I Hand-Painted Freckles Because I Care About Verisimilitude (An Update)

Several months ago I wrote about the gross underrepresentation of ethnic groups in generative AI training data, which also heavily weights models to Anglicize non-white subjects.

It’s still a problem.

Painting of a young Central Asian woman with an intense determined expression, short messy black hair with auburn highlights, and a freckled deep russet-brown complexion with mahogany undertones wearing futuristic black ballistic plate armor standing in a red desert holding a swaddled infant with tawny hair and a wheatish complexion. A large hazy blue gas giant hangs low over the horizon in the distance. Sketchy concept art style. 

This isn’t a one-off either. Multiple attempts with the model all produced the same white woman. Other top-shelf models I tested today fared slightly better to various degrees, but not by much. So, overall some models seem to be improving somewhat when it comes to representation, at least compared to half a year ago, but they’re not there yet.

One new model however deserves special mention. On February 26, 2026 Google released Nano Banana 2 (Gemini 3.1 Flash Image), and I’ve been experimenting with it extensively. While it still pushes to colonize non-white subjects, it’s not nearly as bad as other models (cough, Flux, cough).

It’s also the best model on the market for targeted edits—especially with reference images—which can be leveraged to combat the tendency all models have to whitewash.

I took Flux’s generation into NB2 and did an edit operation with a reference image.

Change the woman’s facial features in @img2 to be more central asian like the woman in @img1 with a freckled deep russet-brown complexion with mahogany undertones. Preserve everything else in @img2 exactly. 

NB2 closed the gap other models still can’t. The last mile—the part where she stops being “close” and starts being Sarai—was countless iterative generations, masked composites, and hand-painting the freckles back into pigment in Photoshop. NB2 got further than any other model has. It still couldn’t finish her.

When I wrote six months ago, my conclusion was a hope: that someone would build a model that could see her. Now someone partly has. NB2 isn’t there—it still lightens, still drifts her features toward a default, still often renders freckles as dirt flecks when it should read pigment. But it’s the first model that got close enough to argue with instead of throw out, and the first whose edit tools let me drag the rest of the way to her by hand. That’s real progress, and I won’t pretend it isn’t.

But notice what that progress actually is. The model got me to close. Everything past close—the russet that reads as warmth instead of mud, the freckles that read as melanin instead of damage, the structure that reads as Silk Road instead of Sun Belt—came from a human who already knew her face and refused the default the machine kept offering. That’s not a prompting trick. It’s the work the tools still can’t do reliably, because it requires seeing her as a person representing tens of millions of women instead of a deviation.

Which is the same thing I said about the artist who was paid (a lot) to paint her, given reference photos and a detailed brief, and reached for a vague Mediterranean default anyway. He didn’t even bother trying to paint freckles—just slapped blood spatter on her cheek and called it a paycheck.

The bias was never really about silicon. It’s about whose face gets treated as the baseline and whose gets treated as a variation to be corrected toward it—and that’s a human problem that humans built into the machines, not the other way around. The good news is it’s fixable, on both sides. NB2 is proof the silicon half is moving. The other half moves when the person at the keyboard decides the character is worth the dozens of generations and several hours in Photoshop.

Sarai was. She always was. The question the new Dark Dominion eARC cover answers isn’t whether the tools can see her yet. It’s whether the person running them bothered to.