acceptodds
Under review as a conference paper at ICLR 2027

Spend Bits Where the Renderer Looks: Observability-Guided Rate–Distortion Optimization for Dynamic Gaussian Splatting Compression

Abstract

Compressing a trained dynamic Gaussian representation requires understanding both how parameter perturbations affect rendered images and when a primitive's contribution can be transferred to a persistent substitute. We introduce **ORDO-GS** (*Observability-guided Rate–Distortion Optimization*), a post-training compression framework informed by *render-space observability*: the expected sensitivity of rendered images to perturbations of the trained representation. Our analysis identifies a pronounced storage–observability mismatch: positions occupy little storage but require high precision to preserve rendering quality, whereas appearance dominates storage and can be compressed more aggressively. ORDO-GS combines render-aware structural reduction with subsequent attribute-specific compression. Structural reduction combines reconstruction-aware importance with trajectory compatibility to determine which primitives to retain and which contributions to transfer. Persistent co-location motivates complementary prune-only and unified delete–merge reduction branches, selected by a lightweight no-fine-tuning probe. For attribute compression, we derive a position-precision prior from a local render-distortion model and adopt channel-specific compact representations for temporal and appearance attributes. At matched primitive counts, ORDO-GS improves OMG4 by 0.30 dB while reducing storage by 12.6% on N3DV, and by 0.73 dB with 3.4% less storage on SelfCap. ORDO-GS also transfers to Ex4DGS, achieving consistent rate–distortion improvements across representations.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Dynamic 3D Gaussian Splatting, Model Compression, Render-Space Observability, Rate–Distortion Optimization

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_yrReOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes ORDO-GS, a post-training compression framework for dynamic 3DGS, aiming to improve rate–distortion performance by considering the rendering sensitivity of Gaussian attributes and the redundancy among primitives. The paper first introduce a render-space observability analysis based on the rendering Jacobian to characterize the sensitivity of different attribute groups, revealing a mismatch between their storage costs and rendering importance. Then the render-aware structural reduction strategy combines reconstruction-gradient-based pruning with trajectory-aware Gaussian deletion and merging. A lightweight probe selects the more suitable reduction strategy for each scene, followed by teacher-guided recovery. For attribute compression, the method employs an observability-based position-precision prior with render-based calibration, together with quantization and compact neural representations for the remaining attributes. Experiments on N3DV and SelfCap demonstrate improved rate–distortion performance over existing dynamic Gaussian compression methods.

Critical strengths
  1. Effective Rate–Distortion Performance: The proposed method achieves favorable rate–distortion performance on both N3DV and SelfCap, consistently improving reconstruction quality while reducing storage compared with the OMG4 baseline. The experiments also demonstrate its applicability to different dynamic Gaussian representations.
  2. Adaptive Structural Reduction: The proposed structural reduction strategy considers both the reconstruction importance of Gaussian primitives and their potential redundancy through trajectory-aware merging. The adaptive selection between pruning and merging provides a reasonable way to accommodate different scene characteristics, and the ablation studies demonstrate the effectiveness of these design choices.
Critical weaknesses
  1. Limited Ablation Studies on Structural Reduction: Although the paper provides ablations on different pruning and merging strategies, comparisons with simpler pruning baselines, such as opacity-based or random pruning followed by the same fine-tuning procedure, are missing. Such comparisons would help better isolate the benefits of the proposed structural reduction strategy from those of the subsequent recovery process. Additionally, given that the complete compression pipeline takes approximately 20 minutes per scene, a breakdown of the computational costs across different stages would be helpful for understanding the efficiency of the proposed framework.

  2. Limited Novelty and Practical Benefits of Attribute Compression: The attribute compression pipeline in Section 4.3 largely builds upon established techniques. For instance, compact latent features and shared MLPs for Gaussian attribute prediction have been explored in Scaffold-GS [1], while vector quantization have been investigated in prior Gaussian compression methods. The main distinctive component is the observability-based position-precision model. However, the analytic prior is followed by render-based calibration to determine the final bit depth. Since a straightforward search over candidate bit depths could potentially achieve the same rate–distortion performance with modest computational overhead, the practical benefit of the proposed observability-based precision selection remains unclear.

[1] Lu, T., Yu, M., Xu, L., Xiangli, Y., Wang, L., Lin, D., Dai, B.: Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)

Questions for authors
  1. Could the authors provide a comparison against a simpler structural reduction baseline, such as opacity-based pruning followed by the same teacher-guided fine-tuning procedure under an identical primitive budget and optimization schedule? This would help clarify how much of the reconstruction improvement comes from the proposed pruning and merging strategy rather than the subsequent recovery process.

  2. Could the authors clarify the key methodological contributions of the attribute compression pipeline beyond existing neural attribute coding and quantization techniques? Additionally, please provide a comparison of the total computational cost between the proposed observability-guided position-precision selection and a straightforward render-based bit-depth search, including the overhead of observability estimation.

3 comments
AuthorsOpenReview ↗

Structural reduction, attribute compression, and efficiency

We appreciate your recognition of the rate–distortion improvements, transfer across representations, and adaptive structural reduction. We have completed the requested pruning and timing comparisons, with additional evidence on the contribution of attribute compression.

W1-1 / Q1. Structural selection improves recovered quality. ORDO-GS gains 0.518/0.234 dB over opacity/random pruning after the same 16K teacher-guided recovery, outperforming random pruning in all six scenes. The table reports N3DV six-scene mean PSNR at 90K before attribute compression (three-run mean ± standard deviation; 500K teacher: 32.77 dB):

Structural reductionRecovered PSNR (dB)
Opacity-based pruning32.177 ± 0.034
Random pruning32.461 ± 0.042
ORDO-GS32.695 ± 0.005

ORDO-GS exceeds both baselines' 16K mean PSNR with one quarter of the recovery iterations. At 4K, it leads both by 0.79 dB and wins in all six scenes:

Recovery iterationsOpacity pruningRandom pruningORDO-GS
4K31.75131.75532.542
8K31.89432.17932.605
16K32.21632.48132.693

The submitted ablations test deletion scores and reduction operators (Table 4), as well as branch selection, recovery, and attribute coding (Table 5).

We further separate the two-stage contributions by encoding each method's recovered models with both codecs, keeping recovery fixed. Entries are six-scene mean PSNR gains on SelfCap/90K:

Recovered modelOMG4 codecORDO-GS codec
OMG4 reductionReference+0.941 dB
ORDO reduction+0.854 dB+1.885 dB

W1-2. Per-stage computational cost. Thank you for suggesting this analysis. A 250K run on corgi2 takes 13.7 minutes on an RTX 4090. We have included the stage breakdown in Table S5 of the revised manuscript (Appendix C.5):

StageComponentTime (s)
Structural reductionBranch-selection probe26.05
Primitive reduction and recovery269.49
Attribute compressionAppearance fitting205.06
Sub-vector quantization (SVQ)138.38
Joint codebook and decoder refinement169.06
Position precision calibration10.05
Serialization and LZMA compression4.21
Total822.31

W2-1 / Q2-1. Methodological contribution of attribute compression. We introduce render-space observability as a diagnostic tool for understanding and compressing dynamic Gaussian representations. The analysis identifies positions as precision-critical despite their low storage cost, while temporal and appearance attributes tolerate more aggressive compression (Table 1). We further derive an analytic model that predicts rendering distortion from position quantization and yields scene-specific precision estimates for a given quality tolerance (Section 4.3).

Measured observability improves distortion prediction across dynamic Gaussian representations (Table S7). At 16 bits, the cross-scene Spearman correlation for FTGS-family models increases from 0.769 with count/range prediction to 0.993 with the full model. The precision prior also generalizes to 4DGS and Swift4D, agreeing with codec-calibrated bit depths to within two bits in all 24 models (Appendix C.8).

The resulting codec improves quality and reduces storage over OMG4 on the same OMG4-selected and recovered models:

Dataset / budgetPSNR gain over OMG4 codingStorage reduction
N3DV / 250K+0.162 dB13.7%
SelfCap / 250K+1.086 dB6.4%

We have also expanded Related Work to discuss connections to prior neural attribute-prediction methods, including Scaffold-GS.

W2-2 / Q2-2. Cost of position-precision selection. Calibration checks and refines the model's predicted precision against the rendering-quality target. Prior-initialized calibration reaches the same final precisions as ascending render-based search with 24.0% less total selection time, including observability estimation. Both searches use the same candidate bit depths, training queries, and PSNR tolerance on the twelve FTGS-family scenes:

Selection methodTotal selection time (s)
Ascending render-based search160.19
Prior-initialized calibration121.77

More broadly, the prediction results discussed in W2-1 (Table S7 and Appendix C.8 of the revised manuscript) validate the quantitative relationship between rendering sensitivity, position precision, and rendering distortion across scenes and representations. This relationship provides scene-specific precision estimates that let calibration find the same final bit depths in less time.

Thanks for the response. The authors have provided additional clarifications on the methodological details, which have addressed most of my concerns. I will raise my rating to Weak Accept.

Thank you for taking the time to review our responses and additional results. We appreciate your constructive feedback and your positive reassessment of our work.

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

This paper introduces ORDO-GS, a post-training compression framework for dynamic Gaussian representations. It analyzes render-space sensitivity across parameter groups, reduces primitive counts through either pruning or joint deletion–merging, and compresses the remaining attributes using calibrated position precision, vector quantization, and shared appearance decoders. A lightweight probe selects the reduction branch, followed by teacher-guided recovery. Experiments on N3DV and SelfCap show improved quality–storage trade-offs over OMG4, with additional results on Ex4DGS.

Critical strengths
  1. The method achieves consistent improvements over its closest baseline. At matched primitive counts, improving PSNR with less storage on N3DV and SelfCap.
  2. Considering both reconstruction contribution and the availability of substitutes is a reasonable approach to dynamic primitive reduction.
  3. The position-quantization analysis provides a useful connection between renderer sensitivity, coordinate range, and required precision. The reported correlations with measured distortion and the comparison between predicted and calibrated bit depths provide supporting evidence for this model.
Critical weaknesses
  1. The practical contribution of observability is less clear than its central role in the paper’s framing suggests. Structural reduction uses reconstruction-loss gradients or residual-weighted footprint scores and trajectory distances, rather than the observability metric defined in Section 4.1.
  2. Merge compatibility relies primarily on projected-center trajectories, which may overlook differences in appearance, temporal support, and footprint shape. The robustness of this proxy and the median-based transfer threshold is insufficiently evaluated.
Questions for authors
  1. Which compression decisions directly depend on measured observability, and how would the pipeline change if this analysis were removed?
  2. How sensitive is deletion–merging to the transfer threshold, and would incorporating appearance, temporal support, or footprint differences improve the results?
3 comments
AuthorsOpenReview ↗

Observability-guided design and merge compatibility

We appreciate your recognition of the consistent quality–storage improvements, the contribution-and-substitute perspective on reduction, and the connection between renderer sensitivity and position precision.

W1-1 / Q1-1. The diagnostic role of observability. Render-space observability provides a diagnostic view of dynamic Gaussian representations by quantifying how parameter perturbations affect rendered images. It identifies which attributes tolerate aggressive compression and which require high precision. For positions, the distortion model further estimates the precision needed to preserve rendering quality. These diagnostic capabilities generalize across different dynamic Gaussian representations, including FTGS-family models, 4DGS, and Swift4D (Table 1; Table S4; Appendix C.8).

These diagnostics guide the following decisions:

Finding or measurementCompression decisionSupporting evidence
O1: view-conditioned positional nullspaceCompare substitutes through multi-view screen-space trajectoriesMatching-feature comparison (Table S8)
O2: low-observability temporal parametersCompact temporal codingFP16 position/time streams require 9.6–13.2% more storage at similar quality (Table 5)
O3: storage–observability mismatchProtect position precision and compress appearanceCalibration vs. uniform 16-bit positions: +0.377 dB for +0.157 MB; explicit SH: 2.08–2.33× storage at comparable quality (Table 5)
Measured position observabilityCompute the analytic position-precision priorPrediction and calibration tests below

Without this diagnosis, attribute encodings would instead rely on empirical comparisons, and render-based search would replace the analytic position-precision prior.

Structural reduction evaluates each primitive's importance to reconstruction. Observability measures how parameter changes affect rendered images, while the prune-only score measures how they affect reconstruction loss. The chain rule connects these two sensitivities through the same rendering Jacobian (Section 4.2).

W1-2 / Q1-2. Measured observability and position precision. Measured position observability enters the position-precision prior through the local render-distortion model (Section 4.3):

E[Dμ]≃2−2bμ12∑k=13Oμ,kAμ,k2,\mathbb E[\mathcal D_\mu]\simeq\frac{2^{-2b_\mu}}{12}\sum_{k=1}^{3}O_{\mu,k}A_{\mu,k}^2,

where bμb_\mu is the position bit depth, Oμ,kO_{\mu,k} is measured axial observability, and Aμ,kA_{\mu,k} is coordinate range.

Measured observability adds predictive information beyond primitive count and coordinate range: at 16 bits, the Spearman correlation with measured position-quantization distortion increases from 0.769 to 0.993 across the twelve FTGS-family scenes (Table S7). The resulting precision prior also transfers to shared-deformation representations: it agrees with codec-calibrated bit depths to within two bits in all 24 models (Appendix C.8).

Prior-initialized calibration reaches the same precisions as ascending render-based search with 24.0% less position-selection time on the twelve FTGS-family scenes, using the same candidate bit depths and including observability estimation.

W2-1 / Q2-2. Additional matching attributes. Thank you for suggesting these additional compatibility cues. We evaluate whether adding appearance, temporal support, or projected-footprint information improves trajectory matching. Keeping the survivor set and recovery schedule fixed, we report mean PSNR changes relative to trajectory-only matching over twelve scenes at 250K before attribute compression:

Added matching informationPSNR change (dB)
Appearance−0.054
Temporal support−0.016
Projected footprint−0.024

All three attribute augmentations slightly reduce mean PSNR in these tests, and we have included this ablation as Table S9 in Appendix E. Trajectory-only matching uses a more compact descriptor and requires less computation, while transfer and recovery accommodate attribute differences between matched primitives.

W2-2 / Q2-1. Transfer threshold robustness. The unified cost selects the survivors, and τm\tau_{\mathrm m} controls contribution transfer to those survivors. The transfer cutoff scales with each scene's median nearest-survivor trajectory distance, allowing a shared threshold. The submitted Appendix E already evaluates $\tau_{\mathrm m}\in \lbrace 0.5,1,2 \rbrace $ on dance1/2 and corgi1.

We extend the same sweep to all twelve scenes. The default τm=1\tau_{\mathrm m}=1 gives the highest mean PSNR and stays within 0.033 dB of the best tested threshold in every scene. Mean recovered PSNR spans only 0.021 dB across these settings.

Thank you for the clarifications and additional experiments, which have largely addressed my concerns. I will maintain my rating of Weak Accept. Please clarify the practical role of observability in the main text, distinguishing design choices motivated by the analysis from computations that directly use measured observability, particularly the position-precision prior.

Thank you for your constructive feedback and continued support. We have revised the Method overview and Section 4.3 to clarify which design choices are motivated by the analysis and how measured position observability is used to derive the position-precision prior.

Reviewer_Eqr4OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents ORDO-GS, a post-training compression framework for dynamic Gaussian splatting representations (FreeTimeGS-family and Ex4DGS) guided by render-space observability : the expected sensitivity of rendered images to parameter perturbations, estimated via Hutchinson's trace identity on the Gauss–Newton (Fisher) metric of the renderer without materializing the Jacobian. This analysis yields three quantified observations — O1: displacements along viewing rays leave projected positions unchanged (motivating screen-space trajectory matching of substitutes); O2: temporal centers and durations carry low aggregate observability (motivating low-rate temporal coding); O3: a storage–observability mismatch (positions: ~97% of observability, ~4.7% of storage; higher-order SH: ~70% of storage, negligible observability). The framework combines (i) render-aware structural reduction with two complementary branches — prune-only (loss-gradient energy) and unified delete–merge decimation (residual-weighted footprint deletion cost + trajectory-compatibility merging) — selected per model by a ~40 s no-fine-tuning probe; (ii) teacher-guided fine-tuning with interpolated views for recovery; and (iii) observability-informed attribute coding: an analytic position-precision prior derived from a local render-distortion model plus render-based calibration, Morton ordering with entropy coding, 8-bit time centers, SVQ for geometry/motion, and shared MLP decoders replacing per-primitive SH/opacity. At matched primitive counts on N3DV, ORDO-GS exceeds OMG4-S by 0.30 dB with 12.6% less storage (90K) and OMG4-L by 0.08 dB with 25.7% less storage (250K); on SelfCap it exceeds a controlled OMG4+interp. adaptation by 0.73 dB (250K); and it transfers to Ex4DGS (+0.28 dB, 8.7% less storage). Extensive ablations, per-scene results, temporal-quality metrics, and rate–distortion curves are provided.

Critical strengths
  1. The analysis is not decorative — it is quantified, used, and validated. Render-space observability is estimated with a practical Hutchinson scheme, each of the three observations is measured (Table 1), and the derived position-precision prior is validated against measured quantization distortion (Spearman 0.944–0.993 across 12 scenes at six precisions; prior within 2 bits of the codec-calibrated precision, Table S4/Fig. S2). Few compression papers close this loop between analysis and coding decisions, and this is the paper's most substantive contribution.

  2. The evaluation is more controlled than most in this area. Matched teachers, budgets, recovery schedules, and seeds; held-out views excluded from all selection and calibration; per-scene results, temporal-quality statistics, and unfavorable findings (RD saturation beyond 250K; the load-time appearance-baking difference affecting FPS) are reported. The two reduction branches are demonstrably complementary, and the no-fine-tuning probe lands within 0.021 dB of the oracle branch choice.

Critical weaknesses
  1. The central empirical claim — a better rate–distortion trade-off than the closest post-training compressor, OMG4 — is not convincingly established. On N3DV, the headline gain shrinks from 0.30 dB (90K vs. OMG4-S) to 0.08 dB (250K vs. OMG4-L), a margin comparable to the reported seed-level variation. On SelfCap, no official OMG4 checkpoint exists, so the +0.73 dB mean is measured against the authors' own adaptation (OMG4+interp.) rather than an independent implementation; per-scene results (Table S8) show this baseline actually wins on corgi1/corgi2 at both budgets (e.g., 28.511 vs. 28.066 dB at 90K), and the aggregate margin is driven almost entirely by dance1/dance2 (+2.54/+1.98 dB) — precisely the high-colocation scenes where the delete–merge branch dominates the prune-only baseline itself. The claimed superiority is therefore regime-specific, and its comparison basis is not independent of the authors. Since "beats OMG4" is the paper's principal evidence of contribution, this directly weakens the case for acceptance.

  2. Generality and practical claims exceed the evidence. The FPS advantage is confounded by a representation-level difference acknowledged in Sec. C.6 (ORDO-GS bakes appearance at load time; the released OMG4 path decodes neural appearance per frame). The cross-representation claim rests on a single additional representation (Ex4DGS) on which only the prune-only branch was exercised (App. C.8); the delete–merge branch and the temporal/motion attribute coding presuppose explicit per-primitive motion and appearance parameters and are untested on deformation-field, anchor-based, or feature-stream representations that dominate the dynamic-GS landscape. Evaluation is further limited to N3DV and six SelfCap clips, the many interacting pipeline components are only partially ablated jointly, and no code-availability statement is provided — so the reported numbers cannot currently be independently verified.

Questions for authors
  1. Per-scene results show OMG4+interp. wins on corgi1/corgi2 at matched budgets. Does the branch probe select the same branch as the oracle there, and what causes the deficit on these low-colocation scenes?
  2. How does the framework instantiate on deformation-field-based representations (e.g., 4DGS, Swift4D), where per-primitive trajectories are produced by a shared MLP and merging entangles the deformation field? Without this, the "handles other dynamic representations similarly" claim (Sec. 3) remains unverified.
3 comments
AuthorsOpenReview ↗

Rate–distortion evidence and branch selection (W1, Q1)

We appreciate your recognition of the paper's analysis–design–validation loop, particularly the measured observability, validated precision prior, and complementary reduction branches, as well as your positive assessment of the controlled evaluation.

The results support three conclusions:

  1. ORDO-GS achieves higher PSNR with less storage at 90K and 250K than OMG4-S and OMG4-L, respectively. Its 50K model also exceeds OMG4-S's PSNR with nearly half the storage.
  2. Results on corgi1/2 and dance1/2 demonstrate RD gains in both low- and high-co-location scenes, with the prune-only and delete–merge branches, respectively.
  3. The validated position-distortion model and controlled comparisons of structural reduction and attribute coding further substantiate the methodological contribution.

W1-1. Rate–distortion evidence on N3DV. Against OMG4's official checkpoints, ORDO-GS improves quality while reducing storage (teacher PSNR: 32.77 dB for ORDO-GS; 32.80 dB reported by OMG4):

ORDO-GSPSNR / MBOfficial OMG4: PSNR / MBStorage reduction
250K32.704 / 4.361L, 284K: 32.620 / 5.87025.7%
90K32.519 / 1.756S, 90K: 32.220 / 2.01012.6%
50K32.263 / 1.074S, 90K: 32.220 / 2.01046.6%

The 250K model is only 0.066 dB below its teacher. It outperforms OMG4-L with 25.7% less storage and fewer primitives. Its six-scene mean ranges from 32.672 to 32.749 dB across the reported runs, consistently above the official checkpoint.

The gain reaches 0.299 dB at the same 90K count; the 50K model also exceeds OMG4-S while nearly halving its storage. The RD advantage therefore extends across compression levels (Appendix D, Fig. S1).

W1-2. Fairness of the SelfCap comparison. To fairly assess the contribution of our compression design, we strengthen OMG4 from 24.687 to 26.338 dB at 250K (adaptation with a matched optimization budget: +0.687 dB; teacher-guided interpolated-view recovery: +0.964 dB). ORDO-GS still gains 0.73/0.92 dB over OMG4+interp. with smaller models at 250K/90K:

CountTeacherORDO-GS (PSNR / MB)OMG4+interp.OMG4 (original budget)
250K27.14027.070 / 4.57926.338 / 4.73724.687 / 4.697
90K27.14026.701 / 1.88425.777 / 1.89324.290 / 1.843

OMG4+interp. follows the released OMG4 implementation, with adaptation details in Appendix C.7.

W1-3 / Q1. Branch selection and corgi scenes. The corgi entries in the submitted Table S8 were obtained with delete–merge. We have re-evaluated the training-view no-fine-tuning probe using the same reduction scores and branch-selection procedure as in the paper. Prune-only achieves 1.250/0.937 dB higher PSNR than delete–merge before recovery, so the probe selects the loss-gradient prune-only branch for corgi1/2.

After recovery and attribute coding:

Scene / countORDO-GS (prune-only)OMG4+interp.
corgi1 / 250K28.941 / 4.47128.874 / 4.893
corgi2 / 250K27.901 / 4.53227.693 / 4.807
corgi1 / 90K28.520 / 1.79828.511 / 1.950
corgi2 / 90K27.742 / 1.83027.598 / 1.908

All four results improve quality and reduce storage.

W1-4 / Q1. Probe accuracy and robustness. Probe-based selection achieves mean PSNR only 0.011 dB below oracle selection of the better branch after recovery, across six SelfCap scenes and both budgets. Branch choices remain unchanged on all six scenes when the probe compares models retaining 40%, 50%, or 60% of the teacher's primitives, or when pruning scores use 6, 12, or 20 time points. These tests demonstrate robust branch selection before fine-tuning and near-oracle quality after recovery.

AuthorsOpenReview ↗

Efficiency, transferability, and component interactions (W2, Q2)

Thank you for your detailed comments on generality and practical utility. We address these points with the following evidence and explanations:

  1. ORDO-GS combines compact neural storage with explicit-Gaussian rendering through one-time appearance decoding.
  2. Both reduction branches and the observability analysis transfer to 4DGS and Swift4D across twelve scenes, with quality–storage gains over adapted OMG4.
  3. Structural reduction and attribute coding each improve quality in the stage-wise comparisons.

W2-1. Rendering efficiency. ORDO-GS combines compact neural storage with explicit-Gaussian rendering. Its appearance decoder runs once at loading to recover SH coefficients and opacities, which are reused across frames. This design avoids per-frame neural decoding while retaining the storage savings. The reported 1.6× throughput compares this inference path with the released OMG4-S implementation, as specified in Appendix C.6.

W2-2 / Q2. Transferability to representations with shared deformation fields. The Ex4DGS results use the prune-only branch selected by the paper's no-fine-tuning probe (Appendix C.8). Persistent co-location depends on the learned representation, and the probe selects the branch for each model. We now extend both branches to 4DGS and Swift4D, which use shared deformation fields.

Let ai(t)=Gϕ(θi,t)a_i(t)=G_\phi(\theta_i,t) map canonical Gaussian attributes to their time-dependent values. Observability follows by the chain rule, ∂Iq/∂θi=(∂Iq/∂ai)(∂Gϕ/∂θi)\partial I_q/\partial\theta_i=(\partial I_q/\partial a_i)(\partial G_\phi/\partial\theta_i), while the native renderer supplies reconstruction scores and projected trajectories for reduction. Transfer preserves each survivor's canonical query position and static/dynamic status. Recovery keeps ϕ\phi fixed, so changing one primitive's inputs does not alter the deformation mappings of the others.

The observability diagnosis also transfers: positions account for over 99% of IQR-scaled observability among canonical Gaussian attributes (motion is encoded by the shared field rather than FTGS's explicit per-primitive temporal attributes). The analytic position-bit estimate is within two bits of the codec-calibrated choice in all 24 models.

ORDO-GS compresses both representations while preserving rendering quality at 50% primitive retention. Results are six-scene means; sizes include the shared field:

4DGS

DatasetMethodPSNR ↑SSIM ↑LPIPS ↓Size (MB) ↓
N3DVTeacher31.7220.94420.053432.639
OMG4+interp.31.4700.94110.05678.045
ORDO-GS31.7480.94420.05528.017
SelfCapTeacher25.8380.86840.140765.642
OMG4+interp.25.3360.85880.16198.595
ORDO-GS25.9510.86710.14708.544

Swift4D

DatasetMethodPSNR ↑SSIM ↑LPIPS ↓Size (MB) ↓
N3DVTeacher31.6250.95100.0416131.886
OMG4+interp.31.4850.94760.044263.858
ORDO-GS31.7110.94980.043163.630
SelfCapTeacher26.0940.89840.0836262.990
OMG4+interp.25.1790.88180.113616.086
ORDO-GS26.0300.89210.097115.101

ORDO-GS improves mean PSNR by 0.23–0.85 dB over OMG4+interp. with smaller models.

Delete–merge gains 1.36–2.03 dB over prune-only after coding on dance1/2, the highest-co-location SelfCap scenes in both representations. This supports the two-regime design under shared deformation.

We have added these results to Section 5.4 (Table 7), with adaptation details in Appendix C.8 (Table S6).

W2-3. Component interactions. The submitted ablations cross deletion scores with reduction operators (Table 4), and test branch selection, interpolated-view and teacher-guided recovery, position precision, position/time coding, and appearance coding (Table 5; Appendix D).

We further examine the two-stage interaction: comparing codecs on identical recovered models isolates coding; comparing reduction methods under the same codec isolates structural reduction. Entries are six-scene mean PSNR gains on SelfCap/90K:

Recovered modelOMG4 codecORDO-GS codec
OMG4 reductionReference+0.941 dB
ORDO reduction+0.854 dB+1.885 dB

Both stages contribute; combining them gives the highest PSNR. ORDO-GS coding reduces storage by 4.7% on SelfCap/90K and 13.7% on N3DV/250K using the same OMG4-selected models.

W2-4. Reproducibility. We will provide the anonymous implementation and configurations in a separate OpenReview comment visible to the reviewers and area chairs. The package covers training, compression, observability diagnosis, rendering, evaluation, and the shared-field adapters.

AuthorsOpenReview ↗

Follow-up on our responses

Dear Reviewer Eqr4,

Thank you again for your feedback. Our responses include:

  • RD evidence and comparison fairness, including how the OMG4+interp. baseline used in our submitted paper improves on the original OMG4 setup through longer optimization and teacher-guided interpolated-view recovery.
  • Probe-selected results on corgi1/2, together with probe accuracy and robustness tests.
  • Transfer to 4DGS and Swift4D, covering both reduction branches, observability diagnosis, and precision prediction.
  • Cross-stage ablations separating the contributions of structural reduction and attribute compression.

We have also shared the anonymous implementation and reproduction instructions in a separate comment for reviewers and ACs.

We would be happy to clarify any remaining points. As the discussion deadline approaches, we would appreciate your updated assessment in light of this additional evidence.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of responses and manuscript updates

We thank the reviewers for recognizing the quantitative analysis, adaptive structural reduction, and quality–storage improvements. We clarify the role of observability in ORDO-GS and provide additional evidence on its rate–distortion performance and transfer across representations.

Methodological contributions and validation. Our observability analysis links rendering sensitivity to attribute-specific compression and yields an analytic position-precision prior. The reduction stage outperforms naive pruning (opacity-based or random) before and after matched fine-tuning. Attribute compression achieves higher quality with less storage than OMG4 on the same recovered models. The position model predicts quantization distortion, and its prior reduces calibration cost.

Rate–distortion performance and branch selection. ORDO-GS improves the quality–storage trade-off against both released OMG4 models and the strengthened OMG4+interp. baseline. Additional compression results on corgi1/2 demonstrate rate–distortion gains with the probe-selected prune-only branch.

Cross-representation transfer and robustness. We evaluate both reduction branches on 4DGS and Swift4D. ORDO-GS improves mean PSNR over OMG4+interp. by 0.23–0.85 dB with smaller models, including the shared fields (Section 5.4; Appendix C.8). Matching-attribute controls in Appendix E and threshold tests in our reply to rydG support trajectory-based matching and a shared transfer threshold.

Manuscript updates. The following updates are highlighted in blue:

  1. Shared-deformation transfer to 4DGS and Swift4D (Section 5.4, Table 7; Appendix C.8, Table S6).
  2. Predictor-factor controls (Appendix D, Table S7).
  3. Matching-feature controls (Appendix E, Table S9).
  4. Per-stage timings (Appendix C.5, Table S5).
  5. Clarification of the branch-selection protocol (Appendix C.3) and expanded related-work coverage of neural Gaussian attribute prediction.

Code availability. We will provide the anonymous implementation and configurations in a separate OpenReview comment visible to the reviewers and area chairs. The package covers training, compression, observability diagnosis, rendering, evaluation, and the shared-field adapters.

Further details are provided in our responses to the individual reviewers.

AuthorsOpenReview ↗

Manuscript update following our responses

We have incorporated the probe-selected SelfCap results into the revised manuscript, as previously discussed in our responses to Reviewer Eqr4. We have updated Tables 3 and S12, Fig. S1, and the corresponding numerical summaries accordingly.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.