Reproduce “3D Dynamic Fluid Assets from Single-View Videos with Generative Gaussian Splatting” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhiwei Zhao, Alan Zhao, Minchen Li, Yixin Hu
Researchers from Tencent and Carnegie Mellon University developed a framework for generating physically consistent and editable 3D dynamic fluid assets from single-view videos. The method leverages generative Gaussian Splatting combined with a differentiable physics-based re-simulation, yielding high-quality dynamic fluid assets that received higher user preference for realism and aesthetic quality compared to purely generative approaches.
Key insights
- Integrating generative 3DGS with physics-based re-simulation effectively overcomes the limitations of purely data-driven approaches, ensuring physical consistency and editability for dynamic fluid assets.
- A differentiable physics engine, incorporating a modified Navier-Stokes solver and a PeRCNN for stable pressure computation, allows for robust optimization of physical parameters to align simulated dynamics with real-world observations.
- Overcoming optical flow limitations in homogeneous fluid textures through mainstream correction and physics-constrained velocity projection is crucial for accurately inferring 3D volumetric velocity from single-view video.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.