DualPhys-GS: Underwater 3D Gaussian Splatting

• The paper introduces DualPhys-GS, a 3D Gaussian Splatting framework that decouples underwater image degradation into attenuation and scattering for high-quality reconstruction.
• It employs a dual-path optimization mechanism, using RGB-guided attenuation and multi-scale depth-aware scattering to tackle wavelength-selective absorption and particle-induced blur.
• Empirical evaluations on SeaThru-NeRF and SaltPond datasets show competitive PSNR, SSIM, and LPIPS metrics, with efficient 0.016s per frame rendering speed.

Searching arXiv for DualPhys-GS and closely related Gaussian-splatting/physics papers to ground the article. {"query":"DualPhys-GS underwater scene reconstruction 3D Gaussian Splatting arXiv (Li et al., 13 Aug 2025)", "max_results": 5} Searching for the exact DualPhys-GS paper and related Gaussian Splatting papers. I’m going to look up the cited arXiv papers to ensure the article is grounded in the latest preprints. DualPhys-GS is a 3D Gaussian Splatting (3DGS)-based framework for underwater scene reconstruction that addresses failure modes specific to the water medium, notably the selective attenuation of light wavelengths and the effect of suspended particle scattering. It introduces a dual-path optimization mechanism in which attenuation and scattering are modeled separately but optimized jointly, together with several physically-guided loss terms and a scene adaptive mechanism that automatically identifies water-body-type characteristics and adjusts optimization accordingly. The stated goal is high-quality underwater reconstruction in settings where atmospheric optical models lead to color distortion, geometric artifacts, and collapsing phenomena at long distances (Li et al., 13 Aug 2025).

1. Problem formulation and domain assumptions

DualPhys-GS is motivated by the observation that underwater image formation differs materially from atmospheric imaging. The framework is built around two physical effects emphasized in the underlying work: wavelength-selective attenuation and suspended particle scattering. According to the paper, traditional methods based on atmospheric optical models cannot effectively handle these effects, and the resulting errors manifest as color distortion, geometric artifacts, and collapsing phenomena at long distances (Li et al., 13 Aug 2025).

The method is therefore specialized to underwater reconstruction rather than being a generic 3DGS refinement layer. Its scene model explicitly distinguishes attenuation-related degradation from scattering-related degradation. This separation is central to the architecture: the attenuation path is guided by RGB imagery and depth, while the scattering path is modeled as a multi-scale, depth-aware process with feature aggregation and attention. The paper also frames water-body variability as a first-class issue, citing examples such as clear coral reef waters and turbid coastal waters, and introduces automatic adaptation rather than fixed global coefficients.

A plausible implication is that DualPhys-GS treats underwater reconstruction as a physically structured inverse problem rather than as purely photometric fitting. In that sense, its use of “physically-guided” refers to optical transport phenomena in water, not to rigid-body, fluid, or deformation mechanics.

2. Dual-path image formation model

The framework decouples underwater image degradation into attenuation and scattering and writes the observed image as

$$\boldsymbol{I} = \boldsymbol{J} \cdot \boldsymbol{A}(\boldsymbol{D}, \boldsymbol{I}_{\mathrm{RGB}}) + \boldsymbol{B}(\boldsymbol{D})$$

where $\boldsymbol{I}$ is the observed underwater image, $\boldsymbol{J}$ is the latent scene radiance, $\boldsymbol{A}$ is the attenuation map, $\boldsymbol{D}$ is depth, and $\boldsymbol{B}$ is the scattering map (Li et al., 13 Aug 2025).

This factorization defines the method’s dual-path optimization mechanism. The attenuation branch uses RGB features and depth information, with explicit edge-guided modulation. The scattering branch uses multi-scale depth features and attention mechanisms to represent spatially varying scattering behavior. The paper states that attenuation and scattering are decoupled and addressed via different features: attenuation uses RGB and depth with edge-aware modulation, whereas scattering uses multi-scale depth features and edge awareness.

This decomposition has methodological significance because the two terms encode different failure modes. Attenuation is tied to wavelength loss and color shifting over distance; scattering is tied to anisotropic, depth-dependent, and spatially variant degradation, especially in the presence of suspended matter. The architecture therefore does not ask one latent field to explain both phenomena simultaneously. This suggests a reduction in ambiguity during optimization, particularly in long-range and turbid scenes.

3. RGB-guided attenuation optimization

The attenuation branch is motivated by wavelength dependence in underwater propagation, especially the faster absorption of red light relative to blue and green. DualPhys-GS extends a depth-based exponential attenuation model with feature-based enhancement:

$$\boldsymbol{A}(\boldsymbol{D}, \boldsymbol{I}_{\mathrm{RGB}}) = \exp \left( -\sum_{c \in \{r,g,b\}} \boldsymbol{w}_c \cdot \beta_c(\boldsymbol{D}) \cdot (1 - \gamma \cdot E(\boldsymbol{I}_{\mathrm{RGB}}, \boldsymbol{D})) \cdot T(\boldsymbol{I}_{\mathrm{RGB}}) \cdot \boldsymbol{D} \right)$$

Here, $\boldsymbol{w}_c$ denotes channel-wise attenuation weights, $\beta_c(\boldsymbol{D})$ the base attenuation coefficient per channel, $E(\boldsymbol{I}_{\mathrm{RGB}}, \boldsymbol{D})$ an edge awareness factor, $\boldsymbol{I}$0 the edge modulation strength, and $\boldsymbol{I}$1 a texture factor (Li et al., 13 Aug 2025).

The paper describes the RGB feature extractor as learning to identify color-consistent regions and boundaries, thereby informing the model about regions likely to require correction due to wavelength loss. The edge-aware mechanism incorporates information from RGB and depth discontinuities, with the stated purpose of preventing artifacts and improving estimates near scene boundaries such as object contours. The reported benefit is mitigation of color loss at long distances, especially in visually complex regions including coral reefs and sharp object boundaries.

In functional terms, the attenuation model is not only depth-conditioned but also feature-modulated. This distinguishes it from a fixed Beer–Lambert-style formulation. The method therefore embeds local structural and texture cues directly into the attenuation field, rather than treating attenuation as a purely scalar function of path length.

4. Multi-scale depth-aware scattering model

The scattering branch targets depth-dependent and anisotropic scattering arising from suspended particles. The paper argues that a single-scale model is inadequate because near-field detail and far-field background fog exhibit different spatial characteristics. To address this, DualPhys-GS uses a three-scale feature pyramid with attention:

$\boldsymbol{I}$2

where features are extracted at original, $\boldsymbol{I}$3, and $\boldsymbol{I}$4 resolutions and fused by an attention module (Li et al., 13 Aug 2025).

The scattering map is then written as

$\boldsymbol{I}$5

where $\boldsymbol{I}$6 is the theoretical background color at infinity, $\boldsymbol{I}$7 is the base scattering coefficient, $\boldsymbol{I}$8 is confidence in local depth, $\boldsymbol{I}$9 is edge weighting, and $\boldsymbol{J}$0 is the modulation for multi-scale features (Li et al., 13 Aug 2025).

The attention module is described as combining channel attention and spatial attention. Channel attention weights physically salient features such as blue-green dominance, edge gradients, and color saturation related to water turbidity, while spatial attention focuses on areas with pronounced scattering transitions. The reported advantage is that the multi-scale architecture preserves fine scene details while correctly simulating global scattering transitions, thereby overcoming blurring and geometric collapse in prior approaches.

This branch is therefore both scale-aware and structure-aware. The model does not merely estimate a global haze term; it constructs a scattering field conditioned on depth hierarchy, edge structure, and learned feature salience.

5. Loss design and scene-adaptive optimization

DualPhys-GS supplements its dual-branch architecture with several physically-guided loss functions. The attenuation-scattering consistency loss is defined as

$\boldsymbol{J}$1

and is motivated by the requirement that scattering should correlate positively with depth while attenuation should correlate negatively with depth. The third term enforces that scattering and transmittance sum to $\boldsymbol{J}$2, described as energy conservation (Li et al., 13 Aug 2025).

The water body type adaptive loss is

$\boldsymbol{J}$3

where $\boldsymbol{J}$4 and $\boldsymbol{J}$5 are dynamically set by water type $\boldsymbol{J}$6. The paper states that this automates weighting adjustment for clear and turbid waters and removes manual parameter tuning (Li et al., 13 Aug 2025).

The edge-aware scattering loss is written as

$\boldsymbol{J}$7

with higher weights near large depth gradients, in order to preserve sharp structural edges and reduce blur-related artifacts (Li et al., 13 Aug 2025).

The multi-scale feature loss is

$\boldsymbol{J}$8

and is introduced to preserve both global structure and local detail by matching outputs and targets across scales (Li et al., 13 Aug 2025).

The scene adaptive mechanism complements these losses with automatic water type recognition:

$\boldsymbol{J}$9

The resulting classification governs an adaptive optimization controller. The paper states that learning rates are reduced for clear water and increased for turbid water; attenuation is weighted higher in clear water and scattering in turbid water; and all adjustments are data-driven and determined solely by the input images (Li et al., 13 Aug 2025).

6. Empirical evaluation, ablations, and reported performance

The evaluation uses the SeaThru-NeRF and SaltPond datasets and reports PSNR, SSIM, and LPIPS as primary metrics (Li et al., 13 Aug 2025). The paper states that DualPhys-GS is consistently among the best in the 3DGS-based category for SSIM and LPIPS and achieves the best PSNR on challenging SaltPond scenes. It further states that the method outperforms existing methods in several metrics, especially in suspended matter-dense regions and long-distance scenes.

The comparative discussion is differentiated rather than uniformly favorable. WaterSplatting is reported to achieve the highest PSNR but to fail in structural consistency and edge preservation, especially in turbid or extreme settings; DualPhys-GS is said to excel especially where scattering dominates, including dense suspended particles and long-range scenes. UW-GS is reported to achieve marginally superior raw PSNR with access to external depth priors, whereas DualPhys-GS achieves competitive performance without any external depth models, which the paper attributes to its physical constraints (Li et al., 13 Aug 2025).

Efficiency is also part of the evaluation. The reported rendering time is $\boldsymbol{A}$0 per frame, described as substantially faster, specifically $\boldsymbol{A}$1, than WaterSplatting (Li et al., 13 Aug 2025). Qualitatively, the method is reported to reconstruct clear, undistorted geometry far from the camera, preserve red channel content and edge sharpness, and maintain strong performance across scenes ranging from pristine to heavily turbid.

The ablation studies attribute performance gains to multiple components rather than to a single module. The paper states that both RGB-guided attenuation and multi-scale scattering are necessary for maximum performance, that the full model outperforms single- or two-component ablations, that the attenuation-scattering consistency loss is especially crucial for stable and realistic reconstructions under varying optical conditions, and that the scene adaptation mechanism secures superiority in cross-dataset and cross-condition scenarios (Li et al., 13 Aug 2025).

7. Position within physically grounded Gaussian Splatting research

DualPhys-GS belongs to a broader movement toward physically grounded Gaussian Splatting, but its notion of “physics” is specifically underwater optical physics. This distinguishes it from work such as PG-3DGS, which couples differentiable physics simulation with 3D Gaussian representations so that generated shapes satisfy tasks such as pouring and aerodynamic lift (Lee et al., 11 May 2026). It also differs from GS-Verse, which directly integrates an object’s mesh with a Gaussian Splatting representation for physics-aware interaction in virtual reality, including stretching, twisting, and shaking manipulations (Pechko et al., 13 Oct 2025).

Other contemporaneous work pushes 3DGS toward general physical interaction at scene scale. “Scene-Level Heterogeneous Physics Simulation with 3D Gaussian Splats” introduces a Representation Abstraction Framework that translates 3DGS assets, meshes, and fluids into a unified physical particle set for scene-level, heterogeneous, multi-solver simulation (Liu et al., 19 Jun 2026). In a different direction, GS-IR adapts 3D Gaussian Splatting to inverse rendering by estimating geometry, material, and environment illumination from multi-view images under unknown lighting conditions (Liang et al., 2023).

This broader context helps clarify a common misunderstanding. DualPhys-GS is not a mechanics-oriented simulator and is not primarily about deformation, collision, or articulated interaction. Its contribution is instead to encode underwater attenuation, scattering, and water-body variability directly into the reconstruction process. A plausible implication is that its main scientific significance lies in demonstrating that 3DGS can be specialized not only by renderer design or geometry priors, but also by domain-specific physical image formation models.