acceptodds
Under review as a conference paper at ICLR 2027

UniOcean: Generative Gaussian Splatting for Unified Underwater 3D Reconstruction

Abstract

Underwater 3D reconstruction with Gaussian splatting has recently achieved remarkable progress. However, existing methods suffer from two critical limitations: (1) These methods typically rely on particle-specific modeling paradigms, limiting their adaptability across diverse underwater scenes with static, dynamic, and uncertain scattering conditions. (2) Underwater attenuation further exacerbates the inherent low-frequency bias of 3D Gaussians induced by their kernel distribution, leading to inferior representation of local high-frequency textures. To address these limitations, we propose UniOcean, a novel generative Gaussian splatting framework for unified and high-fidelity underwater 3D reconstruction. Specifically, we propose an Adaptive Particle Modeling strategy that exploits a learnable neural attenuation function to adaptively estimate depth-dependent absorption and scattering terms, thereby uniformly modeling diverse underwater particle effects. Furthermore, we design a Progressive Gaussian Generation module that leverages optimized coarse Gaussians to generate fine-grained texture Gaussians guided by the generative prior, thus yielding refined Gaussians featuring both global appearance and local textures. Benefiting from the above collaborative strategies, our method achieves high-quality representation while exhibiting excellent adaptability. Extensive experiments demonstrate that our UniOcean significantly outperforms existing methods across diverse underwater 3D scenes with static, dynamic, and varying synthetic scattering conditions.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Underwater 3D reconstruction, Gaussian splatting, Gaussian generation, Particle modeling

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 2 reviews, above 85% of the others.
Reviewer_rMGMOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes UniOcean, a generative Gaussian splatting framework for underwater 3D reconstruction under diverse scattering conditions. The method combines Adaptive Particle Modeling, which estimates absorption and scattering using a neural attenuation function, with Progressive Gaussian Generation, which generates additional texture Gaussians from coarse representations. A pretrained diffusion model provides guidance for recovering high-frequency details. Experiments on three underwater benchmarks demonstrate competitive reconstruction performance.

Critical strengths
  1. Relevant motivation. The paper addresses the challenges of underwater 3D reconstruction, particularly light attenuation, scattering, and loss of fine-grained textures.

  2. Reasonable framework design. Combining attenuation modeling, coarse-to-fine Gaussian representations, and generative guidance is a potentially useful approach to improving underwater reconstruction.

  3. Competitive experimental results. The method achieves strong PSNR, SSIM, and LPIPS results across three datasets. The component and loss ablations provide some evidence for the effectiveness of the proposed framework.

Critical weaknesses
  1. Weak theoretical motivation for progressive Gaussian generation. The paper attributes insufficient high-frequency reconstruction to an inherent low-frequency bias of Gaussian kernels. However, this explanation is not rigorously established, particularly since Gaussian splatting can represent high-frequency details through sufficiently small and dense primitives. Furthermore, the generated texture Gaussians reuse the positions of coarse Gaussians, raising questions about their ability to recover missing geometric details. More direct comparisons with established Gaussian densification and multiscale strategies may be needed to justify the proposed mechanism.

  2. Insufficient validation of generative guidance and reconstruction consistency. The method uses diffusion-refined novel views as auxiliary supervision, but diffusion models may introduce plausible yet inaccurate textures. Although the authors use LPIPS supervision to reduce this risk, no quantitative experiments directly assess hallucination, geometric consistency, or cross-view consistency. Moreover, the claim of handling dynamic underwater scenes is not supported by explicit temporal modeling or temporal-consistency evaluation. These issues weaken the claims regarding high-fidelity reconstruction across diverse conditions.

  3. Experimental and presentation limitations. The evaluation covers only 12 scenes, with limited analysis of challenging conditions such as severe turbidity and spatially varying attenuation. The efficiency comparison also shows that UniOcean requires more training time and memory and achieves lower rendering speed than several competing 3DGS methods. In addition, some methodological details and figures require clarification, particularly the generation and optimization of texture Gaussian attributes and the training procedure for generative guidance.

Questions for authors
  1. Why is generating additional texture Gaussians at existing coarse Gaussian positions preferable to conventional densification or multiscale Gaussian representations?

  2. How is the accuracy and multi-view consistency of diffusion-generated textures evaluated?

  3. Can the authors provide a more comprehensive analysis of reconstruction quality versus computational complexity, particularly compared with 3D-USE and 3D-UIR?

3 comments
AuthorsOpenReview ↗

Official Response by Authors – Part 1

Thank you for your constructive comments and for helping improve our paper. We have carefully considered and addressed them point by point in the following responses.

Weakness 1 & Question 1: Weak theoretical motivation for progressive Gaussian generation.

Answer: We thank the reviewer for these valuable comments. We provide detailed responses to these concerns as follows.

(1) Explanation of frequency-domain properties. We agree that 3DGS is capable of representing high-frequency details using sufficiently dense and appropriately scaled Gaussian primitives. We would also like to clarify that, due to the inherent low-frequency bias of Gaussian kernels, 3DGS requires more primitives to represent high-frequency details than methods using more suitable kernel functions. However, conventional densification strategies often struggle to meet this requirement due to gradient collisions [1] and improper coupling between Gaussian density and scale [2]. Therefore, 3D reconstruction kernels fundamentally determine the upper bound of densification [3]. This limitation becomes even more pronounced in underwater scenes, where high-frequency information is further suppressed by absorption and scattering effects, making effective Gaussian densification more challenging. These observations motivate our PGG, which generates fine-grained texture Gaussians rather than relying solely on conventional Gaussian densification strategies to learn high-frequency details.

(2) Why reuse the positions? We would like to clarify that reusing the positions of coarse Gaussians is a deliberate design choice, as these positions provide relatively accurate spatial information and are continuously refined during training. Based on these positions, our PGG generates fine-grained texture Gaussians with newly learned attributes, including scales, colors, and opacities, to enhance the representation of local high-frequency details. This design avoids learning Gaussian positions from scratch, thereby accelerating convergence and improving fine-grained texture generation.

(3) Comparison with existing densification strategies. To further validate the effectiveness of PGG, we conduct additional comparisons on the SeaThru-NeRF dataset against the original Gaussian densification strategy from 3DGS, the frequency regularization strategy from FreGS [4], and the multiscale densification strategy from ResGS [4]. As shown in Tab. R1 and Fig. 12 of the revised paper, simply increasing the number of Gaussians, introducing frequency regularization, or adopting multiscale densification does not achieve the same reconstruction quality as our PGG. This is mainly because these methods are primarily designed for reconstruction under relatively clear imaging conditions, without explicitly accounting for the severe visual degradation caused by underwater attenuation. In contrast, our PGG progressively generates fine-grained texture Gaussians, enabling more effective recovery of high-frequency details in underwater scenes. We have included this table and related discussion in the revised version.

Table R1: Ablation study of different densification strategies.

AblationPSNR ↑\uparrowSSIM ↑\uparrowLPIPS ↓\downarrow
w/ More Gaussians (3DGS)30.060.9220.126
w/ Frequency Regularization (FreGS)30.110.9230.123
w/ Multiscale Strategies (ResGS)29.250.9170.134
UniOcean (Ours)30.480.9280.117

[1] AbsGS: Recovering Fine Details for 3D Gaussian Splatting, ACM MM 2024.

[2] Frequency-Aware Density Control via Reparameterization for High-Quality Rendering of 3D Gaussian Splatting, AAAI 2025.

[3] Revisiting 3D Reconstruction Kernels as Low-Pass Filters, CVPR 2026.

[4] FreGS: 3D Gaussian Splatting with Progressive Frequency Regularization, CVPR 2024.

[5] ResGS: Residual Densification of 3D Gaussian for Efficient Detail Recovery, ICCV 2025.

AuthorsOpenReview ↗

Official Response by Authors – Part 2

Weakness 2 & Question 2: Insufficient validation of generative guidance and reconstruction consistency.

Answer: Thank you for your constructive comments. We address these concerns from the following two aspects.

(1) Validation of generative guidance and multi-view consistency. We would like to clarify that our UniOcean introduces an LPIPS-based perceptual loss to guide the recovery of fine-grained textures, while the reconstruction loss on the original observed views preserves the underlying scene structure and appearance. Moreover, the generative guidance is applied to a shared 3D Gaussian representation rather than independently optimizing individual views. This design helps reduce the risk of introducing hallucinated details while improving high-frequency texture reconstruction. As shown in Tab. 5 and Fig. 7 of the revised paper, removing the generative guidance leads to a 0.26 dB decrease in PSNR and introduces blurring artifacts in the rendered results. These results demonstrate the important contribution of our generative guidance to recovering fine-grained texture details. To further validate reconstruction consistency, we additionally evaluate the rendered clear views using warped LPIPS (wLPIPS) [4] and photometric reprojection loss (Ph-Loss) [5], measuring cross-view consistency from perceptual and photometric-structural aspects, respectively. As shown in Tab. R2, our UniOcean achieves lower values on both metrics across different scenes, indicating smaller cross-view consistency errors. We have included this table and related discussion in the revised version.

Table R2: Evaluation of multi-view consistency across different scenes.

MethodSea-NeRFU-IWS-U
wLPIPS ↓\downarrowPh-Loss ↓\downarrowwLPIPS ↓\downarrowPh-Loss ↓\downarrowwLPIPS ↓\downarrowPh-Loss ↓\downarrow
WaterSplatting0.4020.2270.4140.2320.3960.227
SeaSplat0.4130.2300.4260.2360.4040.231
3D-UIR0.4020.2210.4050.2280.3870.218
3D-USE0.3590.1940.4120.2300.3960.226
UniOcean (Ours)0.3570.1940.3820.2130.3710.206

(2) Evaluation of temporal consistency. We would like to clarify that our UniOcean does not explicitly model temporal deformation. Instead, we adopt a more general generative refinement strategy to suppress reconstruction artifacts caused by object motion and inconsistent observations. Similar benefits of diffusion-guided refinement have been demonstrated in previous studies, such as Splat4D [6]. To evaluate temporal consistency, we employ the Tem-Con metric following [7], which computes the average cosine similarity between CLIP features of consecutive frame pairs. As shown in Tab. R3, our method achieves higher similarity scores, indicating better temporal consistency. We have included this table and related discussion in the revised version.

Table R3: Evaluation of temporal consistency in U-IW scenes.

MethodTem-Con ↑\uparrow
WaterSplatting0.924
SeaSplat0.921
3D-UIR0.923
3D-USE0.926
UniOcean (Ours)0.927

[4] StylizedNeRF: Consistent 3D Scene Stylization as Stylized NeRF via 2D-3D Mutual Learning, CVPR 2022.

[5] Edit in 2D, Verify in 3D: Reinforcement Learning for Multi-view Consistent Scene Editing, ECCV 2026.

[6] Splat4D: Diffusion-Enhanced 4D Gaussian Splatting for Temporally and Spatially Consistent Content Creation, SIGGRAPH 2025.

[7] FateZero: Fusing Attentions for Zero-shot Text-based Video Editing, ICCV 2023.

AuthorsOpenReview ↗

Official Response by Authors – Part 3

Weakness 3 & Question 3: Experimental and presentation limitations.

Answer: We thank the reviewers for these constructive suggestions. Our detailed responses to the comments are provided as follows.

(1) Additional experiments on more challenging scenes. As suggested, we have conducted additional experiments on the DRUVA dataset, which includes severe turbidity and spatially varying attenuation. As reported in Tab. R4, our method outperforms competing methods on most evaluation metrics in reconstruction quality under these challenging conditions. Meanwhile, the updated Fig. 5 of the revised paper further demonstrates that our method more effectively mitigates underwater attenuation effects. These tables and related discussion are included in the revised version.

Table R4: Quantitative comparison of different methods on the DRUVA dataset.

MethodA2A4A5A13
PSNR ↑\uparrowSSIM ↑\uparrowLPIPS ↓\downarrowPSNR ↑\uparrowSSIM ↑\uparrowLPIPS ↓\downarrowPSNR ↑\uparrowSSIM ↑\uparrowLPIPS ↓\downarrowPSNR ↑\uparrowSSIM ↑\uparrowLPIPS ↓\downarrow
WaterSplatting26.580.8500.33729.630.8590.36426.220.8320.35433.070.9650.356
SeaSplat25.790.8530.34227.060.8570.38825.510.8460.34827.210.9570.417
Plenodium25.920.8550.32628.830.8610.35226.170.8480.34030.260.9610.352
3D-UIR28.650.8650.30430.580.8690.34327.350.8540.32234.430.9640.348
3D-USE26.100.8460.34429.390.8560.37125.990.8230.36832.740.9640.360
UniOcean (Ours)28.810.8670.29630.770.8750.33827.430.8580.31934.620.9650.340

(2) Analysis of reconstruction quality versus computational complexity. We acknowledge that UniOcean requires additional training time and GPU memory and achieves a lower rendering speed compared with several existing 3DGS-based methods, as shown in Tab. 12 of the paper. This is mainly because the generative guidance introduces additional computational costs during training, while the texture Gaussians increase the number of primitives to be optimized. Compared with 3D-USE and 3D-UIR, our method exhibits certain limitations in computational efficiency. Nevertheless, these additional computational costs enable UniOcean to recover finer texture details and achieve higher reconstruction quality. Considering the trade-off between reconstruction quality and computational efficiency, we believe that the additional computational overhead is acceptable, as efficiency limitations can be partially alleviated by more powerful hardware, whereas improvements in reconstruction quality fundamentally rely on effective methodological designs. We have included this discussion in the revised version.

(3) Clarification of methodological details. 1) Regarding the generation and optimization of texture Gaussian attributes, our PGG first employs multi-resolution hash encoding to extract multi-scale spatial features, as detailed in Eqs. (24)–(26) and Fig. 10. These features are then modulated by the attributes of coarse Gaussians and fed into a multi-head MLP to predict the scales, colors, and opacities of texture Gaussians, as formulated in Eq. (11). During optimization, rather than independently optimizing each texture Gaussian, we update the parameters of the MLPs through backpropagation of the rendering losses, thereby progressively refining the generated Gaussian attributes. 2) Regarding the training procedure for generative guidance, we first synthesize unseen pseudo-views by perturbing the camera poses, as formulated in Eq. (16). These pseudo-views are then fed into a diffusion model to suppress reconstruction artifacts and generate fine-grained details, as described in Eq. (17). Subsequently, the diffusion-refined views provide perceptual guidance through an LPIPS-based loss to optimize the shared 3D Gaussian representation via backpropagation, as formulated in Eq. (21). The collaborative optimization of these two components further facilitates the generation and refinement of texture Gaussians.

Thank you again for your valuable feedback. We have revised the paper to incorporate all related experiments and detailed discussions as you suggested. If there are any additional issues, we are happy to provide further explanations.

Reviewer_mmgdOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents UniOcean, a generative Gaussian splatting framework for underwater 3D reconstruction under diverse scattering conditions. The proposed method consists of two main components: Progressive Gaussian Generation (PGG) and Adaptive Particle Modeling (APM). PGG aims to improve high-frequency texture representation by combining coarse Gaussian optimization with texture Gaussian generation, while APM employs a learnable neural attenuation function to estimate absorption and scattering effects. Experiments on three underwater datasets demonstrate competitive reconstruction performance.

Critical strengths
  1. The paper focuses on an important problem in underwater 3D reconstruction, where complex absorption and scattering effects significantly degrade reconstruction quality.
  2. The combination of adaptive medium modeling and progressive Gaussian generation provides a reasonably comprehensive framework for underwater scene reconstruction.
Critical weaknesses
  1. The paper motivates PGG by arguing that Gaussian primitives inherently exhibit a low-frequency bias, which limits their ability to represent high-frequency scene details. However, the frequency-domain analysis in Figure 2 primarily illustrates the properties of individual Gaussian functions, which may not directly reflect the representation capacity of an entire 3D Gaussian Splatting model. Furthermore, the connection between this frequency-domain analysis and the claimed limitations of conventional Gaussian densification remains unclear.
  2. In Lines 95–97, the paper claims to be the first to address diverse underwater conditions through a unified generative paradigm. However, the distinction from existing underwater reconstruction frameworks could be more clearly established. For example, WaterSplatting [1] already combines 3D Gaussian primitives with a learnable volumetric medium representation to jointly model scene geometry and underwater scattering effects. More recently, UDF [2] further introduces a unified underwater field that jointly models dynamic scene geometry and time-varying medium properties.

[1] WaterSplatting: Fast Underwater 3D Scene Reconstruction Using Gaussian Splatting. 3DV 2025. [2] Uncertainty-Aware 3D Reconstruction for Dynamic Underwater Scenes. ICLR 2026.

Questions for authors
  1. Could the authors provide further evidence that 3D Gaussian Splatting struggles to represent high-frequency details, beyond the analysis of individual Gaussian kernels in Figure 2?
  2. In Eq. (1), are the attenuation coefficients βD\beta^D and βB\beta^B scalars or RGB vectors?
  3. In Eq. (16), how are adjacent views selected, camera rotations interpolated, and perturbations determined? How is the multi-view consistency of diffusion-generated details ensured?
  4. Could the authors use consistent capitalization for "Gaussian Splatting" and "Gaussian splatting" throughout the paper?
  5. Could Figure 4 be placed closer to its corresponding discussion to improve readability?
  6. Please correct typos, such as “appearanc” and “chear”.
5 comments
AuthorsOpenReview ↗

Official Response by Authors – Part 1

Thank you for your insightful comments and for helping us improve our work. We have carefully considered your comments and respond to them point by point as follows.

Weakness 1 & Question 1: Relationship between frequency-domain analysis and the 3DGS model.

Answer: We thank the reviewer for this insightful comment. We provide detailed responses regarding the frequency-domain properties as follows.

(1) Clarification of frequency-domain analysis. We agree that the frequency-domain properties of individual Gaussian functions do not directly determine the representation capacity of the entire 3DGS model, as 3DGS can represent high-frequency details through sufficient Gaussian densification. Although Fig. 2 illustrates the frequency-domain properties of individual Gaussian functions, these properties remain relevant to the overall 3DGS representation, as the rendered scene is constructed through the aggregation of Gaussian primitives. Our frequency-domain analysis is intended to explain this underlying relationship.

(2) Connection between frequency-domain analysis and conventional Gaussian densification. Due to the inherent low-pass properties of Gaussian functions, 3DGS require a larger number of densely distributed and appropriately scaled primitives to represent high-frequency details compared with a more suitable kernel function. However, conventional densification strategies often struggle to satisfy this requirement due to gradient collisions [1] and improper linking between Gaussian density and scale [2]. This limitation becomes even more pronounced in underwater scenes, where high-frequency information is further suppressed by absorption and scattering effects. Therefore, the kernel functions fundamentally determine how effectively a 3D model can densify its primitives to capture high-frequency details. This is further supported by the findings reported in [3], which highlight the inherent limitations of Gaussian kernels. Such a connection motivates our PGG, which generates fine-grained texture Gaussians rather than relying solely on conventional Gaussian densification strategies to learn high-frequency details.

(3) Further evidence of the representation limitations of 3DGS. Beyond the analysis of individual Gaussian kernels in Fig. 2, we further provide visualizations of the Gaussian primitives and their corresponding rendered images at different training iterations of the original 3DGS. As shown in Fig. 17 of the revised paper, 3DGS gradually captures high-frequency details by adjusting the density and scale of Gaussian primitives as training progresses. However, the effectiveness of conventional densification gradually diminishes after 15,000 iterations. In particular, in grass regions containing dense high-frequency textures, some Gaussian primitives remain insufficiently densified even at the end of training, resulting in noticeable blurring in the rendered images. This observation is closely related to the inherent low-pass property of Gaussian kernels, which fundamentally determines the upper bound of densification. Furthermore, previous studies have provided additional evidence of the limitations of conventional Gaussian densification strategies in recovering high-frequency details [1-3].

We have summarized the above discussion and included it in the revised version.

[1] AbsGS: Recovering Fine Details for 3D Gaussian Splatting, ACM MM 2024.

[2] Frequency-Aware Density Control via Reparameterization for High-Quality Rendering of 3D Gaussian Splatting, AAAI 2025.

[3] Revisiting 3D Reconstruction Kernels as Low-Pass Filters, CVPR 2026.

AuthorsOpenReview ↗

Official Response by Authors – Part 2

Weakness 2: Distinctions from existing underwater reconstruction frameworks.

Answer: Thank you for this constructive suggestion. Our UniOcean fundamentally differs from these methods in both its problem formulation and generative paradigm. 

(1) Problem formulation. WaterSplatting and UDF formulate underwater reconstruction from the perspective of the physical image formation process, modeling relevant physical variables to represent underwater optical effects:

I(x)=J(x)⋅e−βD⋅z(x)+B∞⋅(1−e−βB⋅z(x)).I(x) = J(x) \cdot e^{-\beta^D \cdot z(x)} + B^\infty \cdot \bigl(1 - e^{-\beta^B \cdot z(x)}\bigr).

In contrast, our UniOcean revisits the overall effects of underwater attenuation and formulates underwater attenuation as a unified modeling problem involving absorption and scattering effects by optimizing a learnable neural attenuation function:

I(x)=J(x)⊙A(x)+S(x).I(x)= J(x) \odot A(x) + S(x).

Therefore, rather than relying on explicit modeling of individual physical variables, we adopt a more unified and adaptive formulation to model diverse underwater attenuation effects.

(2) Generative paradigm. Unlike WaterSplatting and UDF, which primarily optimize underwater medium representations through reconstruction-based objectives, our UniOcean not only adaptively models diverse underwater particle effects but also recovers high-frequency scene details through a generative paradigm. As illustrated in Fig. 4, of the paper, our method achieves superior reconstruction quality with more faithful scene details.

Overall, our UniOcean is distinct from existing methods in both its problem formulation and generative paradigm, offering clear advantages in underwater attenuation modeling and high-frequency detail reconstruction.

AuthorsOpenReview ↗

Official Response by Authors – Part 3

Question 2: In Eq. (1), are the attenuation coefficients βD\beta^D and βB\beta^B scalars or RGB vectors?

Answer: Thank you for raising this important concern. Both βD\beta^D and βB\beta^B are RGB vectors representing channel-dependent attenuation coefficients for direct transmission and backscattering, respectively. We have included this clarification in the revised version.

Question 3: Clarification of pseudo-view synthesis and multi-view consistency.

Answer: We thank the reviewer for raising this important question. We provide detailed explanations of these two aspects as follows.

(1) Pseudo-view synthesis. To synthesize pseudo-views, we first calculate the Euclidean distances between the camera centers of the training views and select the two closest views as adjacent views. This selection ensures similar camera positions, providing a suitable basis for synthesizing a virtual viewpoint through camera pose interpolation. We then calculate the average camera rotation of the two selected views and use it as the interpolated rotation for the synthesized viewpoint. The perturbations Δt\Delta t are randomly sampled from a Gaussian distribution N(0,δ)\mathcal{N}(0, \delta) at each training iteration, and the selected viewpoint is removed from the sampling pool to increase the diversity of synthesized viewpoints.

(2) Multi-view consistency. We would like to clarify that the diffusion-generated outputs are not directly used as strong supervision for clear-view rendering. Instead, we employ an LPIPS-based perceptual loss to provide generative guidance for optimizing a shared 3D Gaussian representation, thereby encouraging consistent appearance across different viewpoints. Meanwhile, the reconstruction loss on the original underwater views constrains the optimization process to preserve the observed scene structure and appearance, further promoting multi-view consistency. Furthermore, we employ warped LPIPS (wLPIPS) [4] and photometric reprojection loss (Ph-Loss) [5] to evaluate the multi-view consistency of the rendered clear views in terms of perceptual and photometric-structural consistency, respectively. As shown in Tab. R1, our UniOcean achieves lower values on both metrics across different scenes, indicating smaller cross-view consistency errors. We have included this table and related discussion in the revised version.

Table R1: Evaluation of multi-view consistency across different scenes.

MethodSea-NeRFU-IWS-U
wLPIPS ↓\downarrowPh-Loss ↓\downarrowwLPIPS ↓\downarrowPh-Loss ↓\downarrowwLPIPS ↓\downarrowPh-Loss ↓\downarrow
WaterSplatting0.4020.2270.4140.2320.3960.227
SeaSplat0.4130.2300.4260.2360.4040.231
3D-UIR0.4020.2210.4050.2280.3870.218
3D-USE0.3590.1940.4120.2300.3960.226
UniOcean (Ours)0.3570.1940.3820.2130.3710.206

[4] StylizedNeRF: Consistent 3D Scene Stylization as Stylized NeRF via 2D-3D Mutual Learning, CVPR 2022.

[5] Edit in 2D, Verify in 3D: Reinforcement Learning for Multi-view Consistent Scene Editing, ECCV 2026.

Question 4&5&6: Minor presentation and typographical corrections.

Answer: Thank you for your careful reading and valuable suggestions. We have carefully revised the manuscript to address these minor issues. Specifically, we have standardized the capitalization of "Gaussian splatting" throughout the paper, moved Fig. 4 closer to its corresponding discussion to improve readability, and corrected the typographical errors.

Thank you again for your insightful feedback. We have incorporated all relevant experiments and discussions into the revised paper as you suggested. If you have any further questions, we are pleased to provide more details.

Comments on the paper

9 comments
AuthorsOpenReview ↗

Official Response by Authors – Part 1

Thank you for your valuable comments and for your contribution to improving our paper. We have carefully considered your feedback and respond point by point as follows.

Weakness 1: Clear appearance and physical decomposition lack direct validation.

Answer: Thank you for your constructive feedback. We address this concern from the following two aspects.

(1) Validation of clear appearance restoration. We agree that reconstruction metrics alone cannot fully demonstrate the accuracy of restored water-free colors and textures. However, since our experiments are primarily conducted on real-world underwater datasets without corresponding clean reference images, full-reference evaluation of restoration quality on these datasets is infeasible. To address this limitation, we employ no-reference metrics to quantitatively assess the visual quality of restored clear views. As reported in Tab. R1, our method generally outperforms existing methods, indicating better overall perceptual quality of the restored clear views. As suggested by the reviewer, we also evaluate restoration quality on synthetic underwater scenes (S-U) with available clean references using full-reference metrics, providing direct validation of restoration quality. We have included this table and the related discussion in the revised version.

Table R1: Quantitative comparison of restoration quality across different scenes.

MethodSea-NeRFU-IWDRUVAS-U
URanker ↑\uparrowMUSIQ ↑\uparrowURanker ↑\uparrowMUSIQ ↑\uparrowURanker ↑\uparrowMUSIQ ↑\uparrowPSNR ↑\uparrowSSIM ↑\uparrowLPIPS ↓\downarrow
WaterSplatting1.36158.520.57357.62-0.87733.6923.610.8020.278
SeaSplat1.55554.470.89555.090.12834.0225.150.8130.261
Plenodium1.57356.221.07655.430.40532.9725.280.8150.260
3D-UIR1.65854.541.14855.460.83932.1226.040.8260.257
3D-USE1.67959.881.08356.810.44632.1825.720.8210.259
UniOcean (Ours)1.68355.161.15257.940.91334.3926.350.8360.244

(2) Underlying mechanism of physical decomposition. We would like to clarify that our modeling of absorption and scattering effects follows the Beer–Lambert exponential attenuation law, as shown in Fig. 8. Specifically, we parameterize each physical component using learnable variables to construct a neural attenuation function that uniformly models absorption and scattering effects. For the absorption component, we combine features encoded from depth information and appearance embeddings with learnable parameters to form the absorption coefficients βa\beta_a, followed by an exponential mapping to obtain the absorption term:

A^(x)=exp⁡(−βa(x)⋅D^(x)).\hat{A}(x)=\exp(-\beta_a(x)\cdot\hat{D}(x)).

For the scattering component, we first parameterize the depth-dependent scattering coefficients βB\beta_B and βJ\beta_J through two independent learnable mappings, and then jointly optimize them with the corresponding learnable coefficients θB\theta_B and θJ\theta_J within an exponential formulation to model the scattering contribution:

S^(x)=θB⋅(1−exp⁡(−βB))+θJ⋅exp⁡(−βJ)).\hat{S}(x)= \theta_B\cdot\left(1-\exp(-\beta_B)\right) +\theta_J\cdot\exp\left(-\beta_J)\right).

Therefore, this physics-guided modeling mechanism helps reduce physically implausible compensation and promotes more accurate physical decomposition.

AuthorsOpenReview ↗

Official Response by Authors – Part 2

Weakness 2 & Question 3: Novel-view evaluation and consistency cannot yet be fully assessed.

Answer: We thank the reviewer for these valuable comments. We have carefully considered all the comments and provided detailed responses below.

(1) Training/test view split. Following the original 3DGS evaluation protocol, we select every eighth image as a test view and use the remaining images for training. No separate validation set is constructed, maintaining consistency with the established evaluation protocol and ensuring fair comparisons with existing methods.

(2) Whether Gaussian attributes vary across views. We would like to clarify that the rotations and scales of texture Gaussians remain unchanged across different camera viewpoints. Specifically, the camera pose in Eq. (11) is used to provide spatial guidance for more accurate prediction of Gaussian attributes and enhance their spatial correlation, rather than introducing view-dependent variations in these attributes. Once optimization is completed, all Gaussian parameters remain fixed in a shared 3D representation. During inference, novel views are directly rendered from the optimized Gaussians using the target camera parameters, without updating Gaussian attributes. Therefore, the rotations and scales of texture Gaussians remain unchanged, allowing all rendered views to share the same underlying geometry and texture representation.

(3) Use of DIFIX3D+. We directly employ the publicly available Difix3D+ checkpoint for refinement, as it effectively suppresses rendering artifacts and blurring. Specifically, Difix3D+ is fine-tuned from SD-Turbo [1] following a strategy similar to Pix2Pix-Turbo [2], with a frozen VAE encoder and a LoRA-based fine-tuned decoder. For training, Difix3D+ constructs 80,000 noisy-clean image pairs from the DL3DV [3] dataset, where the noisy images contain diverse rendering artifacts and blurring, while the corresponding clean images serve as supervision targets.

(4) Appearance embedding in absorption modeling. We would like to clarify that the appearance embedding is represented by γ(ζ)\gamma(\zeta) in Eq. (15) and is not omitted from the formulation. Specifically, it is derived from the camera pose through positional encoding and serves as a conditioning feature for the absorption network, rather than being directly extracted from image appearances. Notably, appearance embeddings are only used during training to model the absorption component. Once optimization is completed, all Gaussian parameters are fixed and shared across different viewpoints. During inference, novel views are directly rendered from the optimized Gaussian representation using the target camera parameters, without regenerating appearance embeddings. Therefore, no appearance embeddings need to be estimated for novel views, and no evaluation images are accessed during inference.

[1] Adversarial diffusion distillation, ECCV 2024.

[2] One-step image translation with text-to-image models, arXiv 2024.

[3] Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision, CVPR 2024.

AuthorsOpenReview ↗

Official Response by Authors – Part 3

Question 1: Clarification of experimental details and evaluation of restoration quality.

Answer: Thank you for your valuable questions. (1) We would like to clarify that the metrics reported in Tables 1–3 evaluate the reconstructed underwater images, consistent with existing studies. (2) Following the original 3DGS evaluation protocol, we select every eighth image as a test view and use the remaining images for training, without a separate validation set. (3) As suggested, we have conducted additional experiments on the S-U dataset using clean reference images to quantitatively evaluate the restoration quality of our method. Please refer to our response to Weakness 1 and the results reported in Tab. R1 for more details.

Question 2: View dependence of Gaussian attributes and evaluation of multi-view consistency.

Answer: We thank the reviewer for raising this important question. (1) During optimization, the camera pose corresponding to each training view is provided for rendering and guiding the prediction of Gaussian attributes in Eq. (11). During inference, novel views are directly rendered from the optimized Gaussian representation using the target camera pose, without re-executing Eq. (11) or updating Gaussian attributes. Please refer to our response to Weakness 2 for more details. (2) To further validate multi-view consistency, we additionally evaluate the rendered clear views using warped LPIPS (wLPIPS) [4] and photometric reprojection loss (Ph-Loss) [5], measuring cross-view consistency from perceptual and photometric-structural aspects, respectively. As shown in Tab. R2, our UniOcean achieves lower values on both metrics across different scenes, indicating smaller cross-view consistency errors. We have included this table and related discussion in the revised version.

Table R2: Evaluation of multi-view consistency across different scenes.

MethodSea-NeRFU-IWS-U
wLPIPS ↓\downarrowPh-Loss ↓\downarrowwLPIPS ↓\downarrowPh-Loss ↓\downarrowwLPIPS ↓\downarrowPh-Loss ↓\downarrow
WaterSplatting0.4020.2270.4140.2320.3960.227
SeaSplat0.4130.2300.4260.2360.4040.231
3D-UIR0.4020.2210.4050.2280.3870.218
3D-USE0.3590.1940.4120.2300.3960.226
UniOcean (Ours)0.3570.1940.3820.2130.3710.206

[4] StylizedNeRF: Consistent 3D Scene Stylization as Stylized NeRF via 2D-3D Mutual Learning, CVPR 2022.

[5] Edit in 2D, Verify in 3D: Reinforcement Learning for Multi-view Consistent Scene Editing, ECCV 2026.

Thanks again for your constructive comments. We have addressed your comments by including all pertinent experiments and clarifications in our revision. If you need further explanations, we are happy to elaborate in more detail.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.