Blonde, brunette, ginger and white all refers to Europeans, which represents less people than the Chinese tag alone, so it doesn't seem like it fetishizes Asians more than whites.

The strange part is why we don't have equally rich categories for Africans. Native Americans not having their categories makes sense since there are so few of them left today, so Africans are the only large group left out.

I thought about that while writing my comment, but while hair color correlates pretty well with ethnicity, the level of granularity given to Asians still seems somewhat unique. There are, for instance, no distinction between French and Italian, nor British and Scottish categories. Blonde and ginger do have fairly ethnic implications, but not many people would answer with their hair color if asked where their family originates. That, and the fact that these hair colors are not restricted to white people, due to multi-ethnicity being a thing, as well as commercial hair dye, made me drop them in my "half the ethnic labels" comment (though I will concede that the image in most people's heads when hearing any of those terms would be a white person).

This is also coming from a perspective where I'm assuming that the audience for this algorithm (outside of academic consideration) would be westerners where white people tend to be the majority. In that case fetishization generally would only apply to racial minorities, due to the fact that the term seems to imply a sexually gratifying quality to the "uncommon". I add that, speaking as an Asian American, that it's not necessarily all that easy for most people to tell Chinese from Japanese from Koreans, whereas it's basically tautologically given that a blonde can be differentiated from a brunette since the terms are more strictly aesthetic rather than geographic.

On your point about Africans, I would extend it to both Latina and the fact that "Indian" seems to be a catch-all for any South Asians. Middle Easterners are also not represented (though this doesn't strike me as a thing where racial equality is necessarily desirable). However, this is where I think the data set is mostly responsible.

I think they took the tags from some uploader website. CJK ethnicities can be inherently identified by types of sources(smuggled amateur films, commercially made, etc) as well as by styles, when appearing in those sites.

That's my guess too, but my argument is less that this arises from a specific place of malice or intention, and rather this phenomena comes from an uncritical approach to handling the dataset. There is, for instance, no reason not to group all the CJK data as "Asian" for the sake of similarity with the rest of the ethnicities. But in uncritically reflecting the dataset, it introduces downstream effects that could potentially be amplified in later work.

> but the fact that there are so many categories for east Asian... says volumes about the sexualizing gaze cast on Asian women

I think you're reading too much intent into 4 labels. From an American perspective, Japanese was the original--commercial, censored--"Asian" porn, "Chinese" is typically amateur uncensored, etc. I... won't elaborate further here, but there are stylistic differences that have nothing to do with race.

The tags should reflect that if it's really about the niche and not the ethnicity.

You raise some very interesting points. For me this also demonstrates how deeply rooted, historical bias and prejudice bleeds through to this technology and could become a perpetuating factor. These models will reflect societal values, for better or worse.

But also what will the next model, and successive models, that are trained on data from AI generate images, be like? It seems you could get an amplified or skewed model that is pretty far from reality. At least currently they are all trained on “real” images (granted some images are heavily edited or photoshopped). I can imagine that In the (not too distant) future most images will be generated and not “real”.

Yeah, I wanted to stress that the categorizations seem more to me to be symptomatic than a reflection of OP's decision making. I loathe the label "problematic", but it seemed worth noting a potentially unintended consequence that could introduce some harms.

That said, it might be too early to speculate on models trained on the outputs of other models. This algorithm seems to benefit greatly from the post-DALLE 2 art-gen renaissance, but still has a number of oddities and uncanny outputs (noted in the rest of this thread) that I imagine are nearly unavoidable given how familiar the human form is. But certainly, if such decisions end up amplified in later work, or through widespread usage, then the consequences are more significant than, say, simply reflecting uncritically a skewed data set.

This is one of the most interesting comments here. I notice a lot of biases in image generation AI, it's so subtle it's almost at an ideological level. The style of art it creates, the color and dress of the people it generates-- there are little things that that make you wonder "why this instead of something else?"

Keep in mind this AI is biased towards white men's interests which means it heavily fetishizes and degrades Asian women.

Even searching for other races shows predominantly Asian women in the results, its quite disturbing.