Traditional automatic face morphing techniques tend to generate blurry intermediate frames when the two input faces differ significantly. We propose a new face morphing approach that deals explicitly with large pose and expression variations. We recover the 3D face geometry of the input images using a projection on a prelearned 3D face subspace. The geometry is interpolated by factoring the expression and pose and varying them smoothly across the sequence. Finally we pose the morphing problem as an iterative optimization with an objective that combines similarity of each frame to the geometry-induced warped sources, with a similarity between neighboring frames for temporal coherence. Experimental results show that our method can generate higher quality face morphing results for more extreme pose, expression and appearance changes than previous methods.