MoTA: Cross-Modal Motion Transfer for Animating 3D Head Avatars via 2D Motion Tokens

We present MoTA , a cross-modal motion transfer framework that animates 3D head avatars using motion tokens extracted from 2D image pairs. Our method decodes these tokens into 3D motion features, which serve as key and query signals in a transformer that modulates the target triplane representation. To preserve the factorized triplane structure, we employ structure-preserving plane-wise cross-attention that processes the XY, YZ, and XZ planes independently. We further employ a teacher–student learning strategy, where a privileged model with access to 3D supervision guides a student trained only on 2D cues. Additionally, a cycle consistency loss is used to enforce reversible transformations and promote globally consistent motion representations. Experiments across four benchmarks demonstrate that our method achieves superior overall reenactment quality, motion control accuracy, and cross-dataset generalization compared with prior methods, while comprehensive ablations validate the effectiveness of its key components.