Firstly just want to say this software is incredible, massive shout out to the developers, and also the community on the forums who are super helpful!
When checking the alignments/masks with the manual tool, I've noticed that the face coverage can be wildly different from face to face. For example, landmark 1 is often where the top of the ear joins the face, but sometimes much higher, and sometimes goes completely out of the face altogether. Is it good to correct these or let the model learn from the differences?
Also, some of the smaller faces are detected but the alignments are pretty funky. The mask partially covers the face but it seems detection hasn't identified the landmarks correctly; is it best to try and correct this with extraction/bounding box and manual alignments, or just leave it up to the model to decide? I'm thinking specifically for convert so I assume every frame needs an alignment file to prevent flickering/blur?