v2.0:

Better (and longer) training, smaller file size, more accuracy. It's much like v1.0, just better.

v1.0:

So, this model seems to like a higher weight (~1.2). I think that it may be more difficult to get an accurate likeness for cosplayers, probably because they're so amazing at their craft. I think I did a decent job by limiting to more casual photos. I also limited to photos of her with blue eyes, which also seemed to help. Any suggestions/advice for future versions are always welcome.