acceptodds
Under review as a conference paper at ICLR 2027

EvenSplat: Coupled 2D–3D Decomposition for Gaussian Splatting under Exposure and Illumination Variation

Abstract

A surface photographed under even light presents nearly the same appearance from every angle; the same surface under uneven light does not. Exposure changes between views, illumination varies within a single image, and locally strong light sources leave one region bright and its neighbor in shadow. Multi-view reconstruction methods such as 3D Gaussian Splatting treat these lighting artifacts as if they were properties of the scene, entangling capture-specific illumination with the geometry and color they recover. We present EvenSplat, a framework that separates the two. EvenSplat couples an image-space illumination decomposition with an illumination field carried by the Gaussians, so that the same explanation of the lighting is shared between the two-dimensional and three-dimensional views of the scene; a camera-response network and a local exposure-compensation module absorb the global and residual differences that remain across training images. Through extensive experiments across multiple datasets and diverse forms of uneven illumination (cross-view exposure, spatial illumination variation, and high-contrast lighting) on both real-world captured and simulated benchmarks, EvenSplat generally outperforms state-of-the-art methods, particularly under high-contrast illumination.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Gaussian Splatting, Uneven Illumination, Exposure Variation, Image Decomposition

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_3uVsOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses 3D Gaussian scene reconstruction and novel view synthesis under challenging exposure and illumination conditions. It considers three settings: cross-view exposure variation (CEV), where exposure differs across images; spatial illumination variation (SIV), where different regions within an image receive uneven illumination; and high-contrast illumination (HCI), where bright regions coexist with deep shadows. The goal is to reconstruct a consistent, more evenly illuminated base appearance from these challenging observations.

To achieve this, the paper proposes a coupled 2D–3D optimization framework that jointly learns an image-space illumination decomposition and a Gaussian-level illumination field. The two branches constrain each other by aligning their illumination estimates and recombining the estimated base appearance with illumination to reconstruct the input images. Global camera-response correction and local exposure compensation further account for residual differences across training images. The resulting shared scene representation supports base-appearance rendering from novel viewpoints. Moreover, this paper presents a benchmark combining synthetic data with real-world data, covering diverse indoor and outdoor scenes under varying lighting conditions.

Critical strengths
  1. The coupled 2D–3D optimization is well motivated, combining complementary strengths of 2D and 3D representations while encouraging consistent illumination estimates.
  2. Global camera-response correction and ILEC address image-wide and local variations, with illumination-guided regularization constraining local adjustments. Jointly represent every capture-specific photometric residual
  3. Quantitative and qualitative results support the method’s effectiveness across datasets and challenging lighting conditions.
  4. Component ablations demonstrate the contributions of the main modules and regularization terms.
Critical weaknesses
  1. The ablation studies of the 2D and 3D optimization modules are insufficient, as the key innovation lies in their mutual guidance and alignment. Additional experiments should be conducted to evaluate the effects of individual alignment and the absence of alignment.

  2. The ground truth images in real-world scenarios only serve as references under specific conditions. Directly relying on quantitative metrics for evaluation may therefore be misleading. For example, reconstructed images may preserve the correct overall structure but differ slightly in global brightness (The overall brightness remains uniform but deviates from that of the GT images), leading to lower metric scores that do not accurately reflect the actual reconstruction quality.

  3. The adaptive curve introduces a strong prior that may brighten inherently dark materials or darken inherently bright materials, potentially compromising the fidelity of material appearance.

Questions for authors
  1. Could the authors provide additional ablation experiments to address the first issue raised in Critical Weaknesses Q1?

  2. Could the authors propose a more appropriate evaluation method for real-world scenarios? Alternatively, if the current evaluation protocol is considered reasonable, could the authors provide further justification?

  3. Could the authors further discuss the trade-off mentioned in Critical Weaknesses Q3 and analyze the scenarios in which it may lead to undesirable results?

  4. Could the authors provide an analysis of the impact of the proposed module on training and rendering efficiency, including training time and rendering FPS?

5 comments
AuthorsOpenReview ↗

Rebuttal by Authors

We thank reviewer 3uVs for the supportive review and the suggestions on evaluation.

W1 / Q1. Ablate alignment individually and the absence of alignment.

We ran both variants: alignment only (λrecomb=0\lambda_{\mathrm{recomb}}=0) and recombination only (λalign=0\lambda_{\mathrm{align}}=0). Results are in the table below and in Appendix Table 12 of the revised paper. On HCI, alignment only and recombination only reach 15.70 and 15.83 dB, respectively, versus 16.95 dB for the full model. This shows that the two terms are complementary, and they impose different weightings. With R2D=I⊘L2D\mathbf R_{2D}=\mathbf I\oslash\mathbf L_{2D}, the recombination residual is I⊙(Ladapt/L2D−1)\mathbf I\odot(\mathbf L_{\mathrm{adapt}}/\mathbf L_{2D}-1), weighting relative errors by image intensity, whereas alignment directly penalizes illumination discrepancies.

Removing both terms removes Lcouple\mathcal L_{\mathrm{couple}} entirely. Since Lcouple\mathcal L_{\mathrm{couple}} is the only path from the image-space branch to the Gaussians, this is the revised Table 5 row "No Image-Space Decomp.", which drops HCI PSNR to 12.01 dB. The coupling is bidirectional by design, and gradients from Lcouple\mathcal L_{\mathrm{couple}} update both the image-space network and the Gaussian illumination branch. We also test one-way coupling (stop-gradient on the 2D outputs), which lowers HCI PSNR to 15.18 dB.

Table: Coupling ablations on real scenes (PSNR/SSIM/LPIPS).

VariantCEVSIVHCI
EvenSplat (full)19.526/0.708/0.30519.511/0.786/0.16916.950/0.682/0.241
No Lcouple\mathcal L_{\mathrm{couple}} (= revised Table 5 "No Image-Space Decomp.")15.924/0.646/0.37419.216/0.710/0.24212.012/0.438/0.324
Alignment only (λrecomb=0\lambda_{\mathrm{recomb}}=0)13.312/0.478/0.40018.494/0.717/0.21915.700/0.554/0.352
Recombination only, no alignment (λalign=0\lambda_{\mathrm{align}}=0)19.389/0.655/0.33419.130/0.726/0.21715.827/0.550/0.353
One-way coupling (stop-grad on 2D outputs)15.271/0.566/0.38316.128/0.667/0.26415.177/0.552/0.403

W2 / Q2. Real-world references' validity; a structurally correct but globally brighter/darker reconstruction is penalised. Justify the protocol.

We agree that absolute metrics might conflate global brightness differences with reconstruction errors, and therefore additionally report oracle per-image, per-channel affine-aligned metrics (yc=acxc+bcy_c=a_cx_c+b_c) on HCI dataset, as used for FourLogs in Fig. 11, to reduce global brightness and colour discrepancies (see the table below). After correction, EvenSplat remains ahead of the strongest baselines by 2.26 dB, also indicating that our HCI improvement cannot be explained solely by global brightness alignment. We retain uncorrected metrics as the primary evaluation in the paper because they directly measure reconstruction accuracy against the reference without GT-dependent post-processing, while the reference-free cross-lighting consistency test (Sec. 4.2) provides complementary evidence.

Metric3DGSGS-WBilateral GridLuminance-GSPPISPEvenSplat
PSNR ↑15.92916.08616.14316.25216.24518.508
SSIM ↑0.3830.4140.4210.3900.4200.663
LPIPS ↓0.5790.5420.5360.5190.5180.276
AuthorsOpenReview ↗

Rebuttal by Authors (continued)

W3 / Q3. The adaptive curve may brighten dark materials or darken bright ones.

Yes. The adaptive curve loss acts below Tdark=0.25T_{\mathrm{dark}}=0.25 and above Tbright=0.65T_{\mathrm{bright}}=0.65, with soft ℓ1\ell_1 targets of 0.45/0.55. However, the intrinsic ambiguity between reflectance and illumination may cause dark materials to be over-brightened and bright surfaces to be suppressed, as illustrated by the dark chair and white cabinet in DogRoom (Fig. 16, row 1); white preservation provides a soft constraint but cannot fully prevent these errors. This limitation is particularly evident under strong local illumination contrast, where the model may incorrectly attribute intrinsic material brightness to illumination variations. The table below further examines the sensitivity to λcurve\lambda_{\mathrm{curve}} and the weight of other priors.

Table: Parameter sensitivity analysis across CEV, SIV, and HCI scenes. Each entry reports PSNR/SSIM/LPIPS, averaged over the scenes in the corresponding category.

ParameterSettingCEVSIVHCI
All parametersEvenSplat (default)19.526/0.708/0.30519.511/0.786/0.16916.950/0.682/0.241
Strongest baselines(PPISP, PPISP, 3DGS+CHROMA)17.794/0.723/0.26918.112/0.771/0.17813.780/0.522/0.352
λcurve\lambda_{\mathrm{curve}} (default = 0.5)0.3520.016/0.720/0.30119.868/0.800/0.16516.716/0.671/0.246
0.6518.413/0.745/0.26919.285/0.779/0.17017.234/0.691/0.239
λsmooth\lambda_{\mathrm{smooth}} (default = 1.0)0.519.284/0.709/0.29819.269/0.787/0.16717.268/0.685/0.245
1.518.957/0.691/0.30319.700/0.787/0.17016.801/0.675/0.244
λwhite\lambda_{\mathrm{white}} (default = 0.2)0.119.515/0.714/0.30319.508/0.786/0.16916.902/0.679/0.244
0.319.893/0.714/0.29719.519/0.787/0.16917.107/0.686/0.240

Q4. Training time and rendering FPS.

The table below reports the computational cost on a single RTX 3090 GPU at 750×500 resolution. Due to the additional 2D decomposition network, illumination field, and image-specific correction modules, EvenSplat inevitably requires 2.21× the training time and 3.22× the peak training memory of 3DGS, trained for 10k iterations. At inference, the 2D network, CRN, and ILEC are discarded, while the illumination field is retained. EvenSplat achieves 74.36 FPS, compared with 289.60 FPS for 3DGS, maintaining interactive rendering performance.

Methodtrain time (min)peak mem (GB)render FPS
3DGS (gsplat)15.661.0428289.60
EvenSplat34.583.354674.36
AuthorsOpenReview ↗

Comment by Authors

As the author-reviewer discussion period comes to an end, we wanted to briefly check in to see whether our responses have resolved your concerns. If anything remains unclear, we would be happy to offer further clarification or additional analysis. We thank you again for your time and for your thorough feedback.

Reviewer_N64uOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses 3D Gaussian Splatting reconstruction from unevenly illuminated images, where lighting differs both within a view and across views. It groups these inconsistencies into three categories: cross-view exposure variation (CEV), spatial illumination variation (SIV), and high-contrast illumination (HCI). It contributes a nine-scene real-world dataset (3 CEV, 4 SIV, 2 HCI) with evenly lit reference captures, and a method, EvenSplat, for recovering an evenly lit 3D reconstruction. Inspired by Retinex decomposition, the method treats each Gaussian’s color as a base color multiplied by an illumination field shared across views. All components are optimized jointly in a coupled 2D- 3D framework: the Gaussians, the illumination field, a 2D decomposition network that predicts per-image illumination, and per-image global and local exposure corrections that absorb capture-specific variations during training and are dropped at test time. Training is constrained by several hand-crafted regularizers, each motivated by a specific purpose. Validation, including ablations, covers the authors’ real-world dataset and perturbed Mip-NeRF 360 scenes.

Critical strengths
  1. Practically relevant problem. This paper addresses three cases together (CEV, SIV, and HCI) and also evaluates against the real data while most prior works are limited to synthetic perturbation. Aligning an online image space Retinex-style estimate with a shared Gaussian illumination field is interesting.

  2. Proposes real-world dataset. The 9 scene dataset covering CEV, SIV, and HCI under realistic camera response, flash fall-off, and saturation provides evaluation in real-world scenarios. Prior works seem to rely mostly on synthetic perturbations.

Critical weaknesses
  1. The proposed framework is an engineering tweak. It is mostly integration of different components such as Retinex-style decomposition (also addressed in works such as L^2DGS, LL-Gaussian), and multiplicative illumination (used in L^2DGS). Though the paper claims to couple 2D and 3D, where image-space network estimates illumination and reflectance, they are used as a supervision of output from 3D path (for example, L_{2D} is estimated in 2D which which is used to supervise L_{adapt} estimated from 3D information), as clearly highlighted in Fig. 2.

  2. The biggest improvement in the paper is on real HCI, about 3.4 dB over the next best method. But it is not entirely convincing that this comes from the coupled 2D-3D design. Two local, hand-crafted components seem to account for gains of a similar size. Removing ILEC costs 2.9 dB on HCI (Table 4) while being nearly neutral on CEV and SIV. The only ablation of the adaptive curve loss (Table 5) also shows a 2.9 dB drop on HCI, but that number is for the 2D branch's own output, not the full model. Both act locally, the curve loss pushes each region toward mid-tone, and ILEC rescales brightness per region in each training image. The Tone-Matched 3DGS baseline only applies one global gain per image, so it doesn't really address this. The ``No Image-Space Decomp.'' ablation removes the coupling and the 2D priors together, so it can't separate them either. Without the full model trained without the curve loss (and ideally without both the curve loss and ILEC), it's hard to tell how much of the HCI result is due to the proposed coupling.

  3. The objective is a weighted sum of 12 losses, plus a staged training schedule, but there is no sensitivity analysis on the weights (at least, I could not find it). This matters more than usual here. With several competing objectives on an ill-posed decomposition, the weights effectively decide which split between base color and illumination the method ends up with. Many weight settings can reach a similarly low total loss with quite different decompositions. Several terms are also never ablated on their own. The alignment and recombination losses are the clearest case. They are the core of the claimed coupling, yet they are only removed together with the whole 2D branch. They also seem to ask for the same thing. Since R2D=I⊘(L2D+ϵ)R_{2D} = I \oslash (L_{2D}+\epsilon), the recombination term compares I⊙Ladapt⊘L2DI \odot L_{adapt} \oslash L_{2D} with II, which is satisfied exactly when Ladapt≈L2DL_{adapt} \approx L_{2D}. That is the same condition as the alignment loss, just measured as a relative error weighted by image intensity rather than as an absolute difference. The two may therefore be largely redundant, and an ablation keeping one and removing the other is needed to see whether both are necessary. Similarly, the image-space priors are ablated only on the 2D output, not on the final reconstruction. So it is difficult to say which parts actually matter for the final result or how sensitive the method is to these choices. Also other ablations are hard to interpret, which should be discussed explicitly in detail.

  4. The equations in Sec. 3.1 - 3.3 need proper motivation, intuition and explanation. The authors should mention the dimension and meaning of each quantity, and how they are estimated. For example, estimation of normal, quantities/notations involved in Eq. 2-5, and motivation/intuition of Eq. 4, estimation of r(t) etc. While some of these are included in the appendix, it is hard to follow the notations and equations in their current form.

  5. The paper does not provide multiple seeds experiments, variance or confidence interval. 3.41 dB HCI claim is based on just 2 scenes. Also, there is a degradation of ~2.7 dB on CEV when removing normals or view-directions. The same removal has almost no effect on SIV and HCI (and no view direction is slightly better on SIV). This is not explained.

  6. There are multiple missing comparisons with methods such as LITA-GS, LL-Gaussian, Gaussian-DK, HDR-GS, and SWAG.

Questions for authors
  1. Following Weakness 2, report the full EvenSplat model trained without the adaptive curve loss on the real HCI scenes (ideally also with ILEC removed). This would help show how much of the HCI gain comes from the coupling itself.

  2. Following Weakness 3, report results with only the alignment loss and with only the recombination loss? Given that both effectively enforce Ladapt≈L2DL_{adapt} \approx L_{2D}, it would help to know whether both are needed.

  3. Compare against other methods: LITA-GS, LL-Gaussian, Gaussian-DK, HDR-GS, SWAG.

  4. Though each scene has a separate evenly lit capture, the dataset size is limited (9 scenes) with only 2 HCI and 3 CEV. The subjects are mostly small objects (logs, a robot, benches, rocks, a cart). That is not rich or diverse, and it does not match the motivation about large indoor spaces with mixed, locally varying lighting. Furthermore, HCI and SIV are both made with one flash/under-exposure setting. There are no windows, mixed color temperatures, multiple light sources, specular materials or moving shadows.

  5. Why do "No Normals" and "No View Directions" result in degradation of ~2.7 dB on CEV but almost none on SIV or HCI?

  6. Table 5 shows that allowing colored illumination can improve some metrics. Have the authors explored a lightly constrained colored variant more thoroughly on the real HCI/SIV scenes?

  7. The base color keeps SH up to degree 3, with the higher-order coefficients only penalized. Have you tried using only the DC term (SH degree 0)? That removes view-dependent effects from the base color by construction. It would be useful to see this compared with the current setup, including on the cross-lighting consistency test.

  8. The paper does now show any failure cases. What would happen, for example, in the presence of intrinsically dark or bright materials?

  9. Clarify what ``Image-Space Decomp. Only'' in Table 4 is exactly.

Minor

  1. Figs. 11–13 do not seem to have been updated. They label one column “Retinex-GS” (which is presumably an earlier name of the method) and another “3DGS+CHROMA”, which is not among the listed baselines and appears only in appendix Table 11, not in the main comparisons.

  2. Equation references in the text appear as “Eq. equation N” throughout.

  3. The ϵ in R2D = I ⊘ (L2D + ϵ) is unnecessary since L2D is already bounded below by 0.05.

5 comments
AuthorsOpenReview ↗

Rebuttal by Authors

We thank the reviewer N64u for a precise and constructive review. We ran all of ablations requests (see Tables A–E below), and revised Sec. 3 and the ablation discussion in the paper.

W1. The framework integrates known components (Retinex decomposition, multiplicative illumination as in L²DGS / LL-Gaussian); the 2D output only supervises the 3D path.

The coupling is bidirectional by design: no stop-gradient is applied, so Lcouple\mathcal L_{\mathrm{couple}} updates both the image-space network and the Gaussian illumination branch (Sec. 3.4). Table A, row "One-way coupling (stop-grad on 2D outputs)", isolates this. What is new relative to L²DGS / LL-Gaussian is (i) the online, per-scene 2D decomposition trained from priors alone and coupled in both directions, and (ii) the split between a transferable explanation (shared field) and a non-transferable one (CRN, ILEC), which is what allows base-appearance rendering at novel views. Table E below (revised paper, Table 16), "Intrinsic-decomp + 3DGS", shows offline 2D priors (Retinex/intrinsic decomposition) do not substitute for this.

W2 / Q1. The 3.4 dB HCI gain may come from ILEC and the adaptive curve loss (both local), not from the coupling; report the full model without the curve loss and without ILEC.

First a structural point: the curve loss acts on R2D\mathbf R_{2D} and reaches the Gaussians only through Lcouple\mathcal L_{\mathrm{couple}} (Eq. 8), so the coupling is the mechanism by which the prior constrains 3D rather than an alternative to it. Table A gives the requested full-model rows on real HCI: no curve loss 16.08; no ILEC 14.07; neither 14.17; no Lcouple\mathcal L_{\mathrm{couple}} 12.01 (Table 5 "No Image-Space Decomp."; this row is the no-coupling model because Lcouple\mathcal L_{\mathrm{couple}} is the only link between the branches). Conclusively, the coupling is the largest single contributor (−4.94 dB when removed), and we agree that ILEC accounts for a substantial share of the HCI gain.

Table A: Coupling and prior ablations on real scenes (PSNR/SSIM/LPIPS).

VariantCEVSIVHCI
EvenSplat (full)19.526/0.708/0.30519.511/0.786/0.16916.950/0.682/0.241
No adaptive curve loss17.023/0.592/0.37219.369/0.752/0.20216.081/0.624/0.353
No ILEC19.599/0.707/0.30819.467/0.782/0.16814.066/0.561/0.266
No curve loss and No ILEC17.780/0.596/0.37619.271/0.743/0.21014.171/0.537/0.401
No Lcouple\mathcal{L}_{couple} (= revised Table 5 "No Image-Space Decomp.")15.924/0.646/0.37419.216/0.710/0.24212.012/0.438/0.324
Alignment only (λrecomb=0\lambda_{\mathrm{recomb}}=0)13.312/0.478/0.40018.494/0.717/0.21915.700/0.554/0.352
Recombination only (λalign=0\lambda_{\mathrm{align}}=0)19.389/0.655/0.33419.130/0.726/0.21715.827/0.550/0.353
One-way coupling (stop-grad on 2D outputs)15.271/0.566/0.38316.128/0.667/0.26415.177/0.552/0.403
AuthorsOpenReview ↗

Rebuttal by Authors (continued)

W3 / Q2. Loss weights with no sensitivity analysis; alignment and recombination enforce the same condition and are never ablated separately.

On redundancy, the two terms share a minimizer but impose different weightings. With R2D=I⊘L2D\mathbf R_{2D}=\mathbf I\oslash\mathbf L_{2D}, the recombination residual is I⊙(Ladapt/L2D−1)\mathbf I\odot(\mathbf L_{\mathrm{adapt}}/\mathbf L_{2D}-1), weighting relative errors by image intensity. Alignment directly penalizes illumination discrepancies. Table A shows that alignment only and recombination only achieve 15.70 and 15.83 dB on HCI, respectively, versus 16.95 dB for the full model, demonstrating their complementary contributions. On sensitivity, Table B below (Table 9 in the revised paper) reports one-at-a-time variations of selected loss weights, training schedules, and prior parameters. The results show that parameter sensitivity varies across illumination and exposure categories. For example, increasing λcurve\lambda_{\mathrm{curve}} from 0.35 to 0.65 reduces CEV and SIV PSNR by 1.603 and 0.583 dB, respectively, while improving HCI PSNR by 0.518 dB. This suggests that stronger curve regularization may benefit scenes with substantial local illumination contrast, while potentially over-constraining brightness adjustment in other categories. The edge-aware smoothness and white-preservation weights show comparatively smaller variations. Changing the three prior weights by ±30–50% keeps PSNR variations within 1.11 dB of the default everywhere, still outperforming the strongest baselines in SIV and HCI. The analysis also shows that our defaults weren’t tuned on the test scenes, for examaple λwhite=0.3\lambda_{\mathrm{white}} = 0.3 and λSH=1.2\lambda_{\mathrm{SH}} = 1.2 slightly beat the default in all categories.

Table B: Parameter sensitivity analysis across CEV, SIV, and HCI scenes (PSNR/SSIM/LPIPS).

ParameterSettingCEVSIVHCI
All parametersEvenSplat (default)19.526/0.708/0.30519.511/0.786/0.16916.950/0.682/0.241
Strongest baselines(PPISP, PPISP, 3DGS+CHROMA)17.794/0.723/0.26918.112/0.771/0.17813.780/0.522/0.352
λcurve\lambda_{\mathrm{curve}} (default = 0.5)0.3520.016/0.720/0.30119.868/0.800/0.16516.716/0.671/0.246
0.6518.413/0.745/0.26919.285/0.779/0.17017.234/0.691/0.239
λsmooth\lambda_{\mathrm{smooth}} (default = 1.0)0.519.284/0.709/0.29819.269/0.787/0.16717.268/0.685/0.245
1.518.957/0.691/0.30319.700/0.787/0.17016.801/0.675/0.244
λwhite\lambda_{\mathrm{white}} (default = 0.2)0.119.515/0.714/0.30319.508/0.786/0.16916.902/0.679/0.244
0.319.893/0.714/0.29719.519/0.787/0.16917.107/0.686/0.240
λSH\lambda_{\mathrm{SH}} (default = 0.8)0.418.784/0.686/0.30419.597/0.787/0.16816.855/0.680/0.242
1.220.049/0.723/0.29819.636/0.793/0.16716.961/0.683/0.242
(ρdark,ρbright)(\rho_{\mathrm{dark}},\rho_{\mathrm{bright}}) (default = (0.45,0.55))(0.35, 0.65)19.363/0.652/0.34218.352/0.713/0.22317.065/0.644/0.279
(0.5, 0.5)18.019/0.633/0.34218.470/0.715/0.22317.103/0.648/0.281
(Tdark,Tbright)(T_{\mathrm{dark}},T_{\mathrm{bright}}) (default = (0.25,0.65))(0.15, 0.75)19.028/0.650/0.34818.915/0.724/0.21716.697/0.653/0.275
(0.35, 0.55)16.674/0.659/0.33317.783/0.701/0.23216.977/0.622/0.297
γedge\gamma_{\mathrm{edge}} (default = 0.2)0.119.308/0.657/0.33518.375/0.713/0.22417.188/0.654/0.273
0.319.492/0.652/0.34518.337/0.712/0.22617.119/0.651/0.277
TwhiteT_{\mathrm{white}} (default = 95)9019.373/0.652/0.34718.414/0.712/0.22417.107/0.649/0.278
9818.562/0.650/0.34118.496/0.716/0.22216.995/0.637/0.282
schromas_{\mathrm{chroma}} (default = 2.4)1.618.919/0.649/0.34318.326/0.712/0.22517.086/0.658/0.271
3.219.450/0.662/0.33518.355/0.713/0.22417.155/0.657/0.277
ILEC start (default iter = 5000)400019.977/0.721/0.29719.217/0.786/0.16817.522/0.690/0.244
600019.610/0.703/0.30320.017/0.800/0.16616.145/0.649/0.254
ILEC regularization (default = 0.01)λmag=0.001\lambda_{\mathrm{mag}}=0.00119.914/0.714/0.29719.714/0.793/0.16816.741/0.671/0.246
λmag=0.1\lambda_{\mathrm{mag}}=0.119.320/0.707/0.29820.932/0.803/0.16815.944/0.631/0.259
λshrink=0.001\lambda_{\mathrm{shrink}}=0.00120.007/0.716/0.30019.813/0.793/0.16916.903/0.680/0.243
λshrink=0.1\lambda_{\mathrm{shrink}}=0.119.821/0.718/0.29320.483/0.798/0.16616.427/0.659/0.247
λpromote=0.001\lambda_{\mathrm{promote}}=0.00119.745/0.713/0.29719.754/0.791/0.16916.913/0.676/0.243
λpromote=0.1\lambda_{\mathrm{promote}}=0.119.448/0.707/0.29520.646/0.820/0.16315.521/0.619/0.263
3D illumination freeze (default iter = 3000)200015.817/0.620/0.33119.699/0.789/0.16916.596/0.660/0.249
500018.617/0.758/0.24019.198/0.783/0.16817.168/0.690/0.240
AuthorsOpenReview ↗

Rebuttal by Authors (continued)

W4. Equations in Sec. 3.1–3.3 need motivation, dimensions and definitions (normals, Eq. 2–5, r(t)r(t)).

The revision adds a notation table (symbol / dimension / definition) at the start of Sec. 3, one motivating sentence before each of Eq. 2–5, and the missing definitions: r(t)=clip((t−5000)/1000,0,1)r(t)=\mathrm{clip}\big((t-5000)/1000,0,1\big); Δloci∈Rh×w×1\Delta^i_{\mathrm{loc}}\in\mathbb{R}^{h\times w\times 1} with h=⌈H/16⌉h=\lceil H/16\rceil and w=⌈W/16⌉w=\lceil W/16\rceil, where HH and WW denote the height and width of the input image, respectively; nj\mathbf n_j = unit axis of the smallest Gaussian scale (A.2).

W5 / Q5. No seeds or variance; HCI rests on two scenes; removing normals or view directions costs ≈2.7 dB on CEV but nothing on SIV/HCI.

Table C reports three seeds across all nine real scenes, with EvenSplat PSNR standard deviations of 0.172/0.107/0.088 dB on CEV/SIV/HCI, respectively, indicating relatively small variation, particularly on SIV and HCI. Beyond the two real HCI scenes, we additionally evaluate two HCI-type scenes simulated from HDR-NeRF dataset, with qualitative comparisons shown in Fig. 17.

Regarding normals and view directions, both provide directional cues for geometry- and view-dependent appearance. In CEV, large exposure variations (±5 EV) may exacerbate view-dependent highlight clipping and appearance inconsistencies that the global CRN cannot fully compensate for, increasing reliance on these cues. In SIV/HCI, local illumination variations can be partially handled by image-specific compensation, particularly ILEC in HCI, reducing this dependence; nevertheless, removing both normals and view directions still decreases PSNR by 4.51/1.65/0.94 dB on CEV/SIV/HCI (Table D), confirming their contribution across all three categories.

Table C: Result stats from three different seeds on real scenes (PSNR mean ± std).

MethodCEVSIVHCI
EvenSplat19.2627 ± 0.172219.3937 ± 0.106517.0285 ± 0.0879
Strongest baseline (PPISP / PPISP / 3DGS+CHROMA)17.8063 ± 0.109518.9537 ± 0.028213.7315 ± 0.0628

Table D: Normal or view direction ablations on real scenes (PSNR/SSIM/LPIPS).

VariantCEVSIVHCI
EvenSplat (full)19.526/0.708/0.30519.511/0.786/0.16916.950/0.682/0.241
No Normals16.839/0.661/0.38319.506/0.782/0.17216.712/0.674/0.246
No View Directions16.916/0.665/0.36619.601/0.786/0.16516.745/0.670/0.247
No Normals, no View Directions15.013/0.487/0.49817.863/0.658/0.30316.011/0.598/0.331

W6 / Q3. Compare with LITA-GS, LL-Gaussian, Gaussian-DK, HDR-GS, SWAG.

Table E reports these comparisons on held-out views. HDR-GS could not be directly evaluated under its original exposure-conditioned protocol, as our dataset contains view-dependent illumination variations without the corresponding calibrated exposure information. SWAG is not included because its official code is unavailable.

Table E: Additional baselines comparison on real scenes (PSNR/SSIM/LPIPS).

MethodCEVSIVHCIEvaluation protocol
EvenSplat (Ours)19.526/0.708/0.30519.511/0.786/0.16916.950/0.682/0.241
Intrinsic-decomp + 3DGS11.916/0.398/0.65811.505/0.410/0.69412.975/0.410/0.6093DGS on Retinex-decomposition inputs (Careaga & Aksoy 2024)
3DGS+CHROMA10.193/0.530/0.41516.932/0.657/0.23913.780/0.522/0.352
LITA-GS14.515/0.329/0.67413.002/0.279/0.65110.482/0.266/0.632
Gaussian-DK9.930/0.321/0.38715.985/0.340/0.24211.703/0.318/0.353
LL-Gaussian11.732/0.229/0.80516.380/0.340/0.72111.944/0.305/0.456
HDR-GS–––Exposure protocol mismatch
SWAG–––Official code unavailable

Q4. Nine scenes, mostly small objects, single flash setting; no windows, mixed colour temperature, multiple lights, specular materials or moving shadows.

We agree. The heritage ceiling (Sec. 4.4) and the HDR-NeRF rooms are our indoor evidence. However, we find that capturing reference/ground-truth photos with uniform light for those natually unevenly illuminationed large indoor scene is extremely difficult based our real data capturing practice. The revision narrows the claim: validated on single-source flash and exposure variation; mixed colour temperatures and moving shadows are stated as future work in the new Limitations section.

Reviewer_FL8aOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes EvenSplat for Gaussian Splatting under exposure and illumination variation. The authors introduces 1. predict illumination for each Gaussian, for which rendered as illumination map. 2. A jointly optimized network predicts a scalar illumination map for each photograph 3. Losses align the image-space and rendered illumination maps 4. A camera-response network predicts a global RGB gain. A low-resolution local gain map, ILEC, handles remaining spatial exposure differences.

Critical strengths
  1. Combining image-space illumination priors with a shared 3D representation is well-motivated.
  2. The proposed method achieves the state-of-the-art results on most metrics in Table 1 across different scenes compared to other baseline methods.
  3. The ablation in Table 4 support the proposed components.
Critical weaknesses
  1. The decomposition priors depend on heuristics, such as adaptive curve regularization, edge-aware smoothness, and white preservation. It is worth discussing and experimenting how sensitive the performance is in response to chosen values for the heuristics.
  2. A qualitative comparison on ablation would be helpful to demonstrate the effectiveness of the proposed components.
  3. The limitation section should be expanded so that the readers understand the limitation boundary of the proposed method.
  4. How an ablation of no loss_couple would perform?
Questions for authors

Please see the comments above.

3 comments
AuthorsOpenReview ↗

Rebuttal by Authors

We thank reviewer FL8a for the positive assessment and concrete suggestions.

W1. Sensitivity to the heuristic values of the adaptive curve, edge-aware smoothness and white-preservation priors.

We conducted a one-at-a-time sensitivity analysis of the adaptive curve, edge-aware smoothness, and white-preservation priors, together with selected regularization weights and training schedules. The table below reports these results across CEV, SIV, and HCI scenes. The results reveal category-dependent sensitivity. For example, increasing the adaptive curve weight from 0.35 to 0.65 reduces PSNR from 20.016 to 18.413 dB on CEV and from 19.868 to 19.285 dB on SIV, while improving PSNR from 16.716 to 17.234 dB on HCI. In contrast, the edge-aware smoothness and white-preservation weights exhibit smaller variations within the tested ranges. Changing the three prior weights by ±30–50% keeps PSNR variations within 1.11 dB of the default everywhere, which still outperforms the strongest baselines in SIV and HCI. The analysis also shows that our defaults weren’t tuned on the test scenes, for exmaple λwhite=0.3\lambda_{\mathrm{white}} = 0.3 and λSH=1.2\lambda_{\mathrm{SH}} = 1.2 slightly beat the default in all three categories. Overall, these experiments highlight the importance of balancing illumination correction and appearance preservation. All default settings are shared across datasets without per-scene tuning. We also add a Table 9 in the revised paper for this sensitivity analysis.

Table: Parameter sensitivity analysis across CEV, SIV, and HCI scenes. Each entry reports PSNR/SSIM/LPIPS, averaged over the scenes in the corresponding category.

ParameterSettingCEVSIVHCI
All parametersEvenSplat (default)19.526/0.708/0.30519.511/0.786/0.16916.950/0.682/0.241
Strongest baselines(PPISP, PPISP, 3DGS+CHROMA)17.794/0.723/0.26918.112/0.771/0.17813.780/0.522/0.352
λcurve\lambda_{\mathrm{curve}} (default = 0.5)0.3520.016/0.720/0.30119.868/0.800/0.16516.716/0.671/0.246
0.6518.413/0.745/0.26919.285/0.779/0.17017.234/0.691/0.239
λsmooth\lambda_{\mathrm{smooth}} (default = 1.0)0.519.284/0.709/0.29819.269/0.787/0.16717.268/0.685/0.245
1.518.957/0.691/0.30319.700/0.787/0.17016.801/0.675/0.244
λwhite\lambda_{\mathrm{white}} (default = 0.2)0.119.515/0.714/0.30319.508/0.786/0.16916.902/0.679/0.244
0.319.893/0.714/0.29719.519/0.787/0.16917.107/0.686/0.240
λSH\lambda_{\mathrm{SH}} (default = 0.8)0.418.784/0.686/0.30419.597/0.787/0.16816.855/0.680/0.242
1.220.049/0.723/0.29819.636/0.793/0.16716.961/0.683/0.242
(ρdark,ρbright)(\rho_{\mathrm{dark}},\rho_{\mathrm{bright}}) (default = (0.45,0.55))(0.35, 0.65)19.363/0.652/0.34218.352/0.713/0.22317.065/0.644/0.279
(0.5, 0.5)18.019/0.633/0.34218.470/0.715/0.22317.103/0.648/0.281
(Tdark,Tbright)(T_{\mathrm{dark}},T_{\mathrm{bright}}) (default = (0.25,0.65))(0.15, 0.75)19.028/0.650/0.34818.915/0.724/0.21716.697/0.653/0.275
(0.35, 0.55)16.674/0.659/0.33317.783/0.701/0.23216.977/0.622/0.297
γedge\gamma_{\mathrm{edge}} (default = 0.2)0.119.308/0.657/0.33518.375/0.713/0.22417.188/0.654/0.273
0.319.492/0.652/0.34518.337/0.712/0.22617.119/0.651/0.277
TwhiteT_{\mathrm{white}} (default = 95)9019.373/0.652/0.34718.414/0.712/0.22417.107/0.649/0.278
9818.562/0.650/0.34118.496/0.716/0.22216.995/0.637/0.282
schromas_{\mathrm{chroma}} (default = 2.4)1.618.919/0.649/0.34318.326/0.712/0.22517.086/0.658/0.271
3.219.450/0.662/0.33518.355/0.713/0.22417.155/0.657/0.277
ILEC start (default iter = 5000)400019.977/0.721/0.29719.217/0.786/0.16817.522/0.690/0.244
600019.610/0.703/0.30320.017/0.800/0.16616.145/0.649/0.254
ILEC regularization (default = 0.01)λmag=0.001\lambda_{\mathrm{mag}}=0.00119.914/0.714/0.29719.714/0.793/0.16816.741/0.671/0.246
λmag=0.1\lambda_{\mathrm{mag}}=0.119.320/0.707/0.29820.932/0.803/0.16815.944/0.631/0.259
λshrink=0.001\lambda_{\mathrm{shrink}}=0.00120.007/0.716/0.30019.813/0.793/0.16916.903/0.680/0.243
λshrink=0.1\lambda_{\mathrm{shrink}}=0.119.821/0.718/0.29320.483/0.798/0.16616.427/0.659/0.247
λpromote=0.001\lambda_{\mathrm{promote}}=0.00119.745/0.713/0.29719.754/0.791/0.16916.913/0.676/0.243
λpromote=0.1\lambda_{\mathrm{promote}}=0.119.448/0.707/0.29520.646/0.820/0.16315.521/0.619/0.263
3D illumination freeze (default iter = 3000)200015.817/0.620/0.33119.699/0.789/0.16916.596/0.660/0.249
500018.617/0.758/0.24019.198/0.783/0.16817.168/0.690/0.240
AuthorsOpenReview ↗

Rebuttal by Authors (continued)

W2. A qualitative comparison of the ablations. W4. How an ablation of no LcoupleL_{\mathrm{couple}} would perform?

New Fig. 8 in the revised paper shows a qualitative comparison of the full model, no LcoupleL_{\mathrm{couple}}, no curve loss, and no ILEC on PlasticCart and RobotHC. For PlasticCart, we show the reconstructed appearance under the four variants. For RobotHC, we additionally show the corresponding rendered illumination maps L3DL_{\mathrm{3D}} beneath the reconstructed appearance. These examples visually support the quantitative trends in the Table below (summarized in revised Table 5 and detailed in Appendix Table 12): removing LcoupleL_{\mathrm{couple}}, ILEC, or the curve loss degrades appearance quality and weakens illumination modeling, with the full model producing the most balanced reconstruction.

LcoupleL_{\mathrm{couple}} is the only path from the image-space branch to the Gaussians (the three priors act on L2D,R2D\mathbf{L}_{2D},\mathbf{R}_{2D} only; Eq. 8), so the revised Table 5 row "No Image-Space Decomp." is the no-LcoupleL_{\mathrm{couple}} model: 15.92 / 19.22 / 12.01 vs. 19.53 / 19.51 / 16.95 (CEV/SIV/HCI). The finer split (alignment only, recombination only, one-way coupling) is in Appendix Table 12.

Table: Coupling and prior ablations on real scenes (PSNR/SSIM/LPIPS).

VariantCEVSIVHCI
EvenSplat (full)19.526/0.708/0.30519.511/0.786/0.16916.950/0.682/0.241
No LcoupleL_{\mathrm{couple}} (= revised Table 5 "No Image-Space Decomp.")15.924/0.646/0.37419.216/0.710/0.24212.012/0.438/0.324
No adaptive curve loss17.023/0.592/0.37219.369/0.752/0.20216.081/0.624/0.353
No ILEC19.599/0.707/0.30819.467/0.782/0.16814.066/0.561/0.266

W3. Expand the limitations.

We have expanded the Limitations section to clarify the scope and remaining challenges of EvenSplat. First, the intrinsic ambiguity between surface reflectance and illumination may cause over-correction of dark or bright materials, as illustrated in Fig. 16. Second, our sRGB-based formulation does not fully resolve saturation or complex lighting effects, including specular reflections and moving shadows. Third, the additional decomposition and illumination modeling inevitably introduce computational overhead, though still with high FPS for infrence, as quantified in the added Table 10. We explicitly acknowledge these limitations and identify broader lighting conditions and more robust intrinsic decomposition as directions for future work.

AuthorsOpenReview ↗

Comment by Authors

As the author-reviewer discussion period comes to an end, we wanted to briefly check in to see whether our responses have resolved your concerns. If anything remains unclear, we would be happy to offer further clarification or additional analysis. We thank you again for your time and for your thorough feedback.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

We sincerely thank the Area Chair for coordinating the review process and all three reviewers for their time, careful assessment, and constructive feedback. We appreciate the recognition of the motivation behind EvenSplat, the value of our real-world dataset with uniform-light references, and the design of coupling image-space decomposition with a shared Gaussian illumination field. The reviewers' comments have helped us substantially strengthen the experimental validation, clarify the methodological contributions, and provide a more comprehensive assessment of the proposed framework.

The main concerns raised by the reviewers and our corresponding revisions are summarized below:

Reviewer FL8a – Sensitivity to heuristic prior values, qualitative ablations, the no-coupling ablation, and limitations: We conducted a one-at-a-time sensitivity analysis of the prior weights, prior thresholds and targets, SH regularization, ILEC schedule and regularization, and illumination freeze iteration. Changing the three prior weights by ±30–50% keeps PSNR within 1.11 dB of the default and above the strongest baselines in every category, with all defaults shared across datasets without per-scene tuning. We added a qualitative ablation figure (full model, no Lcouple\mathcal L_{\mathrm{couple}}, no curve loss, no ILEC) with rendered illumination maps, clarified that "No Image-Space Decomp." is the no-Lcouple\mathcal L_{\mathrm{couple}} model, and expanded the Limitations section. These revisions are in the revised manuscript (Tables 5, 9, and 10; Figures 8 and 16; Limitations section).

Reviewer N64u – Novelty, source of the HCI gain, loss redundancy and sensitivity, notation, variance, baselines, and scope: We clarified that the coupling is bidirectional by design, isolated with a one-way (stop-gradient) ablation, and stated the novelty relative to L²DGS / LL-Gaussian: an online, per-scene 2D decomposition coupled in both directions, and the split between a transferable shared illumination field and non-transferable image-specific corrections (CRN, ILEC). We ran the requested full-model ablations without the curve loss, ILEC, both, and Lcouple\mathcal L_{\mathrm{couple}}, showing that the coupling is the largest single contributor on HCI (−4.94 dB when removed). We ablated alignment and recombination separately, showing complementary contributions, and added the sensitivity analysis above, a notation table in Sec. 3, motivating sentences for Eq. 2–5, and the missing definitions.

We further reported results over three random seeds (EvenSplat PSNR std ≤ 0.172 dB), added two HCI-type scenes simulated from HDR-NeRF, ablated normals and view directions individually and jointly, and compared against LITA-GS, LL-Gaussian, Gaussian-DK, and an offline intrinsic-decomposition + 3DGS baseline. We also added coloured-illumination and SH degree-0 ablations, the latter including the cross-lighting test (15.57 → 14.25 dB), a failure-case analysis, a clarification of the "Image-Space Decomp. Only" baseline, and fixes for all minor issues. These revisions are in the revised manuscript and supplementary material (Tables 1, 5, 9, 12, and 16; Figures 8, 16 and 17).

Reviewer 3uVs – Alignment ablation, evaluation protocol, behaviour of the curve prior, and efficiency: We ablated alignment only, recombination only, both removed (no Lcouple\mathcal L_{\mathrm{couple}}), and one-way coupling, confirming that the two coupling terms are complementary and that the bidirectional coupling is essential. To address the validity of real-world references, we additionally computed oracle per-image, per-channel affine-aligned metrics on HCI, where EvenSplat remains 2.26 dB ahead of the strongest baseline, indicating that the gain is not explained by global brightness alignment alone. We discussed how the adaptive curve can mis-correct intrinsically dark or bright materials, supported by failure cases and the λcurve\lambda_{\mathrm{curve}} sensitivity analysis, and reported computational cost: EvenSplat renders at 74.36 FPS, maintaining interactive performance. These revisions are in the revised manuscript and supplementary material (Tables 9, 10, and 12; Figure 8 and 16).

Overall, these revisions provide a more thorough evaluation of EvenSplat. They disentangle the contributions of the cross-domain coupling and the local corrections, analyse sensitivity to all major hyperparameters, assess variance across seeds and robustness to global brightness differences, compare against additional baselines, and analyse failure cases and computational cost. The major findings have been incorporated into the revised manuscript, while additional results are provided in the rebuttal.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.