CANRig: Variable Local Control Cross-Attention Facial Rigging
CANRig is an automated neural facial-rigging method that uses cross-attention between control handles and mesh vertices to provide continuous, user-defined local control. It also preserves prior edits
Facial rigging is a key process for creating a character’s expressions and speech, but traditional methods require many control devices and skilled manual work. CANRig is a neural-network-based facial-rigging study that automates this process, focusing on creating and editing expressions by moving control handles on a facial mesh.
Its core is cross-attention between control handles and mesh vertices. Users can directly specify the facial regions affected by a handle, allowing them to precisely manipulate narrow areas such as eyelids or the corners of the mouth, or switch to broader deformations. A defining feature is that it does not rely on manually pre-dividing fixed regions or painting weights by hand.
The paper also proposes a workflow that preserves previous shapes so the editing process can be continued step by step. This structure is designed to prevent earlier edits from changing abruptly when the control range is adjusted. The paper presents an example in which the forehead and area around the mouth are refined broadly before the eyelids and corners of the mouth are adjusted more precisely.
The researchers present applications in animation and high-resolution visual-effects pipelines, including rapid pose creation, frame editing for captured and animated clips, and expression transfer. The authors are Arad Mohammadi, Sebastian Weiss, Jakob Buhmann, Loic Ciccone, Robert Sumner, Derek Bradley, and Martin Guay of ETH Zurich and DisneyResearch|Studios.