Research map / 3D vision / Gaussian splatting
MPMAvatar: Learning 3D Gaussian Avatars with Accurate and Robust Physics-Based Dynamics
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds a physics-based 3D Gaussian avatar for loose garments by driving an anisotropic MPM simulator with a mesh-based collision handler, learning physical parameters from multi-view video.
Problem
Realistically animating clothed human avatars with loose garments under novel motions; prior physics-based avatars (PhysAvatar with C-IPC) are brittle to noisy collider meshes, slow, and limited in appearance fidelity.
Approach
Replace C-IPC with a feedforward MPM simulator tailored for garments: an anisotropic codimensional constitutive model (Jiang et al.) plus a mesh-collider velocity projection that transfers collider velocity and normals to grid nodes via B-spline weights (O(N_f) instead of O(N_grid^3)), combined with 3D Gaussians attached to mesh faces and quasi-shadowing, and a rest-geometry parameter alpha compen
Key results
- On ActorsHQ, MPMAvatar reaches CD 0.42e-3, F-Score 95.7, LPIPS 0.033, PSNR 32.0, SSIM 0.963, beating PhysAvatar (CD 0.55, F-Score 92.9, PSNR 30.2).
- Simulation success rate on ActorsHQ is 100% versus 37.6% for PhysAvatar, with 1.1 s per frame vs 170.0 s on a single RTX 4090.
- Cloth-body penetration depth drops to 0.047 mm from 0.294 mm for PhysAvatar, and MPMAvatar supports zero-shot interaction with unseen objects such as chairs and sand.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the anisotropic constitutive model (ablation) | MPMAvatar · ActorsHQ test | CD (×10^3): 6.24 -> 0.42 | Helped |
| Use the anisotropic constitutive model (ablation) | MPMAvatar · ActorsHQ test | F-Score: 90.3 -> 95.7 | Helped |
| Learn physical parameters rather than fixing them to defaults (ablation) | MPMAvatar · ActorsHQ test | CD (×10^3): 0.69 -> 0.42 | Helped |
| Learn physical parameters rather than fixing them to defaults (ablation) | MPMAvatar · ActorsHQ test | F-Score: 92.9 -> 95.7 | Helped |
| Use quasi-shadowing (ablation) | MPMAvatar · ActorsHQ test | PSNR: 31.8 -> 32.0 | Helped |
| Use quasi-shadowing (ablation) | MPMAvatar · ActorsHQ test | SSIM: 0.963 -> 0.963 | No clear effect |
| Optimize rest-geometry parameter α instead of fixing α at 1 (ablation) | MPMAvatar · ActorsHQ | dynamics accuracy | Helped |
| Use 800 rather than 400 time substeps (ablation) | MPMAvatar · ActorsHQ test | CD (×10^3): 0.42 -> 0.42 | No clear effect |
| Use 200 rather than 400 time substeps (ablation) | MPMAvatar · ActorsHQ test | CD (×10^3): 0.42 -> 0.42 | No clear effect |
| Initialize ρ=0.5, E=100 (ablation) | MPMAvatar · ActorsHQ test | CD (×10^3): 0.42 -> 0.42 | No clear effect |
| Initialize ρ=2.0, E=100 (ablation) | MPMAvatar · ActorsHQ test | CD (×10^3): 0.42 -> 0.42 | No clear effect |
| Initialize ρ=1.0, E=50 (ablation) | MPMAvatar · ActorsHQ test | CD (×10^3): 0.42 -> 0.42 | No clear effect |
Surprising findings
- Removing the anisotropic constitutive model raises CD from 0.42 to 6.24 (x10^3), causing tearing artifacts in some cases.
- PhysAvatar's manually relaxed collision check was required to get it to terminate, so the efficiency comparison includes that manual adjustment.
Limitations
- Finite-difference parameter optimization scales linearly with number of parameters, limiting fine-grained (per-vertex) parameterization.
- No relightable rendering support.
- Rendering degrades for occluded or unseen regions not observed in training views; body regions use linear blend skinning, with no strand-based hair simulation.
Experimental setup
Datasets: ActorsHQ 4D-DRESS
Benchmarks: ActorsHQ 4D-DRESS
Metrics: Chamfer Distance F-Score LPIPS PSNR SSIM simulation success rate simulation time per frame penetration depth Key Physical Phenomena Detection score
Baselines: PhysAvatar ARAH TAVA GS-Avatar Gaussian Garments MMLPHuman
Builds on: PhysAvatar Material Point Method anisotropic constitutive model (Jiang et al.) 3D Gaussian Splatting C-IPC
Closest papers by meaning
- SimAvatar: Simulation-Ready Avatars with Layered Hair and Clothing (CVPR 2025 · similarity 0.91)
- SplattingAvatar: Realistic Real-Time Human Avatars with Mesh-Embedded Gaussian Splatting (CVPR 2024 · similarity 0.90)
- Human Gaussian Splatting: Real-time Rendering of Animatable Avatars (CVPR 2024 · similarity 0.90)
- GaussianAvatar: Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians (CVPR 2024 · similarity 0.90)
- 3DGS-Avatar: Animatable Avatars via Deformable 3D Gaussian Splatting (CVPR 2024 · similarity 0.89)
- GaussianAvatars: Photorealistic Head Avatars with Rigged 3D Gaussians (CVPR 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.