when training for face recognition are the training images needed after?

edit

yes, you can dump the images. the prediction is done entirely based on the trained model state.

you can save the trained model to xml/yml, (yml.gz even!), and in the next run, load it instead of re-training on the images again.