scene results.

Critical strengths
  1. A principled, set-aware compression criterion. The reformulation from marginal importance to conditional response coverage, with the exact identity linking the leverage score to the single-row removal loss -log(1-l_i) (Prop. 1) and the diminishing-return behavior under duplicated responses (Prop. 3), is clean, correct, and directly actionable as a budget signal. Theorem 1 crisply shows why per-primitive scalarization is fundamentally ambiguous. This is the paper's core intellectual contribution and it is consequential: it is what distinguishes EchoGS from FastGS/QIRF-style scoring.

  2. Strong and broad empirical support. Evaluations span 4 datasets and 10 baselines, and the Balanced preset Pareto-dominates FastGS on Deep Blending (0.21M vs 0.32M Gaussians, 30.20 vs 30.04 dB, 391 vs 359 FPS), Tanks & Temples, and Waymo, while trading only 0.03 dB PSNR for a 38% smaller count on Mip-NeRF 360. Every component is ablated, and hyperparameter sensitivity, complexity, memory, and per-scene tables are provided. This breadth materially raises confidence that the gains are real.

Critical weaknesses
  1. The "compact" claim is measured in Gaussian counts, not storage. The paper reports count compression (9x-11.6x) but no actual model size in MB or bits per Gaussian, and does not compare against quantization/entropy-coding methods (e.g., HAC) at matched quality. Appearance (SH) and opacity are inherited from a single representative source, so count reduction need not translate into storage reduction, and view-dependent appearance may degrade. This is consequential because storage is the stated motivation of the compact-3DGS line of work.

  2. Attribution of the headline gains is confounded. Stage III continues training from iteration 15,000 after reduction (Alg. 2), and the ablations are explicitly quality-efficiency trade-offs rather than comparisons at a fixed count. Notably, removing conditional leverage costs only 0.02 dB PSNR (Tab. 5), so the isolated benefit of the titular mechanism appears small, with count and LPIPS doing more of the work. Without fixed-count or no-fine-tuning controls, it is hard to verify that conditional credit — rather than the extra fine-tuning budget plus demand gating — drives the reported wins.

Questions for authors
  1. Can you report actual storage (MB) after standard quantization/entropy coding, or at least bytes per Gaussian, for EchoGS and the strongest baselines at matched PSNR?
  2. How much of the final quality comes from post-reduction fine-tuning (from iteration 15,000) versus the allocation itself? Is there a no-fine-tuning result, and do baselines receive a comparable fine-tuning budget?
  3. Can you provide ablations at a fixed final Gaussian count (e.g., retuning the target fraction per variant so all rows end at the same N_GS), particularly for "w/o conditional leverage"?
  4. Balanced uses more peak GPU memory than FastGS/Taming (e.g., 6.03 vs 4.71 GB on Mip-NeRF 360, Tab. 13). Where does this overhead come from, and can it be reduced?

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

Related papers

Discussion (0)

Sign in to comment.