Research map / 3D vision / Gaussian splatting
Gaussian Splashing: Unified Particles for Versatile Motion Synthesis and Rendering
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.
Gaussian Splashing integrates position-based dynamics (solids and fluids) with 3D Gaussian Splatting and PBR-style rendering so reconstructed scenes can be physically animated and re-rendered from novel views.
Problem
Combining physics-based simulation of solids and fluids with 3DGS-reconstructed scenes is hard: large rotational deformation produces spiky artifacts, fluid particles shift substantially, and vanilla 3DGS cannot render translucent, specular fluids.
Approach
Uses a unified particle representation under PBD: an anisotropy loss and per-kernel normals keep solid Gaussians stable under deformation, simulation particles are decoupled from rendering kernels and interpolated via GMLS, and fluids are rendered with specular PBR shading plus thickness-dependent Beer's-law refraction and position-based surface tension.
Key results
- Scenes with up to ~334K solid and ~280K fluid Gaussian kernels are simulated and rendered; e.g., the Waves scene runs at 8.1 s overall simulation time per step on an RTX 3090 (Table 1).
- Anisotropy regularization removes fuzzy and spiky artifacts under large deformation, shown qualitatively in ablations (Fig. 4).
- Specular PBR shading makes fluids look realistic rather than smoke-like, shown qualitatively against diffuse-only rendering (Fig. 5).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add anisotropy regularization during 3DGS training (ablation) | GSP · Deformed solid rendering | render quality | Helped |
| Add specular reflections to fluid kernels (ablation) | GSP · Dynamic fluid rendering | render realism | Helped |
| Add nearly-soft shadow mapping (ablation) | GSP · Dynamic scene rendering | visual realism | Helped |
| Add LaMa inpainting for displaced-object regions (ablation) | GSP · Cup & dog scene | render quality | Helped |
Limitations
- Reduced accuracy of PBD-based simulation compared to more complex constitutive models.
- Fluid rendering is far from perfect; ellipsoid splatting does not model refraction fully.
- Evaluation is largely qualitative, with no quantitative metrics reported in the main text.
Experimental setup
Baselines: PhysGaussian
Builds on: 3D Gaussian Splatting GaussianShader Position-Based Fluids PhysGaussian NeuS
Closest papers by meaning
- Fully Explicit Dynamic Gaussian Splatting (NeurIPS 2024 · similarity 0.91)
- OmniPhysGS: 3D Constitutive Gaussians for General Physics-Based Dynamics Generation (ICLR 2025 · similarity 0.89)
- SuGaR: Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering (CVPR 2024 · similarity 0.89)
- Reflective Gaussian Splatting (ICLR 2025 · similarity 0.89)
- Subsurface Scattering for Gaussian Splatting (NeurIPS 2024 · similarity 0.89)
- HybridGS: Decoupling Transients and Statics with 2D and 3D Gaussian Splatting (CVPR 2025 · 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.