Skip to content

← ArchivePaper2026

CANRIG: Cross-Attention Neural Face Rigging with Variable Local Control

Arad Mohammadi, Sebastian Weiss, Jakob Buhmann, Loic Ciccone, Robert Sumner, Derek Bradley, Martin Guay

EurographicsDisney ResearchFacialRiggingML Deformation

Cross-attention neural network for face rigging enabling variable local control over expression and pose for production face pipelines.

Abstract

CANRig is a fully automated neural facial rigging method that lets artists deform a 3D face mesh by manipulating sparse control handles with continuous, user-defined locality at runtime. It formulates deformation as a cross-attention between control handles and mesh vertices, where the attention matrix is masked by per-vertex influence weights derived from a user-specified radius and falloff, enabling smooth transitions from precise local edits to broad global changes. A shape-preserving workflow built from zero-preserving MLPs, an added base-shape attention column, and a shape-preserving loss guarantees that previous edits are maintained during iterative, non-destructive editing and yields zero-error inversion of captured performances. For high-resolution live-action meshes the method predicts local patch blendshape weights instead of per-vertex deltas, and the authors demonstrate posing, clip editing, dialogue replacement, and expression transfer across both feature animation and visual effects pipelines.

How to read this

Category
Method: neural face rigging with variable local control
Contributions
  • CANRig, a fully automated neural facial rig that deforms a face mesh via sparse control handles with continuous, user-defined locality at runtime
  • A cross-attention formulation between handles and vertices, masked by per-vertex influence weights from a user-specified radius and falloff, spanning local to global edits
  • A shape-preserving workflow (zero-preserving MLPs, base-shape attention column, shape-preserving loss) giving non-destructive iterative editing and zero-error inversion of captured performances, with local patch-blendshape prediction for high-res live-action meshes
Context
Continues neural rig-approximation work such as Accurate Face Rig Approximation with Deep Differential Subspace Reconstruction (Song et al. 2020), adding artist-controllable locality via attention.Builds on: Accurate Face Rig Approximation with Deep Differential Subspace Reconstruction
Correctness
Capabilities (posing, clip editing, dialogue replacement, expression transfer) are demonstrated across feature-animation and VFX pipelines; the zero-error inversion and shape preservation are architectural guarantees, so the practical question is generalization and handle-placement sensitivity rather than reconstruction error.
Clarity
Dense but well-structured; a first pass conveys the handle-and-locality idea, a second pass needed for the masked cross-attention and shape-preserving components.
How to read it
Read first for how locality is expressed via the masked attention weights; second pass on the zero-preserving MLPs and shape-preserving loss if you need the non-destructive-editing guarantees.

Builds on

Built upon by

Nothing yet.

Related work

Keywords

This page summarises the entry and links to its original source. The archive never hosts or redistributes the publication itself.Show it in the full archive list →