acceptodds
Under review as a conference paper at ICLR 2027

Frequency-Decoupled Compression for 3D Gaussian Splatting

Abstract

Existing anchor-based 3DGS compression methods primarily focus on spatial context construction, while the interaction between coding order and frequency statistics remains less explored. This ordering can affect probability estimation and, consequently, compression performance. In context-based entropy coding, each anchor estimates its probability based on the causal history. However, due to the coding order, the finite causal window may contain many statistically inconsistent anchors, which degrades rate estimation accuracy. We propose FDC-GS, a frequency-decoupled compression framework that organizes anchors into frequency-consistent levels. This organization provides a coding-level frequency state to the entropy model and reorganizes the composition of the decoded history, yielding more statistically relevant causal context and tighter probability estimates without redesigning the context backbone. FDC-GS first applies saliency-guided decomposition to separate a low-frequency base from informative high-frequency residuals, and then aligns these components with coarse and refinement levels for base-to-refinement coding. Finally, a dynamic R–D counterbalancing schedule increases the rate penalty as high-frequency supervision grows, controlling the bitrate growth induced by detail refinement. Based on widely used hash-grid and triplane contexts, FDC-GS achieves average BD-PSNR gains of 0.249 dB over HAC++ and 0.443 dB over TC-GS across four datasets.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting compression, frequency-decoupled coding, context entropy model, rate-distortion optimization, progressive compression

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_gm47OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes FDC-GS, a method for compression of 3D Gaussian Splatting with a focus on modeling of high and low frequency components. It relies on 3 components:

  • It computes a saliency mask on the high-frequency images. The final loss is applied differently on low-frequency, high-frequency, and salient pixels.
  • A progress variable gradually increases and controls the entropy model.
  • An additional loss to regularize the number of anchors.
Critical strengths
  • The method is evaluated in depth, with qualitative, quantitative, and ablation results, on numerous datasets and against many baselines.
  • The proposed frequency-modeling is implemented on both hash- and triplane-based representations, bringing confidence that the proposed method transfers across architectures and consistently brings improvements.
Critical weaknesses
  • The information theory and 3DGS backgrounds are insufficiently discussed and poorly introduced. This includes terms such as anchors, conditional likelihood family, residuals, level masks, all of which are commonly used terms but should be clearly defined in this context. Relevant references are discussed approximately and in technical terms that are not explained. For instance, in the first paragraph, Scaffold-GS is described as employing "entropy coding to reduce bitstream size", which is at least not directly true, and "in these methods, the accuracy of probability estimation is a key factor", which does not make sense since there is no probabilistic or entropy modeling in this paper. All in all, these inaccuracies make the paper needlessly hard to read.
  • The improvement over previous methods is somewhat limited, both in terms of memory and visual accuracy.
  • It seems rather handcrafted and poorly justified to split the image into low-frequency, high-frequency and salient regions with hard masks.
Questions for authors
  • What justifies having two levels for low and high frequencies, rather than several bins or some more continuous representation?
2 comments
AuthorsOpenReview ↗

Response to Reviewer gm47

C1. Background and terminology

We agree that the original text conflated anchor representation with the entropy codecs built on it. Scaffold-GS introduces anchors and shared features but does not perform the learned entropy coding discussed here; probability estimation affects R--D performance in subsequent codecs such as HAC, HAC++, ContextGS, and HEMGS.

We now define the essential terms at first use: an anchor groups nearby Gaussians and shared attributes; a coding level contains anchors first activated by a nested mask; a residual is the quantized difference from an available decoded reference; and a conditional likelihood family is the probability model used for entropy coding. These corrections appear in the Introduction and Sec. 3.1 (lines 51--52 and 214--222).

C2. Magnitude and scope of the improvement

FDC-GS achieves clear gains in its directly optimized PSNR--size R--D objective, while LPIPS gains are limited because LPIPS is not part of the training loss. Across four datasets, average BD-PSNR improves by 0.0970.097 dB over HAC++ and 0.4420.442 dB over TC-GS, demonstrating transfer across hash-grid and triplane contexts. Our objective jointly optimizes pixel-domain reconstruction and bitrate, so its gain is primarily reflected in PSNR and size; LPIPS instead measures learned perceptual-feature distance and is not directly constrained. SSIM remains broadly stable, indicating that the R--D gain is not obtained by substantially sacrificing structural similarity. We now state the claim as improved PSNR--size efficiency rather than uniform improvement of every metric (lines 42--46, 625--647, and 710--716).

Q1/C3. Why low/high components and a saliency mask instead of several hard bins or a continuous representation?

FDC-GS does not hard-quantize the spectrum into two coding levels. It uses the continuous decomposition Ilow=Gσ∗II_{\rm low}=G_\sigma*I and Rhf=I−IlowR_{\rm hf}=I-I_{\rm low}; the saliency mask only suppresses weak spatial fluctuations within the residual and does not define a hard frequency band.

These two supervision signals are aligned with the multi-level PCGS hierarchy: low frequency supervises S0S_0, while annealing and level weights progressively distribute high-frequency supervision over S1,…,SL−1S_1,\ldots,S_{L-1}. Additional spectral bins would require arbitrary cutoffs and bin-to-level rules without addressing the base/refinement mismatch. Ablations support this design: removing the saliency mask, high-frequency supervision, annealing, or level alignment increases average BD-BR by 16.95%16.95\%, 15.87%15.87\%, 9.29%9.29\%, and 19.32%19.32\%, respectively. Secs. 3.2--3.3 clarify the decomposition and multi-level alignment (lines 273--413).

Reviewer_gm47OpenReview ↗

Clarifications

Background. Thank you for these clarifications, and the revised paper looks to have improved. I am not directly familiar with most of the methods listed, so I cannot check the rest of the paper. However, the one that I am familiar with (Scaffold-GS), was described in misleading terms, which is not reassuring.

Improvements. These remain very small on all metrics except on BungeeNeRF, and visually the difference is also very small.

Saliency mask. I see, thanks for clarifying.

The paper is evaluated in depth and seems technically sound, but overall I still have limited understanding of the method despite careful reading. This is a borderline paper, but I am still leaning slightly towards rejection and will keep my rating.

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

FDC-GS connects low-frequency and salient high-frequency image supervision to base/refinement coding levels and a synchronized rate-penalty schedule. The level state selects Gaussian or Laplace likelihoods while retaining the host spatial context and rendering decoder. The paper applies the design to HAC++ and TC-GS and reports improved PSNR–size curves across four datasets, supplemented by component ablations and conditional-NLL diagnostics.

Critical strengths
  1. The reported R–D gains across both hosts suggest useful transferability. The evaluation includes strong context and progressive baselines, component and likelihood-family ablations, and a controlled conditional-NLL diagnostic. Multi-rate support and the reported coding/rendering costs also help assess practical value.
Critical weaknesses

W1. The paper attributes compression gains to frequency-consistent causal histories, but Sections 3.3–3.5 define the context as the unchanged hash/triplane query, with coding level selecting the likelihood family. No dependency on previously decoded anchors is specified. With symbols and their predicted probabilities fixed, reordering Eq. (13) cannot change the estimated rate. Eq. (3) motivates state-conditioned modeling, not an ordering benefit; Appendix G does not establish that its diagnostic uses the production codec's conditional inputs. The reported gains may therefore come from frequency supervision and likelihood selection without validating the central ordering claim.

W2. Suspicious result. On Deep Blending, Table 5 reports 30.651 dB at 3.23 MB for FDC-Hash, whereas Table 9's full model gives 29.998 dB at 3.28 MB; the discrepancy persists across the other operating points.

W3. Novelty. The existing discussion of FreGS, PCGS and context-based codecs is useful. However, CodecGS [1] already connects progressive plane-channel organization with frequency-domain entropy optimization, while Plug-and-Play Optimization [2] combines distribution shaping with high-frequency detail compensation. These methods differ from FDC-GS, but should be distinguished to clarify what is new in its combination of frequency supervision, coding organization and likelihood selection.

[1] Soonbin Lee, Fangwen Shu, Yago Sanchez, Thomas Schierl, and Cornelius Hellge. Compression of 3D Gaussian Splatting with Optimized Feature Planes and Standard Video Codecs. ICCV, 2025.

[2] Tian Bai, Zheng Qiu, Haojie Chen, and Ziyang Dai. Plug-and-Play Optimization for 3D Gaussian Splatting Compression: Distribution Regularization, Probabilistic Pruning and Detail Compensation. AAAI, 2026.

W4. The benefit of synchronized rate scheduling is not fully isolated. The counterbalancing ablation does not clearly separate the benefit of synchronizing rate pressure with frequency supervision from the choice of rate-penalty strength.

Questions for authors
  1. How does coding order change the production entropy model's conditional inputs? Does Appendix G use the same inputs?

  2. What explains the discrepancy between Tables 5 and 9? Do the ablations evaluate the same system as the main results?

  3. What specifically distinguishes FDC-GS from CodecGS and Plug-and-Play Optimization?

  4. What rate weights are used in the counterbalancing ablation, and how is their strength separated from synchronization?

2 comments
AuthorsOpenReview ↗

Response to Reviewer Sgxr

Q1. Production entropy inputs and the ordering claim

Our ordering claim concerns frequency-to-level assignment, not permutation of fixed symbols with fixed probabilities. Although spatial lookup cic_i is unchanged, a refinement symbol is conditioned on its earlier-level feature, decoded neighboring features, and level state:

p(Δf^i(s)∣ci,f^i(s−1),Mi(<s),s).p(\Delta\hat{\mathbf f}_i^{(s)}\mid c_i,\hat{\mathbf f}_i^{(s-1)},\mathcal M_i^{(<s)},s).

Reassigning anchors and residuals across levels therefore changes the decoded references and predicted PMF; Morton order is retained within each level. Appendix G records the symbols and final PMFs used by the production arithmetic coder, not an auxiliary head. We clarify these inputs in Secs. 3.1/3.3/3.5 and the production-hook protocol in Appendix G (lines 196--272, 370--413, 490--508, and 980--1032).

Q2. Discrepancy between Tables 5 and 9

This was a result-assembly error, not a difference between the main and ablation systems. Table 5 mistakenly used the single playroom result, whereas Table 9 used the correct two-scene Deep Blending average. The corrected points are (29.998,3.28)(29.998,3.28), (30.220,4.23)(30.220,4.23), and (30.281,5.61)(30.281,5.61) (dB, MB), matching the full-model ablation row.

After regenerating the curve, Deep Blending BD-PSNR changes from +0.533+0.533 to −0.075-0.075 dB and the four-dataset average from +0.249+0.249 to +0.097+0.097 dB. We audited all remaining rows using complete scene sets and revised Fig. 3, Sec. 4.2.1, and Appendix C (lines 546--638 and 806--853).

Q3. Distinction from CodecGS and Plug-and-Play Optimization

The central distinction is the role of frequency information: FDC-GS uses it to organize the conditional anchor-coding hierarchy. CodecGS instead optimizes DCT coefficients of progressive 2D feature-plane channels for standard video coding; it neither aligns rendered-image frequencies with anchor levels nor uses decoded levels for refinement probability prediction. Plug-and-Play Optimization uses distribution regularization, probabilistic pruning, and a separate network to compensate high-frequency details after compression; it does not reorganize progressive conditional inputs. We added both comparisons to Sec. 2 (lines 174--190).

Q4. Rate weights and synchronization versus strength

Dynamic R--D counterbalancing increases rate pressure as high-frequency supervision grows, controlling refinement bitrate. The full model uses η(t)=1+(β−1)r(t)\eta(t)=1+(\beta-1)r(t) with β=2\beta=2 (1 before 15K, increasing linearly to 2 during 15K--25K); the ablation sets β=1\beta=1 with the same λs\lambda_s.

Strength sweeps show that larger terminal pressure consistently shifts R--D points toward lower rates. Fixed-versus-ramped controls show that fixed η=2.5\eta=2.5 preserves PSNR within 0.0100.010--0.0150.015 dB while using 10.610.6--12.2%12.2\% fewer bits than 1 ⁣→ ⁣2.51\!\rightarrow\!2.5. Thus, sufficient rate pressure explains most of the rate reduction, while the dynamic ramp remains a coupled curriculum that increases pressure with detail refinement. We validate the complete counterbalancing mechanism without claiming an independently isolated gain from synchronization alone. Full controls are in Secs. 3.4/4.3 and Appendix D (lines 428--449, 690--695, and 897--928).

Thank you for the clarifications, I still have two concserns:

  1. The production-NLL comparison changes both the learned representation and its organization, so it does not isolate the claimed benefit of frequency-consistent causal histories.

  2. The corrected R–D gains are modest, and Table 3 does not establish a consistent advantage over PCGS under matched training iterations.

I therefore maintain my rating of 2.

Reviewer_pKq2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FDC-GS, a frequency-decoupled compression framework that constructs and supervises different anchor groups to represent different frequency characteristics. To this end, it first separate low-frequency base components and high-frequency residual components. It then aligns these components with coarse and refinement levels for base-to-refinement coding scheme. During optimization, it gradually increase the weight for high-frequency components. Experimental results demonstrate that the proposed method improves the R-D performance compared to baseline entropy modeling methods for 3DGS compression.

Critical strengths
  • The paper introduces a new problem that has not been addressed by existing 3DGS compression methods and a frequency-decoupled coding strategy to address it. Specifically, it considers the relationship between frequency characteristics and coding order for entropy modeling. Also, experimental results show improved R-D performance in terms of PSNR and storage size. The level-wise analysis also shows that successive coding levels tend to capture progressively higher-frequency information.

  • The proposed method can be integrated with diverse entropy modeling methods. In the experiments, the authors demonstrate its application to both hash-grid- and triplane-based compression methods.

Critical weaknesses
  • The main motivation of the paper is insufficiently supported by empirical evidence. The authors argue that processing anchors with different frequency characteristics within the same causal context can degrade probability estimation. However, it does not directly analyze whether such frequency inconsistency actually occurs in existing coding schemes or how significantly it affects probability estimation. A direct analysis of this problem is necessary to justify the core motivation.
  • The construction of the coding levels is insufficiently explained. In the method section, it introduces multiple coding levels using monotonically increasing level masks, but it is unclear how anchors are assigned to each level. Since this level assignment is a key component of the proposed frequency-to-level alignment, the lack of explanation makes the method difficult to understand.
  • The comparison between the baselines and FDC is conducted under unfair training costs. FDC-Hash requires approximately 2.3× training time, while FDC-Triplane requires approximately 1.5×. HAC++ and TC-GS use 30K training iterations in their default settings, whereas FDC is trained for 40K iterations. Since this imbalance can significantly affect R-D performance, the comparison should be evaluated under a fair training setting.
  • Despite the unfair training setup, the performance improvement is relatively small, and some metrics even degrade. Overall, PSNR improves, while SSIM remains nearly unchanged and LPIPS often becomes worse with the proposed method. In particular, when comparing TC-GS and FDC-Triplane, LPIPS degrades substantially. This may indicate that the reconstruction quality of high-frequency regions is not actually improved.
Questions for authors
  • Can you provide an empirical analysis of whether frequency inconsistency actually occurs in existing coding schemes and how significantly it affects probability estimation?
  • Please provide more details on how anchors are organized and assigned to different coding levels.
  • The training setups for the baselines and FDC seem unfair. It is necessary to evaluate both methods using similar training budgets. It would also be helpful to provide further details on the training settings for both the proposed methods and the baselines.
  • While FDC improves PSNR, SSIM remains nearly unchanged, and LPIPS even degrades. In particular, when comparing TC-GS and FDC-Triplane, LPIPS degrades substantially. Could the authors clarify how these results support the effectiveness of the proposed method in improving reconstruction quality? The overall improvement appears relatively limited across different quality metrics.
2 comments
AuthorsOpenReview ↗

Response to Reviewer pKq2

Q1. Frequency inconsistency and probability estimation

Thank you for requesting direct empirical evidence. We address the two parts separately.

(1) Does frequency inconsistency occur in the existing coding scheme? We analyze the unmodified PCGS hierarchy using spectra of its base reconstruction and successive level increments. Across four scenes, only 0.080.08--3.04%3.04\% of base energy lies above the normalized radial-frequency threshold ρ=0.25\rho=0.25, whereas refinement increments contain 21.0821.08--63.14%63.14\%. Each refinement still contains substantial energy on both sides of the threshold, directly showing that an original coding level mixes low- and high-frequency responses. Appendix F reports the FFT protocol and threshold-sensitivity check (lines 937--978).

(2) How strongly does this affect probability estimation? NLL is the entropy model's average ideal code length for the symbols actually encoded; lower NLL means more accurate probability estimation. On bicycle and counter, pooled production NLL over S0S_0--S3S_3 decreases from 1.8771.877 to 1.7531.753 bits/symbol (6.64%6.64\%). For the same model pairs, cumulative size decreases by 9.459.45--12.82%12.82\% and 6.766.76--11.71%11.71\%, respectively. Overall, the 6.64%6.64\% NLL reduction accompanies a 10.17%10.17\% average cumulative bitstream reduction. Sec. 4.2.3 and Appendix G report this production-NLL/R--D analysis (lines 676--686 and 980--1032).

Q2. Anchor organization and coding-level assignment

FDC-GS inherits PCGS's nested level masks rather than introducing a new hierarchy. For anchor ii,

\ell_i=\min\{s:m_{a,i}^{(s)}=1\}.$$ Thus, an anchor is assigned to the first level at which any associated Gaussian becomes active. Because the masks are nested, coarse anchors remain available to all later levels. Anchors are coded in ascending level order, while newly activated anchors within each level retain Morton order. FDC-GS contributes frequency supervision and likelihood-state alignment with this inherited hierarchy. Sec. 3.3 now specifies this rule (lines 370--413). ### Q3. Training-budget fairness We address both iteration count and total training cost. **Iteration count.** We added 30K FDC-Hash results together with official PCGS 30K/40K results, enabling matched-iteration comparisons between progressive methods. FDC-Hash uses 30K on BungeeNeRF in both settings to avoid overfitting. Appendix B reports this sensitivity analysis (lines 750--764). **Total cost.** One FDC-Hash or PCGS run produces the complete progressive R--D curve, whereas HAC++ trains a separate 30K model per rate. Five HAC++ points therefore require 150K aggregate iterations on Mip-NeRF 360 and Tanks and Temples, versus 40K for FDC-Hash; on Deep Blending, the comparison is 90K versus 40K. Secs. 4.1/4.4 and Appendix B now report both per-model and complete-curve costs (lines 516--540, 698--708, and 750--764). ### Q4. SSIM, LPIPS, and reconstruction quality FDC-GS improves PSNR--size R--D performance, but not every metric. SSIM is generally comparable or higher, whereas FDC-Triplane has worse LPIPS than TC-GS. This is consistent with our pixel-domain distortion and rate objective, which contains no LPIPS term. Moreover, $S_0$--$S_2$ denote corresponding codec settings rather than exactly size-matched pairs; the formal comparison is therefore the complete PSNR--size curve. We restrict our claim to improved PSNR--size efficiency while acknowledging the LPIPS trade-off. Sec. 4.2.2 and Appendix Tables 6--7 state this scope explicitly (lines 639--647 and 806--853).

Thank you for the clarifications on Q1 and Q2. However, I still have concerns regarding the performance gain of the proposed method.

The motivation and the problem setting are interesting. However, the performance gain remains insufficient even after the rebuttal. As the authors acknowledge, SSIM is only comparable and LPIPS does not improve. The absence of an LPIPS term in the objective does not justify this result, since, to my knowledge, the compared methods do not use an LPIPS loss either. In terms of PSNR, under the same training budget, the proposed method performs nearly the same as PCGS and is even lower in some cases (Table 3).

Therefore, I will maintain my original rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.