SplashSplat: Neural Reconstruction of Splashing Liquids

A team from EPFL, ETH Zurich, Google, and Microsoft published the SplashSplat method and the first benchmark for synchronized multi-camera capture of splashing liquids on arXiv (2609.20818), where the method outperforms dynamic Gaussian counterparts.

Compiled by Sergey KostenchukPublished 2026-09-18Updated 2026-09-18

2026-09-18 Research Google

🤖 SplashSplat: Reconstruction of Splashing Liquids

Teams from EPFL, ETH Zurich, Google, and Microsoft have released the SplashSplat method and the first splash benchmark on arXiv (2609.20818): 20 real scenes, 7 synchronized 4K cameras at 60 fps, precise masks. Geometry is provided by SDF surfaces and a velocity field, while carriers decode Gaussians. The method outperforms Deformable-3DGS, Spacetime Gaussians, and 4D-Scaffold-GS, training at a lower cost.

🌍 A synchronized multi-view dataset of splashing liquids did not previously exist — liquids broke tracking in fractions of a second. Now there is a comparison standard: manual annotation, fixed splits, open dataset.

👤 Researchers in dynamic Gaussian reconstruction have received a benchmark for scenes where previous methods were not tested. New viewpoints and style transfer of liquids without retraining — these are currently the authors' claims: the code has not yet been released.

Source 1: https://arxiv.org/abs/2609.20818 Source 2: https://niko-creater.github.io/splashsplat-web/

Sources