I've had some challenges with MidJourney adding "stuff" to faces, mostly women's faces, and prompting with "clear skin" doesn't always help. Still, this result for "indian girl, suit and tie, photograph, 32K" turned out pretty nicely. It's using the v4 model candidate.
@pwinn I noticed that the default person is usually a young white healthy skinny cis woman unless you explicitly specify them. 🤔
Probably because that's the subject of the majority of photos and art in the scraped data?
@madelena To be fair, I should probably add that the folks at MidJourney, at least, seem to be aware of this, and have been trying to make their model more welcoming, less stuck on the generic view of the world. But it's hard for a model to be less-biased than its inputs!
In their periodic "office hours" on Discord, they've talked about this at length, prompt me to try to think about what it would be like to try to train a model biased away from white people. I think you'd have to deliberately withhold source images so that the majority of your training data was whichever ethnicity you wish to favor, and probably try to weight tags as well.
When they were doing their pre-release feedback for model v4, this one, I noticed that the "random" source images I was being asked to rate were heavily skewed *away* from white people, but still prominently featured a certain category of women. So I think they're trying!
@pwinn That's great to know that the #Midjourney team understands and is working on it! Thanks for the detailed response.