SplashSplat: Reconstructing Splashing Liquids from Real-World Multi-View Video

Forum topic · 小凯 · 2026-09-19

Summary

SplashSplat (arXiv 2609.20818) addresses the reconstruction of splashing liquids, which exist for only a fraction of a second and lack persistent trackable features. The authors introduce the first synchronized multi-view dataset of real splashing liquids: a benchmark of 20 real scenes, from coherent streams to violent splashes, captured by seven synchronized calibrated 4K cameras at 60 fps, with manually refined per-view liquid and container masks and fixed evaluation splits. SplashSplat imposes physical structure only where observations can constrain it: per-frame liquid SDFs fused from masks provide geometry, level-set advection between adjacent SDFs yields a coarse velocity field, and Lagrangian carriers advected along this flow are corrected against new observations and reseeded where coverage is missing, decoding local Gaussians for differentiable rendering. SplashSplat outperforms state-of-the-art dynamic Gaussian splatting methods on real and synthetic benchmarks with more physically plausible motion and lower training cost, and also supports temporal interpolation and style transfer without re-optimization.

Overview

Research area: Computer Vision Authors: Peiyu Liu, Dingxi Zhang, Federico Tombari, Marc Pollefeys, Christina Tsalicoglou, Daniel Barath Published: 2026-09-17 arXiv: 2609.20818

Abstract (translated summary)

A splash lives for a fraction of a second: liquid sheets tear into ligaments and droplets, appearance is view-dependent and nearly textureless, and almost no features persist long enough to be tracked. As a result, reconstruction research has focused on smoke, synthetic liquids, or gently deforming surfaces. To the authors' knowledge, no synchronized multi-view dataset of splashing liquids exists. The paper therefore introduces a benchmark of 20 real scenes, ranging from coherent streams to violent splashes, captured by seven synchronized, calibrated 4K cameras at 60 fps, with manually refined per-view liquid and container masks and fixed evaluation splits.

The proposed SplashSplat method is built on a single principle: impose physical structure only where the observations can constrain it. Per-frame liquid SDFs fused from the masks provide geometry; level-set advection between adjacent SDFs produces a coarse velocity field; and Lagrangian carriers advected along this flow — corrected against each new observation and reseeded where coverage is lost — decode local Gaussians for differentiable rendering.

Results

SplashSplat outperforms state-of-the-art dynamic Gaussian splatting methods on both the new real-world dataset and synthetic benchmarks.
Reconstructed motions are more physically plausible and training cost is lower than baselines.
The same representation supports temporal interpolation and style transfer without re-optimization.