Wouldn't your pretrained model be lacking in neurons that react to human features though? Seems like a poor choice if your ultimate goal is human images.