Reproduce “GaussianFluent: Gaussian Simulation for Dynamic Scenes with Mixed Materials” 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

Bei Huang, Yixin Chen, Ruijie Lu, Gang Zeng, Hongbin Zha, Yuru Pei, Siyuan Huang

GaussianFluent, developed by researchers at Peking University and BIGAI, introduces a framework for real-time simulation of dynamic 3D scenes using Gaussian Splatting, enabling photorealistic brittle fracture, mixed-material interactions, and fluid dynamics. The approach achieves superior internal texture quality with a CLIP score of 35.4 and 71.43% user preference, while its physical simulations received an 88.46% user preference.

Key insights

  • Combining scale regularization during initial 3DGS training with iterative multi-axis generative inpainting effectively resolves the challenge of generating consistent and photorealistic internal textures for surface-based 3DGS.
  • A continuous return mapping scheme for CD-MPM is crucial for achieving numerical stability and physical plausibility in brittle fracture simulation within the Gaussian Splatting framework.
  • GPU parallelization of the CD-MPM algorithm allows for real-time, high-fidelity physical simulations of complex material interactions, including fracture and fluid dynamics, making interactive applications feasible.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.