Yeah, there’s a bias in these image generator AIs, because model photographs are:
- easy to get in bulk and in relatively high resolution etc.,
- relatively uniform (even lighting, facial symmetry etc.), meaning they reinforce existing patterns in the training data, and
- relatively easy to generate good-looking results with them, for obvious reasons, so more likely to be optimized towards by anyone involved in the training.
WolfLink@sh.itjust.works 1 week ago
It also seems to change the lighting on faces as if they are lit by a studio ring light.