Umbrias , 11 months ago

The point is:

A. Commonly available models generally aren’t, so you’d need to go through the arduous task of retraining specifically for that which is non trivial.

B. That this bias exists in the data itself and thus difficult to flip it

C. That without being aware of these biases you will, as they did even aware of them, generate images which perpetuate the problem.

Yes with enough data you could flip it, but that’s not really here nor there.