Through the Water: Refractive Gaussian Splatting
for Water Surface Scenes

Scenes with water surfaces present a significant challenge for Gaussian Splatting due to the simultaneous presence of refraction and reflection, as well as the difficulty of accurately estimating the geometry of transparent water surfaces.

To address this, we propose a novel framework for reconstructing scenes involving both reflection and refraction caused by water surfaces. The water surface is modeled as a trainable plane, and 2D Gaussian ray tracing is applied to account for refraction through the water. We extend 2D Gaussian Splatting by introducing a soft mask parameter and a dual set of Gaussian primitives, which handle both reflected and refracted effects.

Our method achieves state-of-the-art performance on newly constructed water surface datasets, including both synthetic and real scenes, and significantly outperforms prior approaches in water-interacting regions. Furthermore, we demonstrate the editability of our model by manipulating the index of refraction to suppress or modify refractive effects, enabling scene transformations into different liquids.

We introduce Water Scene Gaussian Splatting (WSGS), which is a novel framework that enables reconstruction in scenes with refraction caused by water surfaces. We initialize the trainable plane using results obtained from Structure-from-Motion (SfM) and define the water region on this plane by using a soft mask. Reflected and refracted rays are then computed according to Snell's law with respect to this plane, after which we perform 2D Gaussian ray tracing along these rays. The reflected and refracted colors are then blended using Fresnel equations, enabling accurate rendering of the water region. Finally, we combine the water region color with the non-water region color from 2DGS using the soft mask to create the final image. We further regularize the soft mask using a pseudo mask loss derived from depth cues of the base Gaussian and the trainable plane, which reduces imperfect separation between water and non-water regions.

We created and collected two datasets named Water Synthetic and Water Real. The Water Synthetic dataset was produced by constructing scenes in Autodesk Maya, adding water dynamics via physical simulation, and rendering the outputs with the Arnold renderer. The Water Real dataset was collected with a Galaxy S25 Ultra in full HD resolution. Each scene was recorded from at least three heights to make refractive differences clear, providing about 200–400 images per scene.

We first focused on the water regions to evaluate how well our framework and the baseline methods handle view-dependent effects, specifically reflection and refraction on water surfaces. As represented in the "water" columns of table above, even for both the synthetic and real scenes, our method consistently outperformed all baseline methods for the water regions. To assess overall scene quality, we extended the evaluation to entire images, including both water and static regions. This tests the model’s ability to reconstruct realistic scenes beyond water effects. As shown in the "entire" columns of table above, our method outperformed others in most cases, demonstrating robustness across both regions.

Qualitatively, as shown in figures above, the baseline approaches that either ignored view-dependent effects or modeled only reflections often produced blurry results with visible visual artifacts in the water regions. In contrast, our method produced sharper and higher-quality reconstructions, successfully capturing the detailed appearance of water surfaces.

Since the refracted Gaussian is explicitly trained to model refraction, adjusting IOR enables the simulation of different liquids or the removal of water by eliminating both refraction and reflection. As shown in figure above, increasing the IOR to 1.46 results in oil-like refraction, whereas setting the IOR to the same value as air and disabling reflection effectively removes the water surface.