Fast 3D reconstruction networks now turn a few images or a short video into a 3D scene in under a second. They do this with one forward pass through a trained network, with no slow fitting step for each new scene. They work well for rigid and deforming solids. They are physically blind, though: each frame they predict a fresh cloud of primitives (small 3D blobs, each carrying color, shape and position), and nothing ties the amount of "stuff" in one frame to the amount in the next. That is harmless for a chair. It is fatal for a soap film or a surfactant-laden droplet, because two quantities live only on the moving interface and must be conserved: the thin liquid layer of the film, whose thickness is visible as interference colors, and the soap-like molecules (surfactant) whose uneven spread changes surface tension and drives flow. We propose a feed-forward reconstructor for such interfaces in which these quantities are conserved exactly by construction, through every frame, every split and every merge, at no measurable cost to inference speed.
The idea is simple to state. Computational fluid dynamics already solved a closely related problem. When interface particles carrying surfactant become badly distributed, A hybrid method for insoluble surfactant dynamics throws them away, seeds a fresh uniform set, and moves the mass across with a constrained moving-least-squares transfer that conserves it to machine precision. A feed-forward per-pixel reconstructor does the same thing every frame: it discards last frame's primitives and predicts new ones. We turn that conservative remeshing step into a differentiable network layer. On top of it, the network may only redistribute mass through antisymmetric exchanges between neighboring primitives, so it can learn diffusion and Marangoni-driven transport but cannot create or destroy mass. Marangoni flow is the motion pulled along an interface by gradients in surface tension.
Core Claim
Exact local conservation of interfacial mass can be built into a sub-second feed-forward 4D reconstructor without a global solver. Doing so yields lower long-horizon concentration error and correct mass partition at topology changes. A soft penalty does not achieve this, and neither does the cheap fix of rescaling total mass each frame.
The distinction from prior work is specific. DGS-LRM is a transformer that predicts per-pixel Gaussians plus their 3D motion from monocular video in about 0.6 s. It carries no conserved quantity, and its Gaussians have no notion of interface area. A physics-constrained neural network for multiphase flows enforces mass conservation with a non-iterative algebraic projection. It does this only for bulk volume fractions on a fixed 2D grid, it models no surfactant, and it does not reconstruct anything from images. Vid2Fluid and HyFluid are physically grounded reconstructions of fluids from video. Both fit each scene by optimization, taking 1.5–2 hours per scene in Vid2Fluid's case, and both model only bulk fluid with no interface physics. Our proposal is the first to put the remeshing machinery of interfacial CFD inside a feed-forward predictor.
Mechanism
Each primitive $i$ at frame $t$ carries a position, a normal, an area weight $A_i$, and conserved masses $M_i$. One mass is film volume, $h\,dA$ with $h$ the film thickness. The other is surfactant mass, $\Gamma\,dA$ with $\Gamma$ the surfactant concentration. The network predicts a displacement for each old primitive and a fresh set of new primitives $j$. Mass then moves in two steps:
$$ \tilde M_j = \sum_i w_{ij} M_i, \qquad w_{ij}\ge 0,\ \ \sum_j w_{ij}=1,\ \ w_{ij}=0 \text{ if } \|x_i+\Delta x_i - x_j\|>r, $$
$$ M_j^{t+1} = \tilde M_j + \sum_{k \in \mathcal N(j)} F_{jk}, \qquad F_{jk} = -F_{kj}. $$
Here $w_{ij}$ are transfer weights and $r$ is a support radius. The weights come from the constrained moving-least-squares problem of the hybrid method: low-order moments of the mass distribution are preserved, subject to the partition-of-unity constraint. That problem is solved in closed form through a small linear system, so it is differentiable. $\mathcal N(j)$ is the set of interface neighbors of $j$. $F_{jk}$ is a learned flux, predicted from features of both primitives (concentration difference, a surface-tension proxy, relative velocity). Its antisymmetry is imposed structurally as $F_{jk}=g(j,k)-g(k,j)$.
Concentration is read out as $M_j/A_j$. Surface tension then follows from a learned equation of state, and that surface tension feeds back as a tangential force on predicted motion, in the spirit of the surface-force term used in standard continuum surface-force solvers. Two properties hold exactly, up to floating point. Total mass is unchanged, because the weights sum to one and the fluxes cancel. Mass cannot jump across a gap wider than $r$ in one step, which is what makes daughter drops after pinch-off inherit the right amounts.
When a film ruptures into droplets, the network spawns droplet primitives and the same transfer moves film mass into them. This mirrors the codimension change in Thin-Film Smoothed Particle Hydrodynamics Fluid, where a particle on a surface becomes a particle in the volume.
Our approach does not resolve one tension in the field, and we say so. HyFluid found that local physics losses collapse to degenerate solutions without a global pressure projection. Our conservation guarantee is local and exact. Incompressibility of the surrounding bulk is not enforced, so we restrict the claim to interfacial mass.
Research Program
Training data must come from simulation, because no real dataset in the file provides ground-truth surfactant or film-thickness fields. We will reimplement two generators. The first is the hybrid Eulerian-Lagrangian surfactant solver, used for drops in shear, translation and breakup. The second is a thin-film particle solver following Thin-Film Smoothed Particle Hydrodynamics Fluid, used for soap films and bubbles with interference coloring. These generators produce a few thousand sequences. Each is rendered from two to four views with known cameras, film thickness is turned into interference color, and surfactant is left invisible. The model must therefore infer surfactant from its effect on motion and on thickness.
The backbone follows DGS-LRM's space-time patch tokenization, with heads added for area weight, the two masses and the flux. A forward pass is cheap, but the thin-film solver's roughly 10 seconds per frame at 160,000 particles bounds dataset size. We will therefore train at modest resolution first.
Baselines
The naive reference is a DGS-LRM-style network with extra per-Gaussian channels for thickness and surfactant concentration. It predicts these independently each frame with no transfer and is trained only on supervised field losses and rendering loss. This shows what leakage looks like when nothing prevents it.
The strongest learned competitor is the same network trained with a soft total-mass penalty. It shares our backbone, data, training steps and hyperparameter search budget (equal number of sweeps over the penalty weight), so any gap comes from the conservation structure. DGS-LRM reports 14.89 dB PSNR on DyCheck and 0.21 ATE-3D on PointOdyssey. Neither is on interfacial fluids, and we found no reported result on Lacuna for any feed-forward method on surfactant or thin-film reconstruction. All learned baselines will therefore be trained by us. Beating this baseline means lower concentration error at 100-frame horizons and lower per-component mass error after breakup, at matched rendering PSNR.
The physics reference is the classical solver itself. A hybrid method for insoluble surfactant dynamics reports surfactant mass variance around $10^{-14}$ and a 3D sheared-drop deformation index $D_{Ma}=0.1888$ versus $D_s=0.1954$ for a clean drop. A level-set method for interfacial flows with surfactant reports about 0.001% surfactant mass error in 2D Stokes flow. We run this solver twice. In the oracle setting it receives the true initial state. In the realistic setting it is initialized from a feed-forward geometry and thickness estimate, with surfactant set uniform and parameters set to dataset means. We do not expect to beat the oracle. Approaching the realistic pipeline's accuracy at a fraction of its runtime would count as success.
The optimization reference is a Vid2Fluid-style per-scene fit through a differentiable version of our simulator. It runs on a small subset because of cost; Vid2Fluid needs 1.5–2 hours per scene for bulk parameters alone. It is compared on error against wall-clock time.
The key ablations remove the antisymmetric flux, leaving transfer only. They also replace the moment-preserving moving-least-squares weights with plain normalized kernel weights. Finally, they replace the whole layer with global rescaling of predicted mass to the initial total.
Why not just…?
Why not just add a soft conservation loss to DGS-LRM? Penalties keep the average close but let errors build up over a rollout. A physics-constrained neural network for multiphase flows saw nearly 34,000 out-of-bounds grid points and drifting total mass without its hard projection. The soft-penalty baseline above tests whether this also holds for interfacial quantities. If it matches us at 100-frame horizons, the core claim fails.
Why not just rescale total mass to its initial value every frame? This is the real threat. It is one line of code and conserves the global total exactly, so on the total-drift metric it will tie us. Its weakness is locality. It moves mass instantly between disconnected pieces, so after a dumbbell pinches off through the X-junction, or a catenoid splits, the daughters can end up with the wrong amounts. It also cannot represent concentration building up at a drop's trailing edge, which is the source of the Marangoni resistance behind $D_{Ma}<D_s$. Per-component mass error after breakup and the recovered deformation index decide this. If rescaling matches us there too, we will report that the flux and transfer machinery is unnecessary.
Why not just reconstruct geometry fast, then hand it to a classical surfactant solver? Those solvers are forward-only. They need a surfactant field and material parameters that video does not give directly, and A level-set method for interfacial flows with surfactant reports breakdown when drops come into near contact at feasible resolution. The realistic-solver baseline measures exactly this pipeline, so it may well beat us on accuracy for clean single-drop cases. We expect it to lose on runtime and on scenes where initial surfactant is far from uniform.
Why not just extend Vid2Fluid's per-scene optimization with surface tension? It would be physically faithful. It already costs 1.5–2 hours per scene for two bulk parameters, and its authors report trouble with chaotic flows and fine splashes, which is precisely the rupture regime here. The optimization baseline reports the error-versus-time curve so readers can see where, if anywhere, it overtakes us.
Decisive Checks
The first check is the hybrid method's breaking-dumbbell and 3D sheared-drop cases, rendered to video. Our model should keep each daughter's surfactant mass within a few percent of the simulator's value, while the naive and soft-penalty baselines drift. It should also recover the sign and rough size of the 0.0066 gap between $D_{Ma}$ and $D_s$.
The second check is thin-film bubble scenes with rupture. Here we measure film-volume conservation through the codimension change and PSNR on interference colors.
The third check compares runtime against the realistic solver pipeline and the per-scene optimizer. A result that conserves mass exactly but leaves concentration error no better than global rescaling would falsify the central claim. So would rendering quality degrading by more than about 1 dB relative to the unconstrained network.
References
- A hybrid method for insoluble surfactant dynamics
- DGS-LRM: Real-Time Deformable 3D Gaussian Reconstruction From Monocular Videos
- A physics-constrained neural network for multiphase flows
- Vid2Fluid: 3D Dynamic Fluid Assets from Single-View Videos with Generative Gaussian Splatting
- Inferring Hybrid Neural Fluid Fields from Videos
- Thin-film smoothed particle hydrodynamics fluid
- A level-set method for interfacial flows with surfactant