Enhancing realism and visual coherence in fluid simulation for computer graphics

Abstract

Fluid simulations are widely used in a broad range of applications within computer graphics, including the entertainment industry, visual effects, interactive games, scientific visualization, and real-time medical simulation. Despite considerable advancements in existing physical models, achieving realistic simulations that combine physical realism with high-quality visual appearance remains a challenging problem. In this thesis, we make contributions toward two aspects of fluid simulation: improving the physical realism and improving the visual quality of animated fluid simulations. Toward improving the physical realism, we propose an approach that enhances the physical realism of real-time shallow water simulations. The approach builds upon the virtual pipes model and extends it using multi-layered heightmaps. It introduces the extended pipes that can resolve flows through fully flooded passages, which is not possible with existing multi-layered techniques. Two variants are proposed: axis-aligned extended pipes, which handle passages aligned with the grid, and unified extended pipes, which handle multi branch passages and allow liquid flow through passages of arbitrary orientation and shape, including around corners. By enabling reliable flow through fully flooded passages, cavities, and multi-branch structures, this approach significantly improves the physical realism of the simulation. Regarding the improvement of the visual quality, this thesis focuses on enhancing the visual richness of fluid simulations by making contributions related to texture synthesis on the animated free surfaces of fluids. Our inputs consist of a 2D exemplar, a sequence of 3D meshes (corresponding to the animated liquid) with per-vertex velocity, and a 3D orientation field on the first frame. We propose an adaptive multiresolution synthesis approach that balances fidelity to the exemplar and consistency with the fluid motion. Our approach advects the texture and the orientation field across frames, yielding a coherent sequence of textures conforming to the per-frame geometry. Adaptivity is guided by local 2D and 3D distortion measures, which determine when to resynthesize or preserve advected content. Additionally, a popping prevention mechanism is introduced to prevent sudden, drastic color changes by enforcing gradual color transitions over time. We also introduce another texture synthesis strategy: a hybrid texture synthesis method for animated fluid surfaces that combines texton-based texture synthesis with our adaptive multiresolution exemplar-based synthesis. The main textons from the input exemplar are extracted, while the remaining background areas are used to create a background exemplar. The textons are advected given the orientation field together with a process to maintain their distribution. In parallel, the background exemplar is used to synthesize the background details with adaptive multiresolution synthesis. The two textures are then merged to produce a final animated texture that preserves both the features of the exemplar and background details throughout the animation. These texture synthesis approaches perform well on both slow-moving and turbulent liquids, and across a variety of texture exemplars, significantly enhancing the visual quality of animated fluids.

Date	26 Jun 2026
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Eric Paquette (Supervisor)