Laplacian Projection Based Global Physical Prior Smoke Reconstruction.

We present a novel framework for reconstructing fluid dynamics in real-life scenarios. Our approach leverages sparse view images and incorporates physical priors across long series of frames, resulting in reconstructed fluids with enhanced physical consistency. Unlike previous methods, we utilize a differentiable fluid simulator (DFS) and a differentiable renderer (DR) to exploit global physical priors, reducing reconstruction errors without the need for manual regularization coefficients. We introduce divergence-free Laplacian eigenfunctions (div-free LE) as velocity bases, improving computational efficiency and memory usage. By employing gradient-related strategies, we achieve better convergence and superior results. Extensive experiments demonstrate the effectiveness of our method, showcasing improved reconstruction quality and computational efficiency compared to existing approaches. We validate our approach using both synthetic and real data, highlighting its practical potential.

Physics-Based Differentiable Rendering for Efficient and Plausible Fluid Modeling from Monocular Video.

Realistic fluid models play an important role in computer graphics applications. However, efficiently reconstructing volumetric fluid flows from monocular videos remains challenging. In this work, we present a novel approach for reconstructing 3D flows from monocular inputs through a physics-based differentiable renderer coupled with joint density and velocity estimation. Our primary contributions include the proposed efficient differentiable rendering framework and improved coupled density and velocity estimation strategy. Rather than relying on automatic differentiation, we derive the differential form of the radiance transfer equation under single scattering. This allows the direct computation of the radiance gradient with respect to density, yielding higher efficiency compared to prior works. To improve temporal coherence in the reconstructed flows, subsequent fluid densities are estimated via a coupled strategy that enables smooth and realistic fluid motions suitable for applications that require high efficiency. Experiments on synthetic and real-world data demonstrated our method's capacity to reconstruct plausible volumetric flows with smooth dynamics efficiently. Comparisons to prior work on fluid motion reconstruction from monocular video revealed over 50-170x speedups across multiple resolutions.