I’m going to use Midjourney to add art to my Substack, but sometimes you get obsessed with stuff. I decided to generate images of female faces with the prompt:
beautiful [ethnicity/nationality] woman, portrait, photograph, cinematic, white wall
Then I picked what I thought was the most neutral face focused of the four images. I upscaled that, and remastered it. I’ve output the ones that aren’t ridiculous. For whatever reason, if an ethnic group is too obscure, the women turn out to look East Asian. No idea why. I’ve put the nationalities on top, and then various Indian ethnicities. Some of the outputs make a lot of sense. Some of the others, not so much. And many, I do not know.
Warning it will take a while to load all the images.
Razib Khan is a Bangladeshi-American geneticist and writer. He is co-founder of Brown Pundits and runs Unsupervised Learning, a Substack on population genetics, evolution, history, and politics with more than 55,000 subscribers, alongside the accompanying podcast. He has blogged at Gene Expression since the early 2000s.
His writing has appeared in The New York Times, The Guardian, National Review, Slate, India Today, Quillette, and UnHerd. He is Director of Operations at FUTO in Austin, Texas, and co-founder of GenRAIT, a life-sciences platform company. Earlier in his career he developed ancestry algorithms for Gene by Gene, the Genographic Project, and Insitome, and was among the first employees at Embark Veterinary.
Born in Dhaka and raised in upstate New York and eastern Oregon, he holds degrees in biochemistry (2000) and biology (2006) from the University of Oregon, and undertook doctoral work in genomics and genetics at UC Davis. He lives in Austin.
View all posts by Razib Khan
Midjourney is fun as a tool lol but damn I don’t think the thing is too accurate for Bangladesh/Bengali and most certainly not for Jatt or something hahaha
I guess yeah on second look the Bangladesh one does look pretty typical. The Bengali one looks too off tho, I still think that.
On another note, have you ever modelled us Bengalis on qpAdmin? What is our breakdown in terms of Onge-related, Iran/IVC Diaspora, Steppe, and Asian..if you have modelled us :0
use the search box. i have. basically we’re bihari peasants + 10-20% east asian. some of us have brahmin ancestry so that will skew, but that’s very small %
yeah I know the model thing, I was only curious about the qpadmin components. ik about the thing you said about the bihar peasants, the 85% patel/15% Burmese, 70%reddy/15% punhabi/15% burmese stuff you mentioned too, saw those
but yeah my bad lol just saw one, and it was from 2020 when I searched up qpadmin. In your opinion, do you think the Bengali percentages are the same today even with newer samples and technology?
Jat, Bengali, Maratha, Bihari …….. girls i’ve seen a lot, the average just is off, with Jaat, too off.
Biharis are much thinner, Bengali more rounder, Marathi have drier Skin.
Jumpingjacksplash
3 years ago
It’s interesting that even with “beautiful” in the prompt, it doesn’t go for eye symmetry (eyeballing it, less eye symmetry than real people have). Christ Pantokrator icons in the training data?
Ugra
3 years ago
@Razib
The trick is – Instead of using endonyms, use typical first names from that region or ethnicity. You will be gratified!
thewarlock
3 years ago
Patel and Gujarati one are decent, especially the prior. Many look off
Midjourney is fun as a tool lol but damn I don’t think the thing is too accurate for Bangladesh/Bengali and most certainly not for Jatt or something hahaha
the jat looks too asian, which means it didn’t have enough input
why is the Bangladesh one off? the right one in particular seems right
maybe as per him the BD lady looks a bit less E.Asian
and yeah this is sure gonna trigger the resident Jatts like Racelearner bhai here lol
I guess yeah on second look the Bangladesh one does look pretty typical. The Bengali one looks too off tho, I still think that.
On another note, have you ever modelled us Bengalis on qpAdmin? What is our breakdown in terms of Onge-related, Iran/IVC Diaspora, Steppe, and Asian..if you have modelled us :0
use the search box. i have. basically we’re bihari peasants + 10-20% east asian. some of us have brahmin ancestry so that will skew, but that’s very small %
yeah I know the model thing, I was only curious about the qpadmin components. ik about the thing you said about the bihar peasants, the 85% patel/15% Burmese, 70%reddy/15% punhabi/15% burmese stuff you mentioned too, saw those
but yeah my bad lol just saw one, and it was from 2020 when I searched up qpadmin. In your opinion, do you think the Bengali percentages are the same today even with newer samples and technology?
pretty much
Jat, Bengali, Maratha, Bihari …….. girls i’ve seen a lot, the average just is off, with Jaat, too off.
Biharis are much thinner, Bengali more rounder, Marathi have drier Skin.
It’s interesting that even with “beautiful” in the prompt, it doesn’t go for eye symmetry (eyeballing it, less eye symmetry than real people have). Christ Pantokrator icons in the training data?
@Razib
The trick is – Instead of using endonyms, use typical first names from that region or ethnicity. You will be gratified!
Patel and Gujarati one are decent, especially the prior. Many look off