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

Forum topic · 小凯 · 2026-09-19

Summary

SplashSplat is a new method and benchmark for reconstructing splashing liquids, one of the hardest dynamic reconstruction problems because splashes last a fraction of a second, are nearly textureless, and lack persistent trackable features. The authors introduce the first synchronized multi-view dataset of splashing liquids: 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 SplashSplat method imposes physical structure only where observations can constrain it: per-frame liquid SDFs fused from masks provide geometry, level-set transport between adjacent SDFs yields a coarse velocity field, and Lagrangian carriers advected along this flow are corrected by each new observation and re-seeded where coverage is lost, decoding local Gaussians for differentiable rendering. SplashSplat outperforms state-of-the-art dynamic Gaussian splatting methods on both real captures and synthetic benchmarks, with more physically plausible motion and lower training cost. The same representation also enables temporal interpolation and style transfer without re-optimization. arXiv: 2609.20818.

Overview

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

Abstract

A splash lives for a fraction of a second: sheets tear into ligaments and droplets, appearance is view-dependent and nearly textureless, and little persists long enough to track. Reconstruction research has consequently focused on smoke, synthetic liquids, or gently deforming surfaces. To our knowledge, no synchronized multi-view dataset of splashing liquids exists. The authors therefore introduce 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.

Method

SplashSplat 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 transport between adjacent SDFs produces a coarse velocity field.
Lagrangian carriers advected along this flow are corrected against each new observation and re-seeded where coverage is lost.
These carriers decode local Gaussians used for differentiable rendering.

Results

SplashSplat outperforms state-of-the-art dynamic Gaussian splatting methods on both the real capture benchmark and synthetic data.
Motion is more physically plausible and training cost is lower.
The same representation supports temporal interpolation and style transfer without re-optimization.

---

*Auto-collected on 2026-09-19.*