Search

A Differentiable Material Point Method Framework for Shape Morphing

νƒœκ·Έ
Pulications
MPM
Polyscope
Houdini
OpenMP
C++
πŸ‘₯ Authors
Michael Xu*, Chang-Yong Song*, David Levin, David Hyde (*Equal Contribution)
🏒 Venue
ACM SIGGRAPH / Eurographics Symposium on Computer Animation (SCA) 2024 Poster β€” Best Poster Award
πŸ“„ Status
Under Review at IEEE TVCG (2nd round)
🧠 Keywords
MPM, Differentiable Simulation, Shape Optimization, Morphing
TVCG Teaser Video:
Rendering results:
β€’
Sphere to Bunny & Duck to Cow:
β€’
D to Dragon:
β€’
TVCG:
β€’
Abstract :
We present an advanced morphing algorithm based on MPM's natural capability to handle dynamic topology changes. By leveraging differentiable simulation and a chained iterative optimization pipeline, our method enables realistic, smooth, and high-resolution morphs between source and target shapes β€” even when significant topology variation is present.
Unlike traditional morphing methods, ours respects the physics of materials, resulting in simulations that are not only visually impressive but also mechanically plausible.
Key Features
β—¦
Β Differentiable Material Point Method (MPM): Backpropagation-friendly simulations based on the classical MPM framework.
β—¦
Β Per-particle Deformation Gradient Control: Enables fine-grained control over local shape changes during optimization.
β—¦
Β Multi-pass Optimization: Reduces artifacts and improves convergence by refining the deformation field over several passes.
β—¦
Β Log-based Mass Loss Term: Helps suppress mass ejection and numerical instabilities.
β—¦
Β Support for Complex Topology Transitions: Seamlessly morphs between drastically different shapes, including changes in connected components.