Model seems heavily biased towards black princesses

#10
by Kafke - opened

Even when entering "white princess", the result is still often many times black princesses. I haven't tried other ethnicities (such as asian), but I'm expecting the results will probably be the same.

Here's my results for "princess, classic disney style", cfg 7, 20 steps, euler a.

tmpxx3_7rzn.png

Likewise, entering "white princess, classic disney style" still results in black princesses being generated.

tmp_99h5xw6.png

03397-3972075890-((white)) princess, classic disney style.png

This is fairly odd to me since a solid 90% of disney princesses are, in fact, white. So having it fairly difficult to generate white princesses seems to mean a biased training/dataset, and as a result makes this model completely useless for people who want to create pictures similar to classic disney films. Nothing wrong with having multiple ethnicities represented, but it's a bit absurd that when I enter "white princess", a large portion of my results are black princesses.

Am I just missing a token/keyword?

I trained a model of mostly white people and mostly only tagged the ethnicities of non-white people in prompts since it seemed like white would just be the default in SD (which has been noted elsewhere is a bias in AI training data sets in general).

Unexpectedly, the model seems to produce way more non-white outputs after. I think tagging a specific feature can somehow cause it to get more associated with other words in general, which may be what happened here. e.g. Maybe most of the training data was white, then say for 'Tianna' the creator used the prompt black woman or something, hoping to teach the style how to draw a black woman in that style since it's a bit of a deviation from the rest of the dataset, and then the same thing happened.

Sign up or log in to comment