SpatialAvatar-0: High-Quality 4D Head Avatar with Multi-Stage Reconstruction

SpatialAvatar-0 uses a FLAME-mesh-bound 3D Gaussian Splatting representation to unify feed-forward prediction and per-subject refinement for 4D head avatars from one or few portraits. Prior feed-forward models required 300K–600K refinement iterations and adaptive densification that broke Gaussian layouts, preventing end-to-end sharing between the two regimes. The shared representation eliminates that incompatibility while reducing per-subject optimization cost.