I'm implementing this project and it is working fine. Now I wonder how is it possible that the training phase uses only a face crop of the image, but actual use can accept a full image with multiple people.
The model is trained to find a face within an image. Training with face crops allows the training to converge faster, as it does not go through the trial-and-error to recognize -- and then learn to ignore -- other structures in the input images. The full capacity of the model topology can go toward facial features.
When you get to scoring ("actual use", a.k.a. inference), the model has no training for or against all the other stuff in each photo. It's trained to find faces, and will do that well.
Does that explain it well enough?