From 'Plastic Man' to 'Flesh and Blood': The Physical Revolution in Character Animation, PhysRig Achieves More Realistic and Natural Deformation Effects}

Author: Hao Zhang, PhD student at University of Illinois Urbana-Champaign, specializing in 3D/4D reconstruction, generative modeling, and physics-driven animation. Currently a research intern at Snap, with previous internships at Stability AI and Shanghai AI Laboratory. PhysRig, developed jointly by UIUC and Stability AI, aims to advance character animation towards more realistic and controllable physics-based deformation.

Personal homepage: https://haoz19.github.io/

Do animated characters often look overly “plastic”? Even with complex skeleton systems, characters still appear as if they are wooden puppets with hinges when walking. This is because current mainstream binding techniques—Linear Blend Skinning (LBS)—are efficient but produce volume loss, distortions, and “candy-wrapping” effects, severely impacting realism.

In the latest ICCV 2025 paper “PhysRig: Differentiable Physics-Based Skinning and Rigging Framework,” researchers from UIUC and Stability AI propose a new framework: integrating “rigid skeleton + elastic soft tissue” modeling into the binding process using differentiable physics simulation, achieving more realistic and natural deformations.

• Project link: https://physrig.github.io
• Paper link: https://arxiv.org/abs/2506.20936
• Code repository: https://github.com/haoz19/PhysRig

1 | Limitations of Traditional LBS

LBS generates animations by weighted averaging of points’ positions based on bones. While widely used in games, film, and research, it is fundamentally linear and non-physical, leading to issues like volume loss, unnatural twists, and poor soft tissue simulation.

Despite attempts to optimize LBS with deep learning, its structural flaws remain hard to fix.

2 | Core Idea of PhysRig

PhysRig pioneers a new approach: embedding bones into deformable soft tissue volumes, where bones drive shape changes through physical simulation rather than direct point control.

The framework has three key components:

• Differentiable physics simulator: based on Material Point Method (MPM), modeling stress, strain, mass, and momentum conservation to simulate natural deformations under forces.
• Material prototypes: a small set (25–100) representing elastic properties of different regions, interpolated in space via Mahalanobis distance to control responses.
• Driving points system: virtual joints controlling deformation via velocity, initialized with traditional rigging tools like Pinocchio and refined through optimization.

3 | Physics Simulation & Optimization

To infer internal skeleton movements and material parameters from observed animations, PhysRig employs an iterative inverse skinning process:

1. Fix skeleton velocity, optimize material parameters;
2. Fix material parameters, optimize per-frame driving point velocities;
3. Alternate until convergence.

This considers temporal consistency of materials and local frame-by-frame skeleton motion, ensuring stable and efficient optimization.

4 | Evaluation & Dataset

Researchers built a dataset with 17 character types (120 animation sequences), including:

• Humanoids (e.g., Michelle, Kaya)
• Quadrupeds (e.g., cheetah, mammoth, stegosaurus)
• Unusual creatures (e.g., sharks, pterosaurs, cobras)

Compared methods include:

• LBS + RigNet initialization
• LBS + Pinocchio initialization
• LBS + Ground Truth initialization
• PhysRig’s initial and optimized results

Metrics such as user ratings and Chamfer distance show PhysRig significantly outperforms traditional methods, producing more realistic dynamic effects.

5 | Extended Applications: Motion Transfer

PhysRig can also perform pose transfer: extracting skeleton motion from a source animation, transferring it to a different object (e.g., different species), and generating natural deformations.

Unlike traditional weight-prediction methods, PhysRig handles large structural differences effectively, such as transferring human motion to a jelly monster.

6 | Summary & Outlook

PhysRig offers a path from traditional rigging to physically realistic binding:

• Overcomes LBS limitations;
• Enables natural deformation of complex, diverse objects;
• Compatible with deep learning, suitable for differentiable optimization and end-to-end training;
• Opens new avenues for animation, gaming, film, and robotics simulation.

The project is now showcased on the official website, with plans to open-source code and datasets around ICCV 2025. Future plans include developing a Blender plugin for artists.

🎬 If you’re interested in physics simulation and character animation, visit the project homepage or contact the authors!