acceptodds
Withdrawn from ICLR 2027

KC-3DGS: Kurtosis-Constrained Gaussian Splatting for High-Fidelity View Synthesis

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

3D Gaussian Splatting (3DGS) enables real-time novel view synthesis, but with sparse training views it produces floaters, color casts and oversmoothed regions at novel viewpoints. Pixel-space losses (L1, SSIM) do not prevent this: they constrain the total reconstruction error, not how it is distributed across scales. We propose KC-3DGS, which regularizes 3DGS with wavelet-domain statistics of natural images. It combines three terms: (1) a scale-weighted alignment of wavelet coefficients, (2) a supervised kurtosis loss that matches the heavy-tailed statistics of each subband to the ground truth, and (3) a supervised cross-band covariance loss that matches how oriented subbands co-occur. A simple analysis shows why each term is needed. Pixel losses cannot distinguish errors at different scales. Over-smoothing and floaters lower subband kurtosis, which only a supervised kurtosis loss penalizes. Cross-band covariance captures joint structure, such as corners, that per-band statistics miss. Because these statistics pool over image positions, they can supervise renders from pseudo cameras near the training views, where sparse-view artifacts arise; all terms start after densification, leaving the Gaussian set unchanged. With 12 training views on five benchmarks (MipNeRF360, Tanks&Temples, Deep-Blending, WRIVA-ULTRRA, MVImgNet), KC-3DGS improves perceptual metrics (PSNR, SSIM, LPIPS, DreamSim) on scenes, and as a plug-and-play regularizer it also improves three existing sparse-view 3DGS methods by up to 28.8% reduction in DreamSim and 1.59 dB increase in PSNR.

Primary area:
generative models
Keywords:
3DGS, natural image statistics, Novel view synthesis

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_A2jVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes KC-3DGS, which augments standard 3D Gaussian Splatting with three wavelet-domain losses. The first aligns rendered and ground-truth wavelet coefficients with scale-dependent weights. The second matches supervised per-band kurtosis statistics. The third matches supervised cross-band covariance statistics. A three-level Daubechies-3 transform is applied to rendered and ground-truth images. Pixel-aligned wavelet alignment is used on training views, while the statistical losses are applied to renders from translated pseudo-cameras after densification. The Gaussian representation and rasterizer are unchanged. The paper argues that these losses improve the distribution of image structure across scales and orientations, reducing artifacts such as floaters, color casts, and oversmoothing.

Critical strengths
  1. The paper addresses a relevant sparse-view failure mode with a multi-scale and orientation-aware representation. Wavelet coefficients can expose frequency-specific errors that are not visible from a single aggregate pixel-space objective, making the motivation for the regularizer meaningful.

  2. The method is modular and preserves the underlying 3DGS representation and rasterizer. The separation between training-view coefficient alignment and pseudo-view statistical supervision gives the method a clear structure and makes its intended integration point easy to understand.

Critical weaknesses
  1. The theoretical justification for the L1 refinement claim would benefit from clarification. Orthogonality directly guarantees L2 isometry, whereas the stated lemma concerns L1 norms; the argument provided does not appear to establish the general inequality and equality condition as written. The discussion also treats boundary effects as negligible, while the implementation uses zero padding.

  2. Kurtosis matching captures one aspect of the subband distributions, but does not by itself fully characterize image structure. Since it constrains standardized fourth-order statistics rather than band energy, phase, edge location, or spatial arrangement, images with different structures may still have similar kurtosis values.

  3. The supervised cross-band covariance loss matches only a scalar Frobenius norm. It loses subband-pair identities and correlation signs, so different orientation-specific covariance structures can have the same loss.

  4. The pseudo-view target may be inconsistent with the rendered viewpoint. A translated pseudo-camera uses statistics from the anchor-view ground truth, but the paper does not establish stability under parallax, occlusion, disocclusion, or view-dependent appearance.

  5. All three losses start after densification has ended, so they cannot add Gaussian primitives to represent missing regions. They can still adjust the existing set, but their capacity to recover structures that are wholly absent is limited.

  6. The plug-and-play claim is restricted by color parameterization. Shared color networks couple many Gaussians to one pseudo-view gradient, and the reported DNGaussian result shows that the losses do not transfer directly.

  7. The supplementary material has several visibly underfilled pages and uneven whitespace around figures and tables, which makes its layout look less polished. This is a minor presentation issue.

Questions for authors
  1. The key lemma claims a general L1 inequality for any orthogonal transform. Since orthogonality guarantees L2 isometry, can the authors provide a valid L1 derivation or revise the lemma for the implemented zero-padded db3 transform?

  2. What justifies using anchor-view ground-truth statistics for a translated pseudo-camera under parallax, occlusion, disocclusion, or view-dependent appearance?

  3. Since CBC matches only a scalar covariance norm, is the intended claim limited to aggregate dependence, or should it preserve individual orientation-specific correlations?

7 comments
AuthorsOpenReview ↗

W1 / Q1. The lemma and the zero-padded db3 transform

We thank the reviewer for this question; the concern is correct. Orthogonality guarantees only an L2L^2 isometry, and no L1L^1 ordering holds in either direction. For db3 with J=3J=3, a detail basis function has wavelet L1=1L^1 = 1 and pixel L1L^1 between 3.43.4 and 14.114.1, which violates the stated inequality. A single-pixel error reverses the ordering: wavelet L1=4.2L^1 = 4.2 and pixel L1=1L^1 = 1. The equality condition fails for the same reason.

We have therefore taken the second option and revised the lemma in L2L^2 for the implemented transform.

Zero-padded db3. With zero extension, the rendered error Δ\Delta is a finitely supported signal on Z2\mathbb{Z}^2, and our transform computes its orthonormal DWT while keeping every nonzero coefficient. Hence W⊤W=I\mathcal{W}^\top\mathcal{W} = I holds exactly, without neglecting boundaries. W\mathcal{W} is an isometry onto its range but is not surjective, since it produces more coefficients than pixels. We verified this numerically: the relative Parseval error is shown in Table below. The revised appendix states this explicitly in place of the earlier "negligible boundary" assumption.

DatasetRelative Parseval error
MipNeRF3602.3×10−72.3\times10^{-7}
Tanks&Temples2.1×10−72.1\times10^{-7}
DeepBlending2.1×10−72.1\times10^{-7}
WRIVA-ULTRRA2.4×10−72.4\times10^{-7}
MVImgNet2.2×10−72.2\times10^{-7}

The relative Parseval error ∣ ∥WΔ∥22−∥Δ∥22 ∣/∥Δ∥22\big|\,\|\mathcal{W}\Delta\|_2^2-\|\Delta\|_2^2\,\big|/\|\Delta\|_2^2 is measured at each dataset's training resolution. This residual comes from storing the db3 filter coefficients in single precision; with double-precision coefficients it falls to 3×10−163\times10^{-16}. This confirms that zero extension keeps all coefficients and W⊤W=I\mathcal{W}^\top\mathcal{W}=I holds exactly.

Revised Lemma 1 (and Lemma 3 in Appendix B). Let e(Δ)e(\Delta) collect the squared L2L^2 norms of all 3J+13J+1 subbands, including the approximation band. Then:

  • ∥Δ∥22=1⊤e(Δ)\|\Delta\|_2^2 = \mathbf{1}^\top e(\Delta), so pixel L2L^2 is a function of the subband error profile;
  • the converse fails: unit-norm detail basis functions at levels 11 and JJ, both supported inside the image, have equal pixel norm but different profiles.

The proof uses only W⊤W=I\mathcal{W}^\top\mathcal{W}=I, and its counterexamples lie in the image interior, so it holds exactly for the implemented zero-padded db3 transform. The lemma states what subband supervision adds: it distinguishes errors that pixel L2L^2 treats identically. We have removed all claims about L1L^1 ordering and "strictly stronger" supervision, and the revised appendix includes the L1L^1 counterexamples above.

These corrections do not affect our experimental conclusions. The ablations (Tables 4 and 10) show that the perceptual gain comes mainly from the pseudo-view kurtosis and covariance terms, not from the wavelet term this lemma concerns.

AuthorsOpenReview ↗

W4 / Q2. What justifies using anchor-view ground-truth statistics for a translated pseudo-camera?

We measured how far the anchor view's ground-truth statistics are from the pseudo view's true statistics. To do this, we trained a 3DGS model on all views to estimate the pseudo view's image, and checked the trained model's statistics against real held-out views where true ground truth exists.

The anchor target is approximate, and the trained model's statistics at the pseudo poses stay near the anchor's. In the held-out real views test, the trained model's statistics are closer to the true ground truth than 3DGS's. Varying the pseudo-camera offset from 0.1 to 0.4 grows the estimated target error 1.983x (kurtosis) and 1.309x (cross-band):

SettingscenesΔPSNR (wins)ΔSSIM (wins)ΔLPIPS (wins)ΔDreamSim (wins)
KC-3DGS (jitter 0.2)9+0.23 (8/9)+0.0128 (9/9)-0.0076 (9/9)-0.0104 (8/9)
KC-3DGS, jitter 0.19+0.20 (8/9)+0.0127 (9/9)-0.0075 (9/9)-0.0089 (8/9)
KC-3DGS, jitter 0.49+0.26 (8/9)+0.0131 (9/9)-0.0069 (9/9)-0.0083 (7/9)

These measurements do not separate parallax, occlusion, or view-dependent appearance. The target is an approximation that is robust in practice.

AuthorsOpenReview ↗

W2 / W3 / Q3. Kurtosis and CBC are not full structure descriptors. Is the intended claim limited to aggregate dependence?

Location, phase, edge position and energy are carried by L_wav, which aligns coefficients pixel-wise on training views. L_KC and L_CBC constrain only aggregate distribution shape and aggregate cross-band dependence. The CBC claim is restricted to aggregate dependence. We do not claim it preserves orientation-specific correlations or their signs.

Reviewer_KX5KOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes KC-3DGS, a plug-and-play method that augments standard 3DGS training with wavelet-domain loss. Extensive experiments on different dataset and 3DGS methods show its effectiveness under sparse-view setting.

Critical strengths

The proposed KC-3DGS is lightweight, can be plug-and-play added to almost any 3DGS pipeline.

Critical weaknesses
  1. Lack discussion with existing frequency/wavelet-based 3DGS methods, such as FreGS[1], Wavelet-GS[2], DWTGS[3] and LGDWT-GS[4]. DWTGS and LGDWT-GS target the same sparse-view setting with DWT-based losses, which are worth comparing with.
  2. In Table 11, results show that with full training set, the proposed approach may harm the performance. In section 5.2, the authors attribute the degradation towards the the pseudo-view term, but this claim is not backed up by any experiment, a per-term ablation on full training set would help.

[1] Zhang J, Zhan F, Xu M, et al. Fregs: 3d gaussian splatting with progressive frequency regularization[C]//2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024: 21424-21433.

[2] Zhao B, Zhou Y, Yu S, et al. Wavelet-gs: 3d gaussian splatting with wavelet decomposition[C]//Proceedings of the 33rd ACM International Conference on Multimedia. 2025: 8616-8625.

[3] Nguyen H, Li R, Le A, et al. Dwtgs: Rethinking frequency regularization for sparse-view 3d gaussian splatting[C]//2025 International Conference on Visual Communications and Image Processing (VCIP). IEEE, 2025: 1-5.

[4] Salehi S, Agashe A, McFarland A J, et al. LGDWT-GS: Local and Global Discrete Wavelet-Regularized 3D Gaussian Splatting for Sparse-View Scene Reconstruction[J]. IEEE Access, 2026.

Questions for authors
  1. Can the authors provide comparison and analysis with existing wavelet-based methods?
  2. Can the authors provide a per-term ablation on full training sets?
2 comments
AuthorsOpenReview ↗

W1 / Q1. Can the authors provide comparisons and analysis with existing wavelet-based methods?

We thank the reviewer for highlighting the gap in comparison against existing wavelet-based methods and the per-term ablation on the full training set. We address them below:

For LGDWT-GS, we use the released code, run through our train/test splits at the same resolution and iteration count. We use the loss weights of the released code, which differ from the paper's. An adaptive running-mean weight on the global DWT term (clamped to [0.1, 10]) instead of a fixed 1.0, and 0.1 on the patch term instead of 0.5. With the paper's weights, a trial run on bicycle (the scene with the most Gaussians) reached 23.6 GB on our 24 GB GPU at iteration 4,400 of 16,000, still inside densification, and was killed due to an out-of-memory error. With the released code's weights, every scene fits (peak 19.6 GB reserved). Against 3DGS it wins PSNR on 7/9 scenes, SSIM 7/9, LPIPS 8/9 and DreamSim 7/9. Against ours it wins PSNR on 6/9, SSIM 7/9 and LPIPS 7/9, and DreamSim on 3/9. It is built on a different 3DGS codebase (graphdeco's current 3DGS, with antialiasing and exposure compensation) and has no Gaussian cap, while 3DGS and ours cap at 3M.

We added our three terms to LGDWT-GS unchanged (same schedule and weights as on 3DGS, from iteration 15k, LGDWT-GS's own losses kept). Against LGDWT-GS alone, SSIM improves on 9/9 scenes (median +0.0051) and LPIPS on 9/9. PSNR improves on 7/9 (median below our run-to-run spread) and DreamSim is unchanged (5/9, below run-to-run spread). Against ours on 3DGS, the combination wins PSNR 8/9, SSIM 8/9, LPIPS 8/9, DreamSim 4/9. Our statistical terms therefore add to a pixel-wise wavelet-coefficient loss rather than duplicate it. The L_CBC share is computed without LGDWT-GS's wavelet terms.

ComparisonscenesPSNR wins (median Δ)SSIM wins (median Δ)LPIPS wins (median Δ)DreamSim wins (median Δ)
LGDWT-GS vs 3DGS97/9 (+0.44)7/9 (+0.0186)8/9 (-0.0109)7/9 (-0.0024)
LGDWT-GS vs KC3DGS96/9 (+0.15)7/9 (+0.0053)7/9 (-0.0030)3/9 (+0.0049)
LGDWT-GS + KC3DGS vs LGDWT-GS97/9 (+0.09)9/9 (+0.0051)9/9 (-0.0032)5/9 (-0.0004)
LGDWT-GS + KC3DGS vs ours98/9 (+0.19)8/9 (+0.0084)8/9 (-0.0085)4/9 (+0.0021)
LGDWT-GS + KC3DGS vs 3DGS98/9 (+0.37)8/9 (+0.0237)9/9 (-0.0136)6/9 (-0.0022)

Per scene, LGDWT-GS + KC3DGS minus LGDWT-GS:

scene (LGDWT-GS + KC3DGS minus LGDWT-GS)ΔPSNRΔSSIMΔLPIPSΔDreamSim
bicycle-0.11+0.0017-0.0048-0.0013
bonsai+0.50+0.0052-0.0021+0.0004
counter+0.28+0.0036-0.0026+0.0031
flowers+0.25+0.0051-0.0075-0.0057
garden+0.16+0.0085-0.0056-0.0004
kitchen+0.04+0.0005-0.0038+0.0005
room+0.09+0.0102-0.0032-0.0019
stump+0.02+0.0011-0.0027-0.0093
treehill-0.07+0.0078-0.0010+0.0001

DWTGS (reimplemented since there is no public code): mixed against 3DGS and not better than ours on a majority of scenes for any metric. Capped at 3DGS's Gaussian count per scene, it keeps its gains over 3DGS and still trails ours on SSIM and LPIPS, with PSNR and DreamSim within our run-to-run spread. KC3DGS beats 3DGS on 7-9/9 scenes per metric at every Gaussian cap from 500k to 3M; DWTGS on 3-7/9 (tables in the response to tSKa, W1 / Q2).

FreGS and Wavelet-GS: no public code for either. Wavelet-GS addresses dense views with a 3D wavelet split of the point cloud. All four will be added to Related Work.

AuthorsOpenReview ↗

W2 / Q2. Can the authors provide a per-term ablation on full training sets?

We ran L_wav alone on the full set of images for all 9 scenes and had neutral results. PSNR is better on 6/9 scenes (median below run-to-run spread). DreamSim better on 4/9. Pseudo-view nor full-view with L_KC alone does not produce better PSNR. The pseudo-view terms therefore account for at least part of the full-split degradation; pseudo-view L_CBC was not run alone. The revision adds this ablation:

Setting vs 3DGS on full-splitscenesPSNR wins (median Δ)SSIM wins (median Δ)LPIPS wins (median Δ)DreamSim wins (median Δ)
L_wav96/9 (+0.06)6/9 (+0.0002)6/9 (-0.0005)4/9 (+0.0002)
pseudo-view L_KC90/9 (-0.37)0/9 (-0.0065)0/9 (+0.0108)1/9 (+0.0012)
all three terms90/9 (-0.82)0/9 (-0.0102)0/9 (+0.0198)0/9 (+0.0022)
Reviewer_tSKaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper regularizes sparse-view 3DGS by aligning wavelet coefficients and matching subband kurtosis and cross-band statistics. The statistical terms are evaluated on nearby pseudo views after densification, with no changes to the Gaussian representation or rendering pipeline.

Critical strengths
  1. The implementation is straightforward to incorporate into existing 3DGS methods.
  2. Results across five datasets show useful perceptual gains; DreamSim improves on 48 of 57 scenes.
  3. The paper provides detailed ablations, repeated-seed results, and an open account of failures and overhead.
Critical weaknesses
  1. I am not yet convinced that the methodological advance is sufficient. The kurtosis component builds on DiffNat, while DWTGS already introduces wavelet-domain regularization for sparse-view 3DGS. Supervised statistical matching on pseudo views may be useful, but its advantage over this prior work needs a direct comparison.

  2. The quality-cost balance is difficult to assess from equal-iteration experiments. Dataset-average PSNR gains range from 0.11 to 0.29 dB, at 1.39-2.67 times the training time and roughly 20-31% more peak memory. The relevant question is whether the method still improves quality when training time is matched.

  3. The zero-loss argument is incorrect. Zero pixel L1 loss necessarily gives identical images and therefore zero wavelet differences. Since the implemented wavelet loss covers only detail bands, zero wavelet loss does not necessarily imply zero pixel loss either. Lemma 1 and its repetition in Appendix B should be corrected, along with the claim of strictly stronger supervision.

  4. The results also place limits on applicability. Benefits largely vanish at 50 views, dense-view PSNR decreases by an average of 0.77 dB, and DNGaussian loses perceptual quality. The dependence on when the losses are activated relative to densification further weakens the broad plug-and-play claim.

Questions for authors
  1. What happens when baseline and regularized models have equal training time?
  2. What advantage remains over DWTGS under matched sparse-view conditions?
  3. What theoretical justification remains after correcting the lemmas?
5 comments
AuthorsOpenReview ↗

W2 / Q1. What happens when baseline and regularized models have equal training time?

We thank the reviewer for highlighting gaps in training time comparisons, direct comparison against existing frequency based methods like DWTGS, and theoretical justifications. We address them below:

We increased the training iterations for FasterGS from 30k to ~51k-92k iterations to match the spread of training times with KC3DGS. Longer training does not help 3DGS in sparse settings. In the table below, we compare against two 3DGS methods: a vanilla 3DGS with 30k iterations and a vanilla 3DGS run with wall-clock matched to KC3DGS. The "better 3DGS" comes from choosing the better score between these two vanilla runs per scene and per metric. Results are reported in the table below.

ComparisonscenesPSNR wins (median Δ)SSIM wins (median Δ)LPIPS wins (median Δ)DreamSim wins (median Δ)
KC3DGS vs 3DGS (time-matched)98/9 (+0.34)9/9 (+0.0242)9/9 (-0.0130)9/9 (-0.0233)
KC3DGS vs 3DGS (30k)97/9 (+0.25)8/9 (+0.0130)8/9 (-0.0079)8/9 (-0.0088)
3DGS (time-matched) vs 3DGS (30k)92/9 (-0.15)0/9 (-0.0093)1/9 (+0.0045)1/9 (+0.0101)
KC3DGS vs better 3DGS (per scene, per metric)96/9 (+0.25)8/9 (+0.0130)8/9 (-0.0079)8/9 (-0.0088)

Breakdown of per-scene budgets of MipNeRF360

sceneloop-time ratio (KC3DGS / 3DGS)3DGS iterations at matched time
bicycle1.7151500
bonsai3.0992600
counter2.6378800
flowers1.8254700
garden1.7853400
kitchen2.2467300
room2.8585600
stump2.0060100
treehill1.7753000

We make a small correction to the training time factors. The reviewed draft's 1.39-2.67x cost figure accounts for total process time which includes data loading, checkpoint saving, and test renders. Timing the training loop alone from the same runs gives us an overhead of 1.41-2.18x with +20-31% peak memory. The revision reports the corrected range.

AuthorsOpenReview ↗

W1 / Q2. What advantage remains over DWTGS under matched sparse-view conditions?

DWTGS has no publicly released code. We made a good-faith reimplementation of their losses on our 3DGS base, at the paper's stated settings under our splits. Against 3DGS, results are mixed. Against KC3DGS, DWTGS wins on PSNR on 4/9 scenes, SSIM 2/9, LPIPS 2/9, and DreamSim 4/9. DWTGS ends at the 3M Gaussians cap on 5 scenes. In other scenes, it ends with 2.90-3.88x 3DGS' Gaussian counts. In comparison, KC3DGS stays below the Gaussian count of regular 3DGS.

To separate DWTGS's losses from its extra Gaussians, we reran it with densification capped at 3DGS's final Gaussian count per scene; it ends at 0.999x that count on every scene. It keeps its gains over 3DGS (PSNR 6/9, median +0.27 dB; SSIM 8/9; LPIPS 5/9; DreamSim 7/9), so they do not come from the extra Gaussians. Its indoor loss beyond our run-to-run spread is on bonsai and room. There, the excess squared error over 3DGS lies in the coarse bands (approximation plus coarsest detail level: 1.09 and 0.97 of it) and mostly in flat regions (0.75 and 0.46, against 0.27 and 0.24 on edges). In the worst views, it sits in regions that no training view covers. At equal Gaussian count it still trails KC3DGS on SSIM (3/9) and LPIPS (1/9). PSNR (4/9) and DreamSim (6/9) are within our run-to-run spread (0.23 dB, 0.006). Adding our three terms to capped DWTGS improves it on PSNR (7/9, +0.15 dB) and SSIM (6/9), but the combination is not better than KC3DGS alone beyond our run-to-run spread (PSNR 7/9, LPIPS 1/9).

ComparisonscenesPSNR wins (median Δ)SSIM wins (median Δ)LPIPS wins (median Δ)DreamSim wins (median Δ)
DWTGS (reimpl.) vs 3DGS95/9 (+0.32)5/9 (+0.0121)4/9 (+0.0054)6/9 (-0.0054)
DWTGS (reimpl.) vs ours94/9 (-0.11)2/9 (-0.0075)2/9 (+0.0125)4/9 (+0.0019)
DWTGS (reimpl., capped) vs 3DGS96/9 (+0.27)8/9 (+0.0114)5/9 (-0.0016)7/9 (-0.0130)
DWTGS (reimpl., capped) vs ours94/9 (-0.08)3/9 (-0.0065)1/9 (+0.0084)6/9 (-0.0021)
DWTGS (capped) + ours vs DWTGS (capped)97/9 (+0.15)6/9 (+0.0061)5/9 (-0.0010)6/9 (-0.0005)
DWTGS (capped) + ours vs ours97/9 (+0.09)6/9 (+0.0012)1/9 (+0.0074)6/9 (-0.0021)
DWTGS (capped) + ours vs 3DGS98/9 (+0.32)7/9 (+0.0177)6/9 (-0.0020)6/9 (-0.0056)
Ours vs 3DGS (30k)97/9 (+0.25)8/9 (+0.0130)8/9 (-0.0079)8/9 (-0.0088)

"Capped" limits densification to 3DGS's final Gaussian count on that scene.

Final Gaussian counts:

sceneDWTGSDWTGS (capped)DWTGS (capped) + ours3DGSKC3DGS
bicycle2,998,2872,996,4252,946,3622,998,1602,922,424
bonsai1,998,703514,933508,463515,281504,603
counter1,831,269594,108583,251594,409574,259
flowers2,998,0392,391,4072,343,3482,392,7652,342,252
garden2,998,6812,045,5562,023,3072,046,3312,011,034
kitchen2,489,137815,379810,361815,613807,595
room2,033,744701,589683,809702,238665,951
stump2,999,0551,615,7271,602,8831,616,1311,590,794
treehill2,997,8972,805,2322,760,5972,807,2442,770,575

Gaussian budget sweep. We also capped densification at 500k, 1M, 2M and 3M Gaussians (3M is the default) for 3DGS, KC3DGS and, DWTGS. KC3DGS beats 3DGS on a majority of the scenes on every metric at every cap. DWTGS beats 3DGS on 3-7/9 scenes per metric. At 500k, DWTGS leads KC3DGS on PSNR and trails it on SSIM (8/9) and LPIPS (6/9). Details in the table below.

CapComparisonmean final countsPSNR wins (median Δ)SSIM wins (median Δ)LPIPS wins (median Δ)DreamSim wins (median Δ)
500kKC3DGS vs 3DGS0.49M / 0.50M8/9 (+0.32)9/9 (+0.0168)9/9 (-0.0044)9/9 (-0.0098)
500kDWTGS vs 3DGS0.50M / 0.50M6/9 (+0.54)7/9 (+0.0103)6/9 (-0.0016)7/9 (-0.0089)
500kKC3DGS vs DWTGS0.49M / 0.50M3/9 (-0.07)8/9 (+0.0065)6/9 (-0.0025)4/9 (+0.0027)
1MKC3DGS vs 3DGS0.83M / 0.85M8/9 (+0.28)8/9 (+0.0160)8/9 (-0.0077)8/9 (-0.0073)
1MDWTGS vs 3DGS1.00M / 0.85M5/9 (+0.05)5/9 (+0.0109)3/9 (+0.0033)5/9 (-0.0070)
1MKC3DGS vs DWTGS0.83M / 1.00M4/9 (-0.03)7/9 (+0.0094)8/9 (-0.0104)3/9 (+0.0022)
2MKC3DGS vs 3DGS1.33M / 1.36M8/9 (+0.25)8/9 (+0.0160)8/9 (-0.0079)8/9 (-0.0088)
2MDWTGS vs 3DGS1.98M / 1.36M5/9 (+0.31)5/9 (+0.0119)3/9 (+0.0016)6/9 (-0.0086)
2MKC3DGS vs DWTGS1.33M / 1.98M5/9 (+0.03)7/9 (+0.0054)8/9 (-0.0123)5/9 (-0.0039)
3MKC3DGS vs 3DGS1.58M / 1.61M7/9 (+0.25)8/9 (+0.0130)8/9 (-0.0079)8/9 (-0.0088)
3MDWTGS vs 3DGS2.59M / 1.61M5/9 (+0.32)5/9 (+0.0121)4/9 (+0.0054)6/9 (-0.0054)
3MKC3DGS vs DWTGS1.58M / 2.59M5/9 (+0.11)7/9 (+0.0075)7/9 (-0.0125)5/9 (-0.0019)
AuthorsOpenReview ↗

W1. What is the novelty relative to DiffNat and DWTGS?

DiffNat applies an unsupervised kurtosis prior to generated images. We match per-band statistics of the ground truth and apply them on pseudo views after densification. DWTGS and LGDWT-GS align wavelet coefficients pixel-wise on training views. DWTGS's novel-view term is an unsupervised sparsity penalty on high-frequency coefficients. Neither matches ground-truth subband statistics. Adding our term to LGDWT-GS improves SSIM (9/9 scenes) and LPIPS (9/9 scenes). So the two are not redundant (details in the response to KX5K, W1 / Q1).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary: Comparisons with wavelet-domain methods, equal training budgets; ablations under full training views; and theoretical justification corrections

We thank the reviewers for their careful reading. Reviewers noted that the method is easy to adopt: it "is straightforward to incorporate into existing 3DGS methods" (tSKa), "is lightweight, can be plug-and-play added to almost any 3DGS pipeline" (KX5K), and "is modular and preserves the underlying 3DGS representation and rasterizer" (A2jV), with "a clear structure" that "makes its intended integration point easy to understand" (A2jV). On the motivation, A2jV wrote that "wavelet coefficients can expose frequency-specific errors that are not visible from a single aggregate pixel-space objective, making the motivation for the regularizer meaningful." On the evidence, tSKa noted that "results across five datasets show useful perceptual gains; DreamSim improves on 48 of 57 scenes" and that "the paper provides detailed ablations, repeated-seed results, and an open account of failures and overhead." KX5K noted that "extensive experiments on different dataset and 3DGS methods show its effectiveness under sparse-view setting."

The revision addresses the main requests. Changes are in blue in the revised PDF.

  • Equal training time (tSKa). 3DGS trained to KC3DGS's wall-clock time does not close the gap; longer training makes 3DGS worse in the sparse setting (Appendix A.3, Tables 7 and 8). The cost range is corrected to training-loop time.
  • Wavelet-domain baselines (tSKa, KX5K). LGDWT-GS (released code) and DWTGS (our reimplementation; no public code), with KC3DGS added on top of each, and a Gaussian-budget sweep from 500k to 3M (Appendix A.13, Tables 2 and 21–24). Our terms add to LGDWT-GS's pixel-wise wavelet loss rather than duplicate it, and KC3DGS beats 3DGS at every Gaussian budget.
  • Per-term ablation on the full training set (KX5K). The full-split degradation comes from the pseudo-view terms, not from LwavL_{wav} (Appendix A.8, Table 15).
  • Anchor-view targets for pseudo views (A2jV). Measured target error and a pseudo-camera offset sweep (Appendix A.5, Table 12).
  • Theory (A2jV). The L1L^1 lemma is restated in L2L^2 for the implemented zero-padded transform.
  • Scope (tSKa, A2jV). The plug-and-play claim is restricted to per-Gaussian color parameterizations (FasterGS, FSGS, FDS-GS, LAP; LAP per dataset in Table 20), with DNGaussian reported as a failure case. The method is stated as a sparse-view regularizer, and the losses are stated to refine an existing Gaussian set.
  • Presentation (A2jV). Supplementary layout restructured.

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

Related papers

Discussion (0)

Sign in to comment.