PhysSkin: Real-Time and Generalizable Physics-Based Animation via Self-Supervised Neural Skinning

How to read this

Category

physics-informed self-supervised neural skinning

Contributions

• Learns continuous, mesh-free skinning weight fields as basis functions that lift a low-dimensional motion subspace of handle transformations to full-space deformation, in the spirit of Linear Blend Skinning but learned rather than authored
• Introduces a neural skinning-fields autoencoder combining a transformer encoder with a cross-attention decoder to produce discretization-agnostic fields that generalize across diverse 3D shapes
• Develops a physics-informed self-supervised training strategy with on-the-fly skinning-field normalization and conflict-aware gradient correction, balancing energy minimization, spatial smoothness, and orthogonality constraints
• Achieves real-time, generalizable physics-based animation from static geometry alone, without per-shape supervision, and was recognized as a CVPR 2026 Highlight

Context

PhysSkin extends the decades-old Linear Blend Skinning idea by replacing its hand-authored, per-vertex weight assignment with a learned continuous field, sitting in the same broad territory as prior neural skinning and physics-based deformation work but distinguished by being mesh-free and discretization-agnostic. It is a separate project from srivastava-physkin-2026 in this same archive batch: that one is PhySkin from Meta, about garments, this is PhysSkin from zju3dv, about skinning weight fields, and the two should not be cross-referenced as related despite the near-identical name.

Correctness

The claims rest on the self-supervised training strategy converging to physically consistent fields, validated through the energy minimization, smoothness, and orthogonality loss terms described in the project page abstract; the quantitative generalization results across shape categories were not read in this pass. Because the method is self-supervised against a physics objective rather than ground-truth skinning data, its accuracy is bounded by how well that objective captures real deformation behavior, and CVPR Highlight status signals peer recognition of the idea rather than confirmation of production-grade robustness.

Clarity

The abstract is technically dense, transformer encoder, cross-attention decoder, conflict-aware gradient correction, and assumes familiarity with both classical Linear Blend Skinning and self-supervised training strategies. A rigger will grasp the learned-skinning-weights, real-time framing quickly, but the loss design needs the full method section to follow.

How to read it

First pass, read the project page abstract and the CVPR poster summary to confirm the framing: continuous, mesh-free skinning fields learned via physics-informed self-supervision. Second pass, read the method section on the autoencoder architecture and exactly what conflict-aware gradient correction resolves between the energy, smoothness, and orthogonality terms, since that is the paper's core technical claim. Third pass, check the released code and any generalization benchmarks across shape categories to judge whether this could replace hand-painted skin weights in a production rig or is still research-stage.