acceptodds
Under review as a conference paper at ICLR 2027

General Post-Training 3DGS Compression: Coordinating Quantization and Search for Quality-Constrained Size Minimization

10 views

Abstract

The large file sizes of trained 3D Gaussian Splatting (3DGS) scenes increase storage and distribution costs. We propose a general post-training compression framework to reduce model size under rendering-quality constraints and a limited evaluation budget. Our key idea is to coordinate nonuniform quantization with configuration search. We apply signed-power companding to higher-order spherical harmonic coefficients within spatial blocks. This reduces quantization error and allows lower precision or larger blocks within the quality limits. The search jointly selects the pretrained base model, block size, and coefficient precision without retraining or fine-tuning. An independent view set checks the selected model against mean and tail PSNR-loss limits before delivery. Failed confirmation triggers a lossless reference fallback. Across 192 matched configurations, companding increases the number of candidates satisfying the search quality criteria from 75 to 119. At predetermined default settings on eight benchmark scenes, our complete system reduces mean model size by 97.37% and 79.60% compared with 3DGS and FastGS, respectively. Mean PSNR improves by 0.257 dB and 0.173 dB, respectively. The average rendering speed reaches 1,285.5 FPS.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Post-Training Compression, Nonuniform Quantization, Budgeted Configuration Search, Quality-Constrained Optimization

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_f47LOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a general post-training 3DGS compression framework that aims to minimize model storage under rendering-quality constraints and a limited evaluation budget, without per-scene fine-tuning. The framework combines spatial block-wise quantization with signed-power companding of higher-order SH coefficients to reduce quantization distortion and expand the set of feasible compression configurations. It further introduces a budgeted configuration search strategy based on Gaussian-process predictions to jointly select the pretrained base model, spatial block size, and SH precision while satisfying mean and tail PSNR-loss constraints. An independent confirmation stage evaluates the selected configuration on disjoint views, with a losslessly encoded reference used as a fallback when the quality constraints are not met. Experiments across multiple 3DGS benchmarks demonstrate reduced model sizes, competitive rendering quality, and high rendering throughput, while ablation studies examine the effects of companding, configuration search, and quality-aware delivery.

Critical strengths
  1. The proposed framework enables 3DGS compression without per-scene fine-tuning, making it suitable for compressing pretrained Gaussian representations. The reported compression time is approximately 58 seconds per scene, excluding base-model construction, while maintaining competitive rendering quality and high rendering throughput.

  2. The paper provides extensive experiments across multiple datasets, including controlled comparisons of uniform and companded quantization over 192 matched configurations. The ablation studies further investigate configuration search, quality constraints, and confirmation mechanisms, providing useful insights into the effectiveness and limitations of individual components.

Critical weaknesses
  1. The proposed framework primarily integrates established techniques rather than introducing fundamentally new compression algorithms. Specifically, the signed-power companding used for SH coefficients is a classical nonuniform quantization technique, while the budgeted configuration search builds upon Gaussian-process-based constrained optimization. Similarly, the independent quality confirmation and lossless fallback mechanisms are largely adaptations of conventional validation and quality-control strategies. Although combining these components into a post-training 3DGS compression framework is practically motivated, the methodological contributions appear largely incremental, and the paper does not sufficiently establish what is technically novel beyond adapting and integrating existing approaches.

  2. Although the paper presents the framework as a general post-training 3DGS compression method, all experiments rely on Gaussian representations produced by FastGS-based training pipeline, with different densification schedules. The proposed codec is not evaluated on representations obtained from other 3DGS training methods. Since the effectiveness of SH companding and the selected quantization configurations may depend on the attribute distributions of the pretrained models, the claimed generality remains insufficiently validated.

Questions for authors
  1. While companding improves quantization distortion and expands the set of feasible configurations, Table 1 shows that the full method achieves a larger average delivered size than Uniform Quantization with the same local-mean search strategy (16.856 MiB vs. 15.680 MiB). Could the authors explain this discrepancy and clarify under which conditions companding consistently improves the final compression performance?

  2. The authors present the proposed method as a general post-training 3DGS compression pipeline. However, all experiments are conducted on Gaussian representations obtained from the FastGS-based training pipeline. Could the authors evaluate the proposed method on 3DGS representations trained using other methods, such as the original 3DGS, to further validate its generality?

9 comments

1. W1: Are the technical contributions only an incremental combination?

We acknowledge that signed-power companding, Gaussian processes, and independent validation all have prior foundations. The technical contribution of this paper is the design of candidate construction and search decisions around the specific interactions in 3DGS compression.

First, the SH quantization design changes the feasible configurations available to the search. We apply companding to the non-DC SH coefficients within spatial blocks and combine it with per-band SH precision settings. This does more than reduce quantization error for a fixed configuration. It also allows some lower-precision configurations to satisfy the quality constraint, which in turn changes the compression setting ultimately selected by the search.

Our initial experiments provide a direct example. For hotdog with seed 19, using the same base model and spatial block size, uniform quantization selected SH band bit widths of 6/5/4, while companding was able to select 4/4/4. Across three seeds in this setting, the average delivered size decreased by about 6.89%, and average PSNR increased by 0.100 dB. This shows that quantization improvements can translate into actual compression gains through configuration selection.

Second, the search is designed around how quality responds to different base models. For each base model, we predict per-view losses separately and then compute mean and tail statistics. We also calibrate the size prediction using measured bitstream sizes for that base model. These predictions only determine which candidate to evaluate next. The final choice must be based on actual encoding, decoding, and rendering measurements. This design allows the search to account for the relationship among base-model differences, quantization error, and actual storage size.

Therefore, our intended contribution is to construct more valuable candidates for low-precision SH compression and to design limited-budget search and delivery rules that make use of them. We do not claim the existing components themselves as new algorithms.

Q2 : Why Companded Can Be Larger After Fallback

Thank you for the question. We clarify that Table 1 compares final delivered size, not the size of candidate bitstreams under the same configuration. The larger final size of Companded mainly comes from fallback after a confirmation failure, rather than from companding directly increasing the bitstream size.

1. Why Companded Has a Larger Final Size in Table 1

Our size calculation is as follows:

  1. Each candidate configuration is actually encoded into an .npz file, and the true file size in bytes is recorded rather than the theoretical entropy or an estimated size.

  2. If the candidate satisfies both a mean PSNR loss of no more than 1.0 dB and a CVaR10 tail loss of no more than 1.5 dB on the independent confirmation views, the final size is the actual bitstream size of that candidate.

  3. If confirmation fails, the entire delivery is replaced by the lossless bitstream of the reference model, rather than the delivery result from Uniform.

Therefore, an improvement in rate-distortion at the candidate level does not necessarily translate into a reduction in delivered file size. For an explanation of why a confirmation failure immediately triggers lossless reference delivery, please see our response here: OpenReview response to SEZk.

2. Common-Budget Supplementary Experiment and Conditions for Consistent Improvement

We further conducted a budget experiment using the same candidate pool, the same local_mean strategy, and the same quality thresholds: Exploration set: 20 matched-quality scene/seed pairs.Validation set: 8 scenes not included in the matched-quality cohort.Budgets: Q=5, 10, and 20.All selections were frozen before accessing the test set, and all methods were evaluated afterward.

Positive values indicate that Companded has a smaller final delivered size.

QExploration Set, 20 Pairs: Mean / MedianWins / Losses / TiesValidation Set, 8 Pairs: Mean / MedianWins / Losses / Ties
5+1.03% / -1.81%8 / 11 / 1-1.13% / -4.59%1 / 7 / 0
10+2.93% / +2.23%12 / 5 / 3+0.40% / +0.99%4 / 3 / 1
20-36.88% / +3.49%14 / 3 / 3-93.20% / +3.19%5 / 2 / 1

Based on the current experiments, consistent improvement requires at least the following conditions to hold simultaneously:

  1. The candidate-level savings of Companded must be sufficient to cover the additional fallback cost.
  2. The independent confirmation pass rate must be no worse than that of Uniform, especially without an increase in tail risk.
  3. The search budget cannot be too small, nor should the search focus only on more aggressive low-bitrate candidates.

In the current implementation, Q=5 underexplores the candidate space. Q=10 is the most suitable operating point on the exploration set, but the validation results remain mixed. Q=20 improves candidate-level performance, yet this improvement does not consistently translate into delivery-level gains. We have not yet identified a budget range that remains consistently stable on the validation set.

W2 + Q2: Does our method work on Gaussian representations from other training pipelines?

Thank you for the suggestion. We added experiments on original 3DGS, Mip-Splatting, Taming 3DGS, and 3DGS-MCMC to test our key designs beyond FastGS. These methods differ in densification, pruning, and training. These differences affect the number of Gaussians and the distributions of SH coefficients, scale, and opacity. Our method quantizes pretrained Gaussian representations and selects compression configurations. It requires neither the FastGS training pipeline nor per-scene fine-tuning. The new experiments directly test whether our designs work with these different attribute distributions.

Configuration selection provides benefits across all four upstream methods. We compare GP/local-mean selection with random selection under the same evaluation budget. Both use SH companded quantization. For the scenes listed in the table, GP/local-mean reduces mean model size by 5.01%–7.44% across the reported groups. The mean PSNR decrease is at most 0.0125 dB in each group. Original 3DGS, specifically requested by the reviewer, achieves size reductions of 6.14%–7.40%. Thus, the benefits of configuration selection extend beyond FastGS.

SH companding also improves the size–quality trade-off on representations from another upstream method. We compare companded and uniform quantization on the four listed NeRF Synthetic scenes trained with 3DGS-MCMC. Both use GP/local-mean configuration selection. Companded quantization reduces mean model size by 4.86% and increases PSNR by 0.1701 dB. It also reduces CVaR10 from 0.775 to 0.576 dB. These results support a key strength of our work: our quantization and configuration selection work with different training pipelines and provide practical compression benefits. We include the detailed results below for ease of review.

Table: Quantization and configuration selection across upstream methods. Bold values indicate improvements over the baseline.

Notes: C denotes SH companded quantization, and UQ denotes uniform quantization. G, R, GP, and Full denote greedy, random, GP/local-mean, and exhaustive configuration selection, respectively. Full does not refer to the complete framework. G, R, and GP each use 9 search evaluations. Full evaluates all 12 candidate configurations.

Reviewer_sY6eOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a post-training 3D Gaussian Splatting compression framework that combines signed-power companding of spherical harmonic coefficients with Gaussian-process-based configuration search. The method selects among pretrained base models, spatial block sizes, and quantization precisions under mean and tail PSNR-loss constraints.

Critical strengths
  1. The paper addresses a practically relevant problem: minimizing model storage under explicit rendering-quality constraints without requiring per-candidate retraining.
  2. The proposed framework achieves competitive overall performance, providing substantial compression while maintaining rendering quality compared with existing 3DGS compression methods.
  3. The paper presents extensive experiments and comprehensive ablation studies, providing useful evidence of the effectiveness of individual components and their complementary roles in the overall framework.
Critical weaknesses

1.Concerns regarding the clarification of contributions. The authors highlight “formulating post-training 3DGS compression as a budgeted delivery problem” as one of their main contributions in the Introduction. However, constrained compression optimization has already been explored in prior work. For example, SizeGS (ACM MM 2025) optimizes compression configurations under a storage-size constraint, while FlexGaussian (ACM MM 2025) supports compression configuration selection under a rendering-quality loss constraint, although this is not emphasized as its primary contribution. Despite differences in their specific formulations, the paper neither discusses these closely related works nor clearly distinguishes its problem formulation from them. The authors should clarify what is fundamentally new about the proposed budgeted delivery formulation and justify why it constitutes a core contribution.

  1. The necessity and effectiveness of GP-based search. The search space contains only 24 predefined configurations, with a default evaluation budget of 10 queries. According to Section 4.6, the complete search and confirmation procedure takes approximately 58.25 seconds on average. Given this relatively modest computational cost, it is unclear why exhaustive evaluation of all 24 configurations would be impractical or why a GP surrogate is necessary. Moreover, Appendix C.6 reports mean search-only oracle gaps of 6.98% and 7.39% for the proposed strategy and greedy search, respectively, indicating only a marginal improvement. The authors should provide a direct comparison with exhaustive search in terms of computational cost and compression performance to better justify the proposed search strategy.

  2. Insufficient evidence of consistent end-to-end compression benefits. Although the proposed companding strategy effectively reduces quantization errors and expands the feasible configuration space, these improvements do not consistently translate into better final compression performance. In Table 1, the complete framework produces larger delivered models than Uniform + Local Mean (16.856 vs. 15.680 MiB), with more reference fallbacks. Furthermore, the main comparisons with existing compression methods are conducted at predetermined operating points rather than under matched rendering-quality constraints. Since minimizing model size under quality constraints is the central objective of this paper, comparisons at equivalent quality thresholds would be more convincing. The current results do not sufficiently establish whether the proposed framework consistently provides better compression efficiency than existing methods.

Questions for authors
  1. What fundamentally distinguishes the proposed budgeted delivery formulation from prior constrained compression frameworks.

  2. Given the small search space and relatively low evaluation cost, why is GP-based search necessary? How does it compare with exhaustive search in runtime and performance?

  3. Why do the improvements in quantization feasibility not consistently translate into smaller delivered models? Can the authors provide comparisons with existing methods under matched rendering-quality constraints?

6 comments

1. W1 + Q1: What is new relative to SizeGS and FlexGaussian?

We thank the reviewer for pointing out the connection to existing constrained compression methods. We agree that constrained compression itself was not first proposed in this paper. Our work differs in the following three aspects.

First, we treat the number of candidate evaluations as an explicitly constrained resource. SizeGS mainly optimizes quality under a given storage budget, while FlexGaussian can adjust its configuration according to a quality or compression target. We further restrict how many candidate evaluations can be performed. By default, only nine candidate evaluations are allowed, together with at most one independent confirmation. We must therefore decide not only which configuration to adopt. We must also decide which configurations should receive the limited budget of encoding, decoding, and rendering evaluations.

Second, our quality constraint covers both average and tail loss. We check both the mean PSNR loss and the mean loss over the worst 10% of evaluated views. This allows the acceptance rule to reflect cases with similar average quality but different levels of degradation on individual views.

Third, we explicitly link configuration selection to actual delivery. After the search selects a candidate, we freeze the configuration and confirm it on views that were not used in the search. If confirmation fails, we deliver the lossless reference model. We also include the actual size of this fallback file in the results. Thus, we do not simply report the small model found during the search. We also account for whether it passes confirmation and for the actual storage cost of what is delivered.

In our implementation, we jointly select the pretrained base model, spatial block size, and precision of different SH bands. This allows us to compare trade-offs between different base models and different levels of quantization. We position our contribution as the concrete design of the constrained evaluation and delivery pipeline described above. We do not present it as merely reversing the objective and constraint in a classical rate-distortion problem.

2. W2 + Q2: Is GP-guided search justified for 24 configurations?

Under a bounded deployment/search budget, GP provides a better cost-performance trade-off than full-pool exhaustive search. We compared both methods using delivered archives across seven scenes: three seeds on hotdog and bonsai, and one seed on each remaining scene. GP uses Q=10, consisting of nine search queries plus one confirmation. Full-pool search uses Q=25, consisting of all 24 candidates plus one confirmation. Times exclude the shared base construction.

The key results are:

  • GP uses 10 rather than 25 evaluations and saves 50.7%–63.7% of search and confirmation time, with a mean saving of 57.7%. Full-pool search is 2.0–2.75× slower and adds 17.4–214.7 s per scene.
  • GP matches full-pool delivered size and quality exactly on 3/7 scenes: bonsai, room (LLFF), and steamtrain. On these scenes, exhaustive search provides no size or quality benefit.
  • On the other 4/7 scenes, full-pool search is only 5.9%–6.2% smaller, but GP has 0.12–0.16 dB higher test PSNR and approximately 0.12–0.16 dB lower mean and CVaR10 losses, while using less than half the search time.
  • At the same Q=10 budget, GP delivers files 18.3%–25.3% smaller than greedy search, with a mean saving of 20.3%.
  • All GP deliveries remain within the paper’s mean and CVaR10 limits, with better quality margins than full-pool search on the four non-tied scenes.

In short, the additional 15 evaluations required by exhaustive search produce no size gain on three scenes and only about 6% size reduction on four scenes, while more than doubling search time and reducing the quality margin. Full-pool search is the size-only upper bound and may be preferable for offline byte-only compression. However, when search cost is part of the delivery objective, GP is the better practical operating point. This is why we use GP as the default policy and report full-pool search as the size-only oracle. GP does not always reach the exhaustive size optimum. It is 5.9%–6.2% larger on four scenes. We therefore claim a better cost-performance trade-off, not universal size dominance.

SceneSeedsDelivered size GP / greedy / full (MiB)GP size overhead vs. fullTime GP / full; extra full time (s)Test RGB8 PSNR GP / full (dB)GP floating-point mean / CVaR10 loss (dB)Full floating-point mean / CVaR10 loss (dB)
hotdog31.655 / 2.026 / 1.5645.87%16.88 / 34.27;+17.3937.198 / 37.0420.349 / 1.1390.505 / 1.299
bonsai314.934 / 19.987 / 14.9340%95.53 / 214.88;+119.3532.528 / 32.5280.471 / 0.9370.471 / 0.937
chair12.472 / 3.034 / 2.3296.15%18.44 / 42.60;+24.1635.324 / 35.1980.209 / 0.6510.336 / 0.814
kitchen118.962 / 23.265 / 17.8836.03%122.55 / 337.23;+214.6831.917 / 31.7960.413 / 1.0190.533 / 1.148
room (LLFF)11.662 / 2.043 / 1.6620%12.13 / 32.00;+19.8723.896 / 23.896-0.272 / 0.570-0.272 / 0.570
Bike11.729 / 2.139 / 1.6296.10%15.80 / 39.01;+23.2139.752 / 39.6310.499 / 0.8950.620 / 1.047
Steamtrain11.410 / 1.848 / 1.4100%15.72 / 35.57;+19.8437.122 / 37.1220.767 / 1.2350.767 / 1.235

3. W3 + Q3: Do codec gains become smaller delivered models at matched quality?

We thank the reviewer for suggesting a direct comparison of delivered model sizes under the same quality constraints.

Why were the gains inconsistent in the original table? The final delivered size depends on both the compression configuration and fallback to the reference model when confirmation fails. In the original Table 1, Ours had more fallbacks. Their cost outweighed the savings from smaller candidates. As a result, its mean delivered size was 16.856 MiB, compared with 15.680 MiB for Uniform. We retain this result. The new experiments further test the delivery gains achievable under shared constraints.

The new controls use the same bases and candidate space. Ours and Uniform use the same small/reference FastGS bases, the same 24 candidate configurations, and the same encoding pipeline. Only the SH coefficient transform changes. Both methods evaluate every candidate. A candidate must pass both search and confirmation checks, with mean loss at most 1.0 dB and tail CVaR10 loss at most 1.5 dB. We select the smallest feasible delivery, freeze the selection, and then evaluate it on the test set. We measure actual file sizes, including bitstream metadata and the reference file when fallback occurs. Thus, this experiment directly tests compression and delivery gains. It does not depend on whether a search with a limited budget happens to find the best candidate.

The gains span multiple scenes and are not driven by a single extreme case. Table 2 lists 24 scenes with gains from the new and existing experiments. The median saving is 6.41%. The savings on Robot and Steamtrain are 16.63% and 20.37%, respectively. On chair, the 88.01% saving comes from Uniform falling back after confirmation fails, while Ours successfully delivers a compressed file. Even without chair, the remaining 23 scenes with gains still show a mean saving of 7.51%.

Comparisons with other methods also show clear delivery gains. On Jade, Robot, 360 room, and Family in Table 1, Ours is smaller than every other reported result that meets the constraints. Compared with the smallest feasible result from an external method in each scene, Ours saves 48.19%, 68.04%, 23.81%, and 40.65%, respectively. The smaller HAC and ContextGS files fail the shared loss limits. They therefore cannot be counted as better results under the same constraints. For FCGS, we report both stream size with a shared decoder and standalone size including decoder weights. Ours retains its advantage in the table under both accounting schemes.

Table 1: Four cases under the same mean and tail loss limits (MiB)

Bold marks the smallest reported delivery that meets the constraints in each row. F means that the result fails the shared mean/tail constraints. -- means no result is available and is not counted as a win for Ours.

SceneOursUniformHACContextGSC3DGSFCGS streamFCGS standaloneFlexGaussian
BlendedMVS/Jade3.0243.0740.396 F0.487 F6.1985.83715.587--
Synthetic_NSVF/Robot1.9322.3170.603 F0.684 F2.943 F6.04615.796--
360_v2/room11.35913.2055.471 F4.644 F14.90814.96624.71621.620
TanksAndTemple/Family2.0972.4421.071 F1.101 F3.533------

HAC, ContextGS, and C3DGS use the tested default configurations. For FCGS, we select the smallest feasible delivery among the tested configurations. FCGS stream assumes that the decoder is already installed. Standalone size also includes decoder weights. Losses for Ours and Uniform are measured relative to the FastGS reference. Losses for external methods are measured relative to their 3DGS reference.

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

This paper proposes a post-training compression framework for 3DGS that coordinates nonuniform quantization with budgeted configuration search. The method applies signed-power companding to higher-order spherical harmonic coefficients and uses surrogate-guided search to select the pretrained base model, spatial block size, and quantization precision under empirical rendering-quality constraints. An independent confirmation stage checks the selected representation and triggers a lossless-reference fallback if the quality limits are violated. Experiments across multiple benchmarks demonstrate improved quantization fidelity and storage–quality trade-offs, while maintaining high rendering throughput.

Critical strengths
  1. The matched experiments provide convincing evidence that signed-power companding improves reconstruction fidelity compared with uniform quantization. Results demonstrate that improved quantization fidelity can enable smaller feasible representations.

  2. The paper considers complete compressed-model storage, including metadata and lossless-reference fallback, rather than relying solely on nominal quantization rates.

  3. The experiments include multiple datasets, repeated training seeds, comparisons with established compression approaches, and analyses of search budgets, quality constraints, and computational costs. Evaluation provides useful insights into the practical trade-offs of post-training 3DGS compression.

Critical weaknesses
  1. The proposed framework primarily combines established signed-power companding and Gaussian-process-guided configuration search, without a clearly identifiable new joint optimization mechanism.

  2. The search acquisition prioritizes mean PSNR feasibility while final acceptance considers both mean and tail distortion, and failed confirmation triggers an immediate lossless-reference fallback without reconsidering alternative candidates.

Questions for authors

What is the fundamental methodological difference between the proposed coordination framework and a straightforward application of nonuniform quantization followed by conventional constrained black-box optimization?

Why is the companding exponent fixed globally at p=0.75p=0.75, rather than being adapted to different scenes, spatial blocks, or SH degrees?

Why does the acquisition function primarily consider mean PSNR feasibility when the final selection requires both mean and CVaR constraints?

Why does a failed confirmation immediately trigger lossless-reference delivery instead of reconsidering other feasible candidates already evaluated during search?

4 comments

2. Q2: Why is the exponent fixed globally at 0.75?

Thank you. We fixed p=0.75 before search as a reproducible operating point, keeping our method and the Uniform p=1 control under the same Q=10 protocol without tuning p by scene.

Across 192 matched configurations, p=0.75 lowered SH distortion and search view PSNR loss in every pair and raised search feasible candidates from 75 to 119. Appendix C.5 gives the initial four cases, 16 more cases where p=0.75 delivered smaller files than p=1 in 11/16, and a two seed study on eight scenes prelisted before training. The three exponents shared bases, candidates, camera splits, and the same search and confirmation budget. Test views were evaluated only after delivery was locked.

All 48 confirmations passed. Across the two seeds, p=0.75 delivered smaller files than p=1 in 15/16 pairs. On the new seed, it lowered both test losses versus p=0.5 in 6/8 scenes for a median 2.2% size increase. Versus p=1, it saved 3.1% median size with only 0.008/0.006 dB higher average test losses. These results support p=0.75 as a practical fixed size and quality tradeoff. The appendix reports seed variation and the materials tail exception.

Delivered size (MiB) | mean test PSNR loss | test tail loss (dB)

SeedScenep=0.5p=0.75p=1 (Uniform)
47counter9.73 / 0.66 / 1.2410.43 / 0.61 / 1.1910.88 / 0.46 / 0.98
47room (360)10.62 / 0.30 / 1.1711.49 / 0.01 / 0.6411.90 / 0.17 / 0.95
47materials2.38 / 0.75 / 1.942.41 / 0.57 / 1.582.58 / 0.38 / 1.14
47mic2.17 / 0.29 / 0.722.08 / 0.34 / 0.822.13 / 0.37 / 0.89
47fortress3.20 / 0.54 / 0.743.49 / 0.36 / 0.503.58 / 0.38 / 0.56
47horns5.94 / -0.30 / 0.626.12 / -0.30 / 0.625.84 / -0.16 / 0.66
47Caterpillar4.22 / 0.53 / 1.094.26 / 0.18 / 0.324.41 / 0.41 / 1.12
47Ignatius2.31 / 0.18 / 0.622.23 / 0.51 / 1.342.27 / 0.19 / 0.64
9counter9.76 / 0.61 / 1.2310.46 / 0.57 / 1.2010.91 / 0.41 / 1.02
9room (360)17.76 / 0.49 / 0.8617.07 / 0.63 / 1.1618.06 / 0.68 / 1.19
9materials2.39 / 0.77 / 2.002.43 / 0.57 / 1.582.60 / 0.39 / 1.15
9mic2.16 / 0.37 / 1.152.07 / 0.40 / 1.182.11 / 0.45 / 1.24
9fortress3.28 / 0.31 / 0.943.58 / 0.15 / 0.773.67 / 0.17 / 0.78
9horns5.22 / -0.46 / 0.475.60 / -0.59 / 0.335.82 / -0.56 / 0.37
9Caterpillar4.24 / 0.45 / 0.794.35 / 0.83 / 1.684.41 / 0.50 / 1.09
9Ignatius2.30 / 0.62 / 1.832.23 / 0.64 / 1.602.27 / 0.33 / 1.25

All 48 independent confirmations accepted compressed deliveries. Test views were evaluated only after delivery was locked. The p=0.75 materials result at seed 47 has an unrounded test tail loss of 1.575 dB, above the 1.5 dB test limit.

3. W2, acquisition + Q3: Why prioritize mean feasibility if final acceptance uses mean and tail limits?

Acquisition only ranks the next measurement. The final measured candidate guard and independent confirmation both check mean and CVaR10 loss. The paper's quality control ablation shows that confirmation lowers average test mean loss from 0.332 to 0.244 dB and CVaR10 loss from 1.134 to 0.944 dB, although it increases delivered size from 6.927 to 9.232 MiB.

Mean-guided acquisition is used only to prioritize evaluations, candidate selection and independent confirmation still enforce both mean and tail constraints. In the four matched cases, adding predicted tail feasibility to acquisition yielded no delivery benefit. We therefore retain the simpler mean-guided acquisition strategy.

SceneMean-guided: delivered MiBMean-and-tail-guided: delivered MiBIdentical delivered file
hotdog1.6671.667Yes
ficus15.388 F15.388 FYes
Family2.3062.306Yes
bicycle32.55532.555Yes

(Both acquisition rules used the same ten-evaluation budget: nine search evaluations and one confirmation. Only the acquisition feasibility score changed; final selection and confirmation still enforced both mean and tail constraints.)

the revision: Appendix C.7 tests acquisition using predicted mean and tail feasibility at the same ten evaluation cap. It delivers the same file as mean guided acquisition in all four studied cases.

4. W2, fallback + Q4: Why not confirm the next measured feasible candidate after a failure?

Immediate fallback is a conservative policy. We then conducted an additional experiment: the total budget remained ten evaluations, but eight were allocated to search and up to two to confirmation. All measured feasible candidates were ranked by file size, and the smallest candidate was confirmed first. If it failed, the second-smallest candidate was confirmed next.

As the table shows, for the ficus scene, the Backup policy also failed on the first confirmation, but it proceeded to confirm the second-smallest feasible candidate, which passed and ultimately delivered a 1.791 MiB compressed file. Its mean and tail losses on the locked test set were 0.286 dB and 0.658 dB, respectively, both below the 1.0 dB and 1.5 dB limits specified in the paper. For the hotdog scene, the Backup policy delivered 1.758 MiB, which is 5.48% larger than the original method's 1.667 MiB. The Backup policy reduced the number of search evaluations from nine to eight, measuring one fewer candidate. This narrowed the search space and, in turn, led to a larger selected file. For the Family and bicycle scenes, the results remained unchanged.

Although the ficus result appears promising, the Backup policy has a statistical issue. The second candidate is no longer subjected to a fully independent single-pass check, because it is selected only after observing the failure of the first candidate. Testing multiple candidates in sequence and accepting the first one that passes introduces selection bias or a multiple-testing problem: a candidate may pass because of random fluctuations in the confirmation views and is not necessarily reliable on new views.

the revision: Appendix C.7, Table 14, reports the backup policy's delivered files and locked test losses.

SceneOriginal MiBBackup MiBConfirmationsBackup test mean / CVaR10 loss, dB
ficus15.388 F1.79120.286 / 0.658
hotdog1.6671.75810.280 / 0.966
Family2.3062.30610.365 / 0.595
bicycle32.55532.55510.112 / 0.382

1. W1 + Q1: What problem does our method address beyond straightforward black-box search?

Thank you for this important question. Our motivation is that improving configuration search alone does not resolve two compression-specific bottlenecks: whether the codec provides sufficiently small feasible candidates, and whether the selected candidate can actually be delivered.

1. Search cannot compensate for a codec's distortion bottleneck. Lower SH precision and larger spatial blocks save storage, but can violate rendering-quality limits. A more effective optimizer cannot make those configurations feasible without changing their encoding. Our companding addresses this bottleneck directly. Holding the base, block size, and precision fixed across 192 matched configurations, it reduces mean search-view loss in every pair and increases guarded-feasible configurations from 75 to 119. Consequently, the smallest search-feasible stream becomes 1.64–13.98% smaller across all eight scenes. These results establish why improving the candidate representation matters before spending the limited search budget.

2. Search feasibility alone does not establish successful compression at delivery. A candidate selected on search views can fail confirmation, causing a much larger reference fallback. Our protocol checks the frozen candidate on disjoint views and counts the file actually delivered. Table 1 illustrates why this matters: companding saves 13.478 MiB in selected candidates, but additional fallback increments add 41.711 MiB. Although this does not eliminate the delivery bottleneck, it prevents candidate-level improvements from being mistaken for unconditional final-storage gains.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.