Short answer is you can't.
Longer answer is slightly more complicated than that. This assumes that the extracted face sets are from Faceswap. If they aren't, then see the short answer.
If they are of the new extract type (ie from the past week or so) then the alignments information is in the png header, so you can just drop the files into the training folder.
If the face sets are older than that, but you have an alignments file, then you can actually updated the faces to the newest format. The easiest way to do that would probably be to do a "dummy" train (i.e. run a training session for 1 iteration with the imported faceset). By doing this, Faceswap will automatically update the Faceset to put the alignments data in the PNG header, then you can just use the faces as per the last paragraphs.
Assuming the 256px images are "legacy" facesets (i.e. they are cropped to the face rather than to the head) then you're probably out of luck for them. It is not possible to mix "legacy" and "full face" facesets for training.
The absolute final thing you can do is run "Extract" against the already extracted faces to get an alignments file. This really isn't recommended though.
Your best option will always be to source more faces yourself and extract them from video/images.