Reproduce “DiffWind: Physics-Informed Differentiable Modeling of Wind-Driven Object Dynamics” on OpenResearch

About this paper

Yuanhang Lei, Boming Zhao, Zesong Yang, Xingxuan Li, Tao Cheng, Haocheng Peng, Ru Zhang, Yang Yang and 5 more

DiffWind introduces a physics-informed differentiable framework that jointly reconstructs invisible spatio-temporal wind fields and visible object dynamics from sparse-view RGB videos. The system, developed by researchers from Zhejiang University and other institutions, integrates 3D Gaussians with differentiable MPM simulation and LBM-guided wind reconstruction, outperforming state-of-the-art dynamic scene reconstruction methods (e.g., PSNR 47.15 on synthetic data, 28.27 on real-world data) and enabling realistic forward simulation and wind retargeting.

Key insights

• Integrating physics-informed constraints via the Lattice Boltzmann Method (LBM) for wind and Material Point Method (MPM) for objects into a differentiable pipeline enables physically plausible joint reconstruction of invisible environmental forces and visible object dynamics.
• The densification of internal object points for 3D Gaussian representations, combined with MLLM-based physical property inference, significantly enhances the robustness and accuracy of deformable object simulation from visual input.
• Guiding the reconstruction of the wind force field with an LBM simulator, informed by inferred wind source directions, improves physical consistency and leads to more precise dynamic scene reconstructions than purely visual approaches.

Summary by alphaXiv.