Physics3D: Learning Physical Properties of 3D Gaussians via Video Diffusion

• Maybe put text guidance somewhere in Figure 2?

• Section 4.1 should have more references to Fang et al 2019, the main literature that did this viscoelastoplasticity splitting for MPM. The same goes for C.4. These pieces are basically exactly the same with Fang et al 2019. A contribution of this paper is the introduction of it into the ML community.

• Section 4.2 should clarify that the clampping elastoplasticity model comes from Stomakhin et al 2013. It should also clarify that this model is non-physical and invented for snow animation. A more physically grounded elastoplasticity model would be models like von Mises and Drucker Prager, with proper references of mpm literature for those models.

• I’m not exactly getting Figure 11. The bread has not been torn apart. Why do the internal particles matter?

• What’s the text input for fig 4 video diffusion model?

• line 426: “1e-4” -> “10^{-4}”

• The comparisions with PhysGaussian needs to specify the parameters used and how they are chosen. It is unfair to pick a bad parameter that makes the motion bad, and call the method or the model “unrealistic”. You can say that the manual tuning is time-consuming, and diffusion models allow you to automatically get good parameters. Please modify similar statments about physgaussian throughout the paper.

• I don’t quite understand the experienmnt in Figure 6. The ball is supposed to be elastic wihtout permanent deformations. Why do you need to turn on plasticity for this one? (A metal or a cake would make more sense for this study.)

• What’s the differentiable MPM used here? Taichi? Warp?

1. Regarding weakness 2, since PhysDreamer employs only an elastic component, does this imply that it performs significantly worse on hyper-elastic materials such as fluids?
2. Beyond video-quality metrics, I wonder if it's possible to assess the performance of physical property estimation on synthetic scenes?
3. The authors compare with DreamGaussian4D in the qualitative results, yet it does not appear in the quantitative results. Could the authors clarify this choice?

• Is there any novelty in the viscoelasticity modeling? To my understanding, it appears identical to the approach in [1]. The 1D rheological model depicted in Figure 2 and the 3D finite strain multiplicative decomposition of the deformation gradient in Figure 3 align precisely with Figure 4 in [1]. However, this framework for modeling viscoelasticity is claimed as a contribution here, which does not seem appropriate.

• The elastoplastic modeling (Line 281 - 300) is exactly copied from [2]: the choice of base elastic model (fixed-corotated elasticity), the notation of $\theta_c, \theta_s$ are identical. But there is no reference to [2] included there.

• The viscoelastic modeling (Line 306-323) is exactly copied from [1]: the choice of base elastic model (stVK elasticity), the expression of trial strain $\epsilon_N$ are identical. But there is no reference to [1] included there.

[1] Fang, Yu, et al. "Silly rubber: an implicit material point method for simulating non-equilibrated viscoelastic and elastoplastic solids." ACM Transactions on Graphics (TOG) 38.4 (2019): 1-13. [2] Stomakhin, Alexey, et al. "A material point method for snow simulation." ACM Transactions on Graphics (TOG) 32.4 (2013): 1-10.

1. Could the authors break down the quantitative results into each sample? Since different real-world test examples and synthetic samples can represent various types of material, arranging the quantitative results into a table at the instance level can demonstrate the superiority of Physics3D.
2. Could the authors add Frechet Video and Inception Distance metrics for quantitative evaluation following PhysDreamer[a]? The metrics used in the paper are low-level pixel-wise error evaluations. However, all the existing test samples do not have too many dynamics, and most of the pixels in the background are static, which leads to marginal quantitative improvement in PSNR and SSIM. Maybe the high-level semantic evaluation results can make the performance of Physics3D more convincing and distinguishable.
3. The reviewer is a little concerned about whether the material modeling will degrade the capacity of the visual modeling of 3D Gaussian. According to the carnation result in Fig.5, only Physics3D fails to model the white texture of the left curtain. So, the reviewer suggests that the authors can compare synthetic datasets, e.g., fitting the same object model with different texture mapping (like checkboard) by different methods.
4. Could the authors add the quantitative ablation of the learnable internal filling strategy in PSNR/SSIM/FID/FVD? The authors should provide sufficient results to ensure the effectiveness of their contribution.
5. How sensitive is Physics3D to the simulation sub-step compared with other martial modeling? If the elastoplastic + viscoelastic can achieve more robust dynamics, the improvement in training efficiency can also contribute to Physics3D. Perhaps the authors could test this on substep durations of 1e-3 to 1e-5 seconds.
6. Could the authors add the material visualization following PhysDreamer[a] by using heat maps? Although the quantitative results show that Physics3D can achieve the best visual quality, the reviewer thinks the author should visualize the material property of the fitting results to demonstrate the plausible material distribution learned by Physics3D.

Reference: [a] Zhang, Tianyuan, et al. "Physdreamer: Physics-based interaction with 3d objects via video generation." European Conference on Computer Vision. Springer, Cham, 2025.