**Physics3D: Learning Physical Properties of 3D Gaussians via Video Diffusion** This paper introduces Physics3D, a novel method for learning the physical properties of 3D objects using a video diffusion model. The goal is to simulate realistic dynamics by predicting and incorporating the physical properties of materials into the behavior prediction process. Current 3D generative models often focus on surface features like color and shape, neglecting the inherent physical properties that govern real-world object behavior. **Key Contributions:** 1. **Physics3D:** A generalizable physical simulation system that learns various physical properties of 3D objects. 2. **Viscoelastic Material Point Method (MPM):** Designs a highly generalizable physical simulation system based on a viscoelastic material model to simulate a wide range of materials with high fidelity. 3. **Score Distillation Sampling (SDS):** Leverages a video diffusion model to distill physical priors, optimizing physical parameters through iterative optimization. **Methodology:** - **Continuum Mechanics:** Models the deformation of materials using a deformation gradient and stress tensor. - **Material Point Method (MPM):** Discretizes materials into deformable particles, updating their properties over time to simulate elasticity and viscosity. - **Score Distillation Sampling (SDS):** Optimizes physical parameters by distilling physical priors from a video diffusion model, improving the realism of simulated dynamics. **Experiments:** - **Qualitative Results:** Demonstrates the ability to simulate realistic and physically plausible movements of complex textured objects. - **Quantitative Results:** Shows superior performance in video quality metrics compared to baselines. - **Ablation Study:** Highlights the importance of both elastoplastic and viscoelastic components in capturing realistic dynamics. **Conclusion:** Physics3D effectively learns and simulates the physical properties of 3D objects, bridging the gap between the physical world and virtual environments. The method is promising for applications requiring realistic and physically accurate simulations.**Physics3D: Learning Physical Properties of 3D Gaussians via Video Diffusion** This paper introduces Physics3D, a novel method for learning the physical properties of 3D objects using a video diffusion model. The goal is to simulate realistic dynamics by predicting and incorporating the physical properties of materials into the behavior prediction process. Current 3D generative models often focus on surface features like color and shape, neglecting the inherent physical properties that govern real-world object behavior. **Key Contributions:** 1. **Physics3D:** A generalizable physical simulation system that learns various physical properties of 3D objects. 2. **Viscoelastic Material Point Method (MPM):** Designs a highly generalizable physical simulation system based on a viscoelastic material model to simulate a wide range of materials with high fidelity. 3. **Score Distillation Sampling (SDS):** Leverages a video diffusion model to distill physical priors, optimizing physical parameters through iterative optimization. **Methodology:** - **Continuum Mechanics:** Models the deformation of materials using a deformation gradient and stress tensor. - **Material Point Method (MPM):** Discretizes materials into deformable particles, updating their properties over time to simulate elasticity and viscosity. - **Score Distillation Sampling (SDS):** Optimizes physical parameters by distilling physical priors from a video diffusion model, improving the realism of simulated dynamics. **Experiments:** - **Qualitative Results:** Demonstrates the ability to simulate realistic and physically plausible movements of complex textured objects. - **Quantitative Results:** Shows superior performance in video quality metrics compared to baselines. - **Ablation Study:** Highlights the importance of both elastoplastic and viscoelastic components in capturing realistic dynamics. **Conclusion:** Physics3D effectively learns and simulates the physical properties of 3D objects, bridging the gap between the physical world and virtual environments. The method is promising for applications requiring realistic and physically accurate simulations.