TOPOLOGY-AGNOSTIC FACIAL RIGGING

TopoRigTopology-Agnostic Facial Rigging
via Multi-Source Supervision

A topology-preserving deformation model trained with complementary mesh and image supervision.

Paper PDF Code Dataset Demo

Andrew Fleet^1,4Soroush Mehraban^2,3,4Vida Adeli^2,3,4Cole Clifford²Babak Taati^3,4

¹ Queen’s University² Pickford AI³ University of Toronto⁴ Vector Institute

Inner brow up · rightAU03

04 / EXPRESSION

Textured 3D sample 00250 with a left nose sneer

Nose sneer · leftAU51

AU-conditioned deformations on four sample identities.Input mesh connectivity is preserved.

OVERVIEW

Topology-preserving
facial deformation

Automatic facial rigging across heterogeneous mesh topologies is limited by template-dependent supervision and errors introduced during deformation transfer.

TopoRig predicts action-unit-conditioned displacements directly on the input vertices. The model combines common-topology rigs, transferred custom-topology rigs, and targeted image-based cues to generalize to unseen identities while preserving input connectivity and UVs.

3,496generated identities in the corpus

45non-gaze expression controls

3complementary supervision sources

1topology-preserving model

INTERACTIVE VISUALIZATION

Action-unit visualization

Select a sample and expression control.
Adjust its intensity to inspect the deformation.

SAMPLE 00033

Loading 3D face…Preparing the sample and its blendshapes

Drag to orbit / Scroll to zoom

Eyeball movement

Separate gaze blendshapes from the sample assets.
These controls are not labeled as facial action units.

Gaze direction and eye

Gaze intensity0.00

0.0 · Neutral1.0 · Full movement

Facial controls are shown individually. Eyeball movement has independent controls.About the controls

THE METHOD

Multi-source supervision

Joint training with common-topology targets,
custom-topology targets, and image-based cues.

01 / CONSISTENCY

Common-topology rigs

Fitted ICT FaceKit rigs provide accurate, consistent expression targets on a shared topology.

Reliable expressions · template-biased geometry

02 / DIVERSITY

Custom-topology rigs

Transferred expressions retain each generated character’s geometry and introduce varied mesh connectivity.

Diverse topology · noisier transferred targets

03 / REFINEMENT

Image-guided cues

Targeted image supervision refines eight eye-region controls where transferred 3D targets are less reliable.

Localized cues · stronger eyelid closure

Conditional deformation network

THE TOPORIG ARCHITECTURE

TopoRig network architecture: geometric and landmark features enter vertex and global mesh encoders; action-unit controls condition surface blocks that predict per-vertex displacements.

Figure 2 from the paper.

Local surface geometry, landmark-relative features, global shape context, and AU controls predict a displacement for every input vertex.

V_expressed = V_neutral + ΔV

THE RESULTS

Evaluation results

Evaluated on 350 held-out identities,
across shared and unseen mesh topologies.

PIXAL3D-100K · TRANSFER FIDELITY78.5%

lower mean absolute error

Relative to the strongest evaluated baseline on identity-specific topologies.

0.274 mm MAE

Deformation error

LOWER IS BETTER ↓

Deformation error in millimeters, reproduced from Table 1 of the paper.
Method	ICT		Pixal3D-100k
Method	MAE ↓	Q95 ↓	MAE ↓	Q95 ↓
NFS	1.476	4.358	1.416	4.260
NFR	1.928	5.480	1.273	4.145
TopoRig Ours	0.400	1.257	0.274	0.975

Errors in mm; heads normalized to 240 mm height. Table 1 in the paper.

These errors measure agreement with reference expressions. Pixal3D references are automatically transferred rigs, so lower error measures transfer fidelity and topology generalization, rather than perceptual rig quality alone.

Jaw-opening results comparing TopoRig with NFR and NFS on the ICT template and two Pixal3D identities, with close-ups and error maps. — QUALITATIVE COMPARISON
Jaw opening across the ICT template and two identity-specific topologies. TopoRig produces more consistent deformation than the evaluated baselines. Figure 4 from the paper.

PUBLICATION

Paper and supplementary material

Read the full method, evaluations, ablations,
and supplementary material in the paper.

Read the paper

RESEARCH PAPER