We propose GLVD, a fast and accurate hybrid method for 3D face reconstruction that combines per-vertex neural fields with dynamically predicted 3D landmarks to achieve state-of-the-art quality from just one or a few images.