Abstract
We present a novel method to interpolate smoke and liquid simulations in order to perform data-driven fluid simulations. Our approach calculates a dense space-time deformation using grid-based signed-distance functions of the inputs. A key advantage of this implicit Eulerian representation is that it allows us to use powerful techniques from the optical flow area. We employ a five-dimensional optical flow solve. In combination with a projection algorithm, and residual iterations, we achieve a robust matching of the inputs. Once the match is computed, arbitrary in-between variants can be created very efficiently. To concatenate multiple long-range deformations, we propose a novel alignment technique. Our approach has numerous advantages, including automatic matches without user input, volumetric deformations that can be applied to details around the surface, and the inherent handling of topology changes. As a result, we can interpolate swirling smoke clouds and splashing liquid simulations. We can even match and interpolate phenomena with fundamentally different physics: a drop of liquid and a blob of heavy smoke.
Introduction
In visual effects (VFX) and engineering, fluid simulations are notoriously computationally expensive. Traditional workflows require artists or scientists to tweak parameters and restart simulations from scratch to achieve a desired result, such as moving a splash or adjusting smoke viscosity.
The core challenge in reusing fluid data lies in its "unstructured" nature over time—topologies break, droplets merge, and surfaces swirl. Simple linear blending (averaging two simulations) typically results in "ghosting" or double-images rather than a physically plausible intermediate state. This research addresses the problem of automated data reuse: taking existing simulations and "morphing" them into new, valid simulations that represent intermediate parameter states.
Methodology
The proposed method moves away from traditional mesh-based matching, adopting an implicit Eulerian representation that leverages 4D Signed-Distance Functions (SDFs).
4D Signed-Distance Functions and 5D Optical Flow
By representing 3D fluid sequences as 4D volumes (3D space + time) and converting them to SDFs, every point in the grid provides a distance to the fluid surface. This representation solves the "Aperture Problem" common in standard optical flow; because SDFs have a constant gradient, the optimizer always has a clear path to find a match, even far from the surface. The concept of optical flow is extended to five dimensions: three spatial, one temporal, and one "parameter" dimension (e.g., viscosity or initial position).
Pipeline and Alignment
The pipeline consists of a pre-computation phase and a runtime phase. To handle large deformations, a multi-scale approach is used. Large-scale features are matched first, followed by "residual iterations" and a "surface projection" step to snap fine details of the source surface onto the target. A key mathematical contribution is the Eulerian Deformation Alignment, which concatenates multiple "backwards-looking" Eulerian deformations into a single, smooth linear map, preventing fluid breakup during complex interpolations.
Results and Discussion
The method was evaluated on various datasets, including buoyant smoke plumes, liquid drops impacting basins, and complex 2D parameter spaces.
- Efficiency: Once the deformation is precomputed, generating a new interpolated frame is nearly instantaneous, taking between 0.5s and 0.9s per frame.
- Robustness: The system successfully matched simulations where features moved more than 150 grid cells apart, a distance where traditional registration algorithms often fail.
- Visual Fidelity: The approach produces clean surfaces with fewer artifacts like stretching or "dissolving" fluid compared to standard hierarchical optical flow. It can even interpolate between fundamentally different physics, such as a falling liquid drop and a rising smoke blob.
Conclusion
While highly effective, the method currently relies on geometric correspondences and does not strictly enforce mass conservation or Navier-Stokes equations. Future work involves integrating physical constraints directly into the optical flow optimization and addressing the memory intensity of 5D solves. Despite these limitations, the approach offers a powerful tool for efficient, data-driven fluid animation.
Summary updated