That post is great! thanks for the link, i hadn't stumbled across it yet.

From what I can tell from that link though, it is exactly as I suspected, and openface 0.2 significantly outperforms openbr for the LFW test set. (whereas openfaces 0.1 was largely equivalent). Moreover, I think that there is clearly room for drastic improvement with the openface architecture, either by tweaking the layout of the NN's, or simply throwing more data at it. It is not clear to me how openbr could be improved.