MISTA: Compact Multi-Identity Structure-Aware Tensorized Avatars

This paper addresses the challenge of learning a shared and compact representation of multi-identity human avatars from monocular videos. Most existing approaches rely on per-identity optimization, which lacks generalizability and is costly in both time and storage. We propose MISTA1, a method that applies Tensor Train (TT) decomposition to approximate the global tensor of 3D Gaussian (3DG) parameters for multiple avatars in canonical pose. Before decomposition, the parameters are reordered and reshaped in a structure-aware manner to better exploit spatial correlations. Experiments demonstrate that MISTA achieves reconstruction and animation quality comparable to the state-of-the-art method MIGS while using only 6.5% of its parameters, making it particularly suitable for deployment on resource-constrained devices.