I am not sure if this needs a new post or not, but since we are discussing problems with the manual tool, I thought it would be relevant to bring up a topic that was a problem for me and I found at the same time as this thread.
Once you move on from the included basic masks (I like the results from Bisenet-FP), if you edit the landmarks in the manual tool you must then use the mask tool located under tools.
The description is located in this thread.
Otherwise all of your work for the conversion video will be wasted and you will have even worse problems for your training set.
Masks not based on landmarks (anything other than the default masks) are not automatically regenerated when you edit the face. This means that if you do not regenerate the missing masks with the mask tool, you will have to spend a lot of time deleting individual faces for training as the faces that you manually edited will not have the mask for training and the program will not run.
You will either have to use a default mask or manually locate the offending faces in your training set if you want to use your chosen mask (VGG, Bisenet, etc), You could also regenerate your entire training set from your alignments files, but this would also require a lot of work.
If you still want to use your chosen mask with your current training set, the training program will tell you, one at a time, the offending face that needs to be removed from your faces training set.
For me the program began crashing every 15 seconds, then every 30 seconds, up to a minute, then two minutes as I deleted about 50 faces from my training set before I removed all of the offending mask-less faces.
I do not believe this is mentioned in the Extract guide (if it is I did not understand the importance) and caused problems for my second swap.