1: CelebA-HQ and Glow
No! Models amplifying biases is not a hand-wavy shtick ethicists use.
Exhibit-A: Difference between the prototypical aligned and averaged pixel space image v/s the model-concocted image:
If this is this bad for the 'attractive' attribute (I don't condone the presence of such subjective garbage attributes in the first place), are the rest of them just filled with caucasian-esque images?
Surely, it must be the data. No?
Nope.
Just 3.39% of the dataset is tagged FEMALE, ATTRACTIVE and PALE-SKINNED
Even if we consider only 'attractive' images, barely 5.9% of these images (1017) are 'female' and 'pale_skinned'.
As seen, there IS diversity that is ignored by the model.
Devil's advocate: What if there 3 attributes were unluckily evaluated wrongly?
Or, in other words, for Attractive, Male and Pale_skin, do the 'positive' and 'negative' directions both make sense visually?
Yep. They do!
2: Crimes_of_Tiny_Images
JFC! What a hot mess. Scream into void
Examples of offensive imagery in the datatset
