>The objective is similar, but it's: (1) A different approach - they also fine tune the model itself, and they get much much better identity preservation!