3D Gaussian Splatting gives you high-quality real-time novel view synthesis and assumes the scene is static. Coherent motion — a person, a vehicle — can be captured and reconstructed if you set up for it. Flowing water cannot, because it is high-frequency and casual capture does not provide the dense, multi-view, high-speed acquisition that would be needed.
Make It Flow: Animating Water in 3D Gaussian Splatting Scenes, posted 9 October 2026, animates water in a scene that was captured static — without dynamic capture and without physical simulation.
How it works
Three steps, and the middle one is the contribution:
Estimate 2D optical flow from reference images or from renderings of the model itself. Ordinary flow estimation on flat images.
Lift it into a consistent 3D Eulerian motion field. This is the step that matters. Optical flow is per-camera and 2D; a splat scene is 3D and viewable from anywhere. Lifting one view’s flow into a field defined over space — and keeping it consistent — is what lets the motion hold up when you walk around it.
Apply the field to deform the Gaussian primitives via a cinemagraph-inspired algorithm producing temporally seamless looping animations.
The result retains real-time rendering performance and preserves free-viewpoint navigation. Evaluated on a newly captured dataset of in-the-wild waterfall scenes, with reported improvements over prior dynamic-Gaussian and 3D-cinemagraph work in both visual quality and computational efficiency.
Why Eulerian is the right choice here, where it was the wrong one elsewhere
This is worth drawing out, because it directly contradicts something from last week and both positions are correct.
OuroWorld explicitly rejected Eulerian methods as “limited to fluid-like motion” and built a general deformation framework to capture object motion and illumination change instead.
Make It Flow embraces Eulerian — because the subject is literally fluid.
An Eulerian field describes velocity at fixed points in space and advects whatever is there through it. For water that is not an approximation, it is the standard formulation; fluid dynamics is written in exactly those terms. The method is “limited to fluid-like motion” and water is fluid-like motion.
The pair makes a decent general point about choosing a representation. OuroWorld needed generality because it takes any scene and infers what should move. Make It Flow knows it is looking at water and can therefore use the representation that fits water exactly — and gets better efficiency for it, since a velocity field is far cheaper than a learned general deformation.
Why this matters for immersive work specifically
Splat capture has become the default way to bring a real place into a headset, and the recurring complaint is that a captured place is convincing and completely dead. Water is the single worst offender: a frozen waterfall or a glassy river reads as wrong instantly, in a way that still foliage does not, because nobody has ever seen water hold still.
So the narrow version of this result — water, specifically, in splat scenes — is disproportionately useful. A large share of the outdoor locations anyone wants to capture have moving water in them, and it is the element that most reliably breaks the illusion.
Two things to establish before relying on it. The method needs to know which Gaussians are water — the paper describes estimating and applying flow but the abstract does not say how the water region is identified, and that is the step most likely to need manual help. And “visually plausible” is the stated bar, not physically accurate; this produces convincing motion, not a simulation you could measure.