Physics-based fluid simulations produce breathtakingly realistic smoke, water, and fire by numerically solving the Navier-Stokes equations. However, as noted in a comprehensive survey by Xiaokun Wang et al. in Physics-based fluid simulation in computer graphics: Survey, research trends, and challenges, this physical realism comes at the cost of artistic control. If an animator wants a plume of smoke to form the shape of a character or a logo, adjusting raw physical coefficients like viscosity or temperature is an exercise in futility.
This problem—often called "parameter hunting"—stems from the chaotic nature of fluid dynamics: a tiny change in initial forces can completely alter the downstream simulation in unpredictable ways. To give artists control, researchers have developed techniques that bridge the gap between deterministic artistic intent and chaotic physics.
Guiding Fields and Kinematic Constraints
The earliest solutions sought to combine rigid geometric control with fluid physics by introducing external guiding forces. The foundational philosophy dates back to Paul M. Isaacs in Controlling dynamic simulation with kinematic constraints, who proposed combining kinematic constraints (user-defined paths) with dynamic simulations.
To make fluids follow arbitrary shapes, Lin Shi designed a framework in Controllable smoke animation with guiding objects where target 3D meshes act as "guiding objects." The simulation calculates a driving force that coaxes smoke particles toward the target geometry. Crucially, to prevent the smoke from looking like a rigid solid, the system blends these guiding forces with the fluid's natural velocity field. This allows the smoke to inhabit a shape (like a horse or a genie) while maintaining its characteristic turbulent, swirling details.
Later, Tiffany Inglis et al. refined this optimization problem in Primal-Dual Optimization for Fluids using a fast Primal-Dual optimization method. This allowed artists to guide fluid motion smoothly without introducing unphysical artifacts or compromising realism at solid boundaries.
Space Deformation: Bending the Grid
While guiding forces work, they still suffer from predictability issues: a force applied at one step might overcorrect or cause unnatural oscillations. Dinesh Manocha introduced a radically different intuition in Editing smoke animation using a deforming grid: instead of pushing the fluid through space with invisible "ghost forces," why not deform the space itself?
In a standard Eulerian fluid simulation, smoke is advected across a static, rectilinear grid. The author proposed deforming this underlying simulation grid using standard mesh-editing tools. By modifying the mathematical advection operator to account for the grid's local stretching and bending, the fluid naturally follows the deformed coordinates. Because the underlying solver still enforces incompressibility and preserves vorticity within this warped space, the smoke retains its intricate physical details while perfectly conforming to the artist's macro-level geometric edits. This eliminates the trial-and-error of force tuning entirely.
Transport-Based Neural Style Transfer
As computer graphics advanced, artistic control expanded from macro-shapes to fine-grained stylistic textures—such as making smoke look like an oil painting or a line drawing. Simply applying 2D style transfer to rendered frames causes severe temporal flickering because the texture does not move with the fluid.
To solve this, Byungsoo Kim et al. introduced Transport-Based Neural Style Transfer (TNST) in Transport-Based Neural Style Transfer for Smoke Simulations. The core insight is to optimize the velocity field of the fluid rather than modifying individual voxels or pixels directly. By forcing the velocity field to "transport" smoke density into configurations that match the structural motifs of a style image, the resulting stylization is inherently physically plausible and temporally smooth.
However, optimizing a full 3D grid is computationally expensive. To overcome this, the authors shifted the optimization to a particle-based framework in Lagrangian Neural Style Transfer for Fluids. Because particles naturally track fluid identity over time, optimizing artistic attributes directly on particles yields intrinsic temporal consistency and a 20x speedup.
Fabienne Christen et al. further extended this paradigm in Neural Smoke Stylization with Color Transfer by adding a differentiable RGB renderer and a dedicated color optimization pass. This allowed artists to map both complex structural textures and vibrant color palettes onto volumetric fluids. Similarly, for static volumes, Oliver Klehm et al. formulated volume painting as a tomographic reconstruction problem in Volume Stylizer: Tomography-based Volume Painting, enabling multi-view consistent stylization from a few reference sketches.
Real-Time Semantic Control via Machine Learning
The latest paradigm shift addresses the speed and interactivity of fluid authoring. Historically, artists had to wait hours for a simulation to bake before seeing the results of their control parameters.
To achieve real-time interactivity, Hong Huang introduced a hybrid Neural-MPM (Material Point Method) framework in Hybrid Neural-MPM for Interactive Fluid Simulations in Real-Time that utilizes a diffusion-based model. Instead of forcing an artist to guess abstract force coefficients, the user simply draws high-level sketches—such as directional arrows or target shapes. The diffusion model instantly translates these visual inputs into low-level, physically accurate dynamic force fields.
Combined with subspace fluid simulation techniques like those proposed by Eitan Grinspun et al. in Fast Subspace Fluid Simulation with a Temporally-Aware Basis, which compress the simulation space into a temporally-aware basis, modern pipelines allow artists to interactively steer and sculpt complex fluid dynamics on the fly.
Key Foundational Work
Controllable smoke animation with guiding objects (2005) — contributed 1 concept elements to this direction.
Meshless animation of fracturing solids (2005) — contributed 1 concept elements to this direction.
Controlling dynamic simulation with kinematic constraints (1987) — contributed 1 concept elements to this direction.