So, in conclusion: If your face is large, you’re a conservative; if it’s skinny, you’re a liberal; and facial recognition is bad—we all know that. That seems to be all you need to know.
The paper:
Our results, suggesting that stable facial features convey a substantial amount of the signal, imply that individuals have less control over their privacy. The algorithm studied here, with a prediction accuracy of r = .22, does not allow conclusively determining one’s political views, in the same way as job interviews, with a predictive accuracy of r = .20, cannot conclusively determine future job performance.
r=0.22 is a weak to moderate correlation, btw. An actual predictor will need more data than just one's face in order to have a decent chance.