The system first reconstructs canonical 3D Gaussians from four unposed multi-view identity images. A Transformer-based Implicit Neural Animator then separates global rigid motion from local non-rigid dynamics and predicts posed-space deformations. Hybrid supervision distills structural priors from an LBS teacher while allowing training with both fitted data and large in-the-wild video collections.
LUNA is useful for avatar animation, virtual characters, telepresence, and research on controllable human motion. Its project page demonstrates RGB-, keypoint-, and sketch-driven animation, cross-identity generalization, clothing motion, and Gaussian trajectory control for cases where parametric body models can limit expressivity or introduce fitting artifacts.
