Compensating for ensemble-specific effects when building facial models