Is Automatic Mask Generator "trainable"?

Hello!
First of all just wanted to say great job by the developers!

I am trying to fine-tune SAM2 but I noticed that all the online tutorials use the image predictor method instead of the automatic mask generator as inference on the fine-tuned model.

I am somewhat confused as to whether fine-tuning SAM2 (with masks, points etc.) can produce something that can be used with the automatic mask generator. Or is it the case that you can only use the image predictor?

Thanks in advance!