Performance different between Vertices based head and parameter based head

Hey, Great work, congratulations! In your demo, you showed difference between vertices based decoder and parameter based decoder. We also tested the model on different decoder head and noticed that vertices based decoder's resule is much more expressive compare to result from parameter based decoders, was wondering if it is because of the limitation of the ARKit blendshape itself and wondering if there is a way to get more expressive result for ARKit blendshapes? Thank you!