I was asked to do a example of alignments and smoothing.
I chose a clip, a full screen face, to amplify the differences.
All were done with a model trained at 192px, converted with the extended mask.
You may need to make them full screen to see the differences, mainly in the jitter of the eyes and around the edges of the face.
No adjustments at all.
Spatial and temporal smoothing of alignments.
Manual adjustments, by hand with bounding box.
Manual adjustment, then Spacial and temporal smoothing.