In case you missed it, in PART 1 of my AI Bias Bypass (TM) series, I addressed 4 types of bias that I have uncovered in Midjourney.
If you haven't figured it out by now, I'm a brown girl living in an AI world (insert Barbie commercial tune). One of the most exciting things about generative AI is that I can create more visual assets that look like ME and people that I know.
Or at least that was the plan when I started out. But generating images of people of color proved to be a lot harder than I had expected. Imagine my frustration when my images seemed to come out with brown people being the same two shades, beige or mocha. I kept getting complaints from users in my community about difficulty generating "light skin" people. Or that "LatinX people are all fair-skinned."
The truth is that most ethnicities in the world, have a plethora of skin tones and shades. But it seems to have been lost on the Midjourney algorithm. There are many reasons that skin tone gets lost in translation.
Data Sets - Giving credit to MJ for including a number of diverse artists in its training data. However, it seems that the training engineers for the language model may not know enough about skin tone dialects.
User Feedback Loops - MJ feeds user feedback directly into the model to make user-driven adjustments. About 80% of MJ users are male. However, women of color tend to experience more bias regarding their complexion. So these women are more sensitive to visuals that lack diverse skin tones.
Default Skin Tone - When I fail to specify ethnicity or complexion, MJ defaults to Caucasian skin tone. Although the team at MJ says that they have improved the default diversity setting, I have not experienced that.
Now that we've done a deep dive into the issue let's look at the solution.
Skin Tone Bias Bypassâ„¢
Since I've been consistently successful at diversifying skin tones, I'll share what I have learned. The MJ Bot is very smart, but it doesn't quite understand terms like light, dark or fair regarding skin tone. So be more specific and use the colors of items the bot recognizes. Here are a few examples of phrases I use in my prompts:
Black woman with chestnut skin tone
Asian man with caramel skin tone
Hispanic woman with olive skin tone
The key is to use ethnicity AND skin tone in the prompt. This is important because, as I stated earlier, most ethnicities have a plethora of skin tones.
Here are a few skin tones that have worked for me.
North European/ European - alabaster, porcelain, ivory
Central/ Southern European - sand, beige, tan
Mediterranean/ Asian/ Latino - caramel, honey, golden, olive
East Indian/ Native American - almond, pecan, bronze, mahogany
African/ Aboriginal - chestnut, walnut, sable
I'm going to wrap up here on this one. But I would love for you to test some of my prompt suggestions with your Midjourney images and let me know how it turns out. Catch you on the flip side
Comments