TL;DR
DiffWind is a physics-informed differentiable framework that models wind-driven object dynamics from video, integrating fluid dynamics, object deformation, and simulation for accurate reconstruction and wind retargeting.
Contribution
The paper introduces DiffWind, a novel method combining physics-based modeling, differentiable rendering, and simulation to recover and simulate wind-driven object dynamics from video.
Findings
Outperforms prior methods in reconstruction accuracy.
Enables realistic forward simulation under new wind conditions.
Supports wind retargeting and scene editing applications.
Abstract
Modeling wind-driven object dynamics from video observations is highly challenging due to the invisibility and spatio-temporal variability of wind, as well as the complex deformations of objects. We present DiffWind, a physics-informed differentiable framework that unifies wind-object interaction modeling, video-based reconstruction, and forward simulation. Specifically, we represent wind as a grid-based physical field and objects as particle systems derived from 3D Gaussian Splatting, with their interaction modeled by the Material Point Method (MPM). To recover wind-driven object dynamics, we introduce a reconstruction framework that jointly optimizes the spatio-temporal wind force field and object motion through differentiable rendering and simulation. To ensure physical validity, we incorporate the Lattice Boltzmann Method (LBM) as a physics-informed constraint, enforcing compliance…
Peer Reviews
Decision·ICLR 2026 Poster
1. The motivation is quite clear. The paper is well written and easy to follow. 2. The creative combination of LBM + MPM + 3DGS removes limitations of prior work that either modeled only visible dynamics or only simple forces. 3. The wind retargeting demonstrates a strong capability to generalize the wind to other objects. 4. On both synthetic and real data, DiffWind outperforms state-of-the-art dynamic 3DGS baselines on novel view synthesis.
1. The method explicitly optimizes only the wind force field while keeping material parameters fixed after MLLM “physical agent” reasoning; this makes it hard to disentangle whether observed motion comes from wind magnitude or material stiffness/damping. 2. Evaluation on real data relies on image metrics (PSNR/SSIM/LPIPS) and a user study, but no direct wind-field measurements are reported. Given that there are no public datasets for wind-driven dynamics, it would strengthen claims to instrument
1. The framework allows joint optimization of wind forces and object dynamics and leverages differentiable physics simulation for accurate reconstruction. The use of LBM ensures the wind dynamics adhere to fluid mechanics laws. I think this should be effective and novel. 2. The method outperforms state-of-the-art dynamic scene modeling approaches in both reconstruction accuracy and simulation fidelity. 3. It introduces wind retargeting to enable wind dynamics to be applied to novel objects. This
1. As the authors stated, the current implementation focuses on modeling object-level dynamics without accounting for interactions between multiple objects. 2. This method requires accurate segmentation for optimal performance. This may limit its application in less controlled environments and practical scenarios. 3. This paper focuses on continuum objects. What will happen when the method is extended to simulate behaviors in other types of objects?
1. The coupling of differentiable physics (LBM + MPM) with 3DGS for realistic wind–object interaction modeling is novel. 2. Experiments on synthetic and real-world datasets demonstrate clear performance gains over state-of-the-art methods. 3. The introduction of WD-Objects and the novel “wind retargeting” task broaden research potential.
1. This work can be viewed as exploration in generative simulation. However, the experimental results in the paper are still toy examples. The author should justify the practicality of the proposed method in real-world applications. What can this method be used for in practice? 2. The novelty of the proposed method should be further emphasized. How does it compare to existing approaches in the literature? Based on PhysGaussian, the authors should discuss more about the differences and improveme
Videos
Taxonomy
TopicsLattice Boltzmann Simulation Studies · Computer Graphics and Visualization Techniques · 3D Shape Modeling and Analysis
