Skeletal-Driven Animation of Anatomical Humans via Neural Deformation Gradients

How to read this

Category

Neural physics-inspired skeletal-driven animation of layered anatomical humans (skin, muscle, bone)

Contributions

• Trains a neural network to predict per-element deformation gradients, rather than vertex displacements, for volumetric anatomical animation, which the paper shows is inherently more robust to element inversion and volume drift
• Builds a layered volumetric body model (bone wrap, muscle wrap, skin, connected as tetrahedral and prism elements) derived from InsideHumans, trained on high-quality FEM simulation data, and generalizes to new body shapes and poses without retraining
• Runs at 30 frames per second on consumer hardware while closely approximating full FEM simulation results, extending prior surface-only neural Jacobian field work into a multi-layer volumetric setting

Context

The work extends deformation-gradient neural representations, previously used for surface-only problems like garment and facial deformation (Neural Jacobian Fields), into a volumetric, multi-layer anatomical setting. It directly targets the joint-collapse and self-intersection failures of geometric skinning models such as SKEL and HIT, aiming for the coordination quality of full FEM simulation at interactive speed.

Correctness

Claims are grounded in comparisons against FEM ground truth for volume preservation, inversion counts, and generalization, plus architecture ablations, published in a peer-reviewed venue (Eurographics, Computer Graphics Forum). Correctness is bounded by the diversity of the FEM training data and the fixed InsideHumans-derived body topology used to build the layered mesh.

Clarity

A long, carefully organized Eurographics paper with a thorough related-work section covering skinning, data-driven correctives, and volumetric simulation. The volumetric terminology (wrap meshes, prisms, embedded deformation) takes some upfront effort, but the writing itself is precise.

How to read it

First pass: abstract, Figure 1's animated skeleton-muscle-skin sequence, and Figure 2's template mesh diagram. Second pass: Sections 1 and 2 to place it against SKEL, HIT, and FEM or projective-dynamics prior work, plus Section 3.1 for the layered body model definition. Third pass: Sections 3.2 and 3.3 on the deformation-gradient network and the results in Section 5 for volume-preservation and generalization numbers.