The whole idea is that your input(test) data should be as close to validation data with which you trained your model. Please check validation data for FaceNet pre-trained model.

You need to have face detector, which detects faces. Face detector will mostly give you cropped image of the face and and cropped image can then be feed to facenet. Facenet use MTCNN for face detection, you can use the same or whichever you wish to, face alignment is not required because according to David Sandberg (FaceNet implementor) the network will learn alignment if unaligned faces are given. Normalization is not really required, at least i got good results without it.