A splash is close to the worst possible subject for 3D reconstruction, and the paper’s opening sentence says why better than a summary can: 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.
Every assumption reconstruction relies on fails at once. There’s no texture to match between cameras. What you see depends on where you stand, because water’s appearance is mostly reflection and refraction. Nothing lasts long enough to follow from frame to frame. Which is why, as the authors of SplashSplat (arXiv, Sept 17) note, the field has concentrated on smoke, synthetic liquids and gently deforming surfaces instead — and why no synchronised multi-view dataset of splashing liquids existed before this.
The dataset
They built one: 20 real scenes, from coherent streams through to violent splashes, captured with seven synchronised, calibrated 4K cameras at 60fps, with manually refined per-view liquid and container masks and fixed evaluation splits.
That’s an unglamorous, expensive, slow piece of work — seven-camera rigs are a pain to synchronise and calibrate, and hand-refining liquid masks across 20 scenes at 60fps is enormous labour. It’s also the contribution most likely to outlast the paper. A field can’t make progress on a problem it can’t measure, and this is the first common yardstick for splashing liquids.
The method, in one principle
SplashSplat’s stated design rule is worth quoting for anyone who builds systems: impose physical structure only where the observations can constrain it.
Concretely:
- Per-frame liquid SDFs fused from the masks provide geometry — a signed distance field being a way of representing a shape as “how far am I from the surface,” which handles topology changes (a sheet tearing into droplets) without breaking.
- Level-set transport between consecutive SDFs yields a coarse velocity field.
- Lagrangian carriers are advected along that flow, corrected against each new observation, and reseeded where coverage is lost — the reseeding being the acknowledgment that in a real splash, parts of the liquid simply become unobservable.
- Those carriers decode local Gaussians for differentiable rendering.
It outperforms state-of-the-art dynamic Gaussian splatting on their real captures and on a synthetic benchmark.
Why creative people should care
Water is the most expensive thing in visual effects. Simulating it convincingly costs enormous artist time and compute, and simulation is what you do because capture wasn’t an option. A pipeline that reconstructs a real splash into a renderable, reusable 3D asset is a different way of getting there: film it once, relight and re-shoot it forever.
There’s a second, more interesting use for artists rather than studios. A reconstruction of a splash is a sculptural object made of a moment — a form that existed for a fraction of a second, now available to rotate, scale, print or project. Artists have chased that with high-speed photography for a century. Having it as geometry rather than as an image is new.
The project page is linked from the paper; whether anyone can use this outside a lab depends, as always, on whether the code and dataset actually land.