It’s far more than her eyes, she is bilaterally asymmetrical. With real people you can generally take a reflection of one side and it will look fairly close to the other. This woman has so much asymmetry it is off-putting. Her eyes are different heights and shapes, her cheek bones are different, the outer part of her nostrils are at different heights, her lip sides are shaped differently, her jawlines are different, her suprasternal notch(the divot at the base of the neck) is WILDLY different. The easiest thing to spot is her different skin tones. At first, you’ll want to chalk it up to shading, but the light source isn’t to her side but in front and to the upper right, that does not allow for such a radical change if you look at her forehead.
AI notes: make face and body images more symmetrical, but not 100%. Got it.
The only reason it hasn’t done that yet is because it’s not really AI, but large model probability with training feedback, and so far the feedback has enforced the “close enough” aspect. The next versions will cross the lines that still let us sense something isn’t quite right.
It’s far more than her eyes, she is bilaterally asymmetrical. With real people you can generally take a reflection of one side and it will look fairly close to the other. This woman has so much asymmetry it is off-putting. Her eyes are different heights and shapes, her cheek bones are different, the outer part of her nostrils are at different heights, her lip sides are shaped differently, her jawlines are different, her suprasternal notch(the divot at the base of the neck) is WILDLY different. The easiest thing to spot is her different skin tones. At first, you’ll want to chalk it up to shading, but the light source isn’t to her side but in front and to the upper right, that does not allow for such a radical change if you look at her forehead.
AI notes: make face and body images more symmetrical, but not 100%. Got it.
The only reason it hasn’t done that yet is because it’s not really AI, but large model probability with training feedback, and so far the feedback has enforced the “close enough” aspect. The next versions will cross the lines that still let us sense something isn’t quite right.