I don’t know what batch settings you were using. However, the images seem to be working normally on my computer.

I want you to run a test. Create a preset the uses these settings

Face Size = 100

Top Margin = 0

Horizontal Position = 50

Width = 10

height = 10

The batch using the “Head To Top of Subject” Algorithm.

I think you were probably using the “Head Only” Algorithm.  The Chin to Top of Subject Algorithm is generally much more accurate because it runs select Subject in order to get the exact position of the top of the subject. The only time you don’t want to use Chin to Top of subject is in the case where there parts of the subject that are well above the head that would throw this off.  Say if it is sports images and the people are hold a baseball bat the is  above the head. This will throw of the Select Subject because it would now be using the top of the bat instead of the top of the head.  This is the only reason I included the “Head Only” algorithm… because a lot of people are using this for sports images too where people are holding equipment above the head.  For school images, I don’t see a reason to ever switch the algorithm to Head Only as it is far less accurate.

The face detection algorithm only finds face features. The top of the head is just an estimate. So for the image where there head is tilted slightly down, the forehead/top of the head is much taller in relationship to the face features. In that case, the “head Only” Algorithm can be off by quite a bit.  In your images, I can see that happen with the “Head Only” Algorithm on the 2 boys images.  However, when using the Chin To Top Of Subject Algorithm (which is the default), the extra step using Select Subject is able to much more accurately figure out where the top of the head is so it can scale properly.

Anyway, please run the test as mentioned. Here is a screenshot of the results you should get. Assuming you get these same results, [please try running the same preset you ran before, but run with the Chin To Top of Subject algorithm instead of the Head Only algorithm. I think you will then get much more accurate results.