AI Portraits is making people white

just manyally fix it back sheesh

where can i test this program?

Site is dead.
aiportraits . com

They don't need to make a GAN to tell you that when it's trained on a data set, its entire perception of the world will be confined to that data set. If anything looks foreign to it, it'll try its best to approximate it to the data it has learned. It's performing as expected. You'd probably see a similar issue if you trained a GAN to generate and place household kitchen items based on seeing a image of empty kitchens, but then in the test data add empty laundries.

It actually highlights a problem with not being able to self-induce biases in pattern recognition by comparing correlations of multiple features. When a person learns a task, they learn what patterns are meaningful or not (regardless of how strong the coincidence may be) in order to learn the boundaries of the task. Another method may be designing GANs to learn how to produce counterfactual examples of what their data sets contain (hallucinate what could be), and have another discriminator filter out the bad results until it has the desired data it needs. However that's only useful if it can self-train.

...

Just train an AI on 15th century african art. I’m sure there’s plenty of examples out there filled with realism and subtle techniques which would suit their needs.

realism and subtlety are colonialist impositions

Damn why is this already dead? I wanted to have some fun

...

no such thing as people of color
there are races of people… european is not a race, white is not a race, asian is not a race, african is not a race, black is not a race, yellow is not a race…