So I did the thing. The AI avatar thing. I'd been ignoring it for a week while it climbed the App Store charts, mostly because I assumed it was one of those Snapchat-filter fads that shows up in my feed for four days and then vanishes, but my sister sent me her set on Sunday night and I caved by Monday morning.
The app is Lensa, and the feature everyone's using is called Magic Avatars. You upload 10 to 20 selfies (they want a mix — different angles, different backgrounds, at least one full body if you've got it), it churns for anywhere from twenty minutes to a couple hours depending on server load, and it spits back 50 stylized portraits of you as a fantasy warrior, an anime character, a "cosmic" glowing space person, a corporate headshot, whatever. The base pack is $3.99 for 50 images in 10 styles. There's a $5.99 tier for 100 and a $7.99 one for 200 if you want more. I paid for the middle one, which feels like a very small amount of money to hand over for pictures of myself that don't actually look like me, but here we are.
And they don't look like you, not really. That's the part nobody quite prepares you for. Mine came back looking like a slightly-too-symmetrical stranger who happens to share my hair color. Every jawline gets sharpened, every skin texture gets smoothed into something airbrushed, and if you're a woman uploading photos there's a very well-documented pattern where a chunk of the results drift toward armor-bikini fantasy-art territory whether you wanted that or not. My sister got three separate "warrior queen" outputs that were, let's say, generously proportioned in ways her actual photos are not. She thought it was funny. I thought it was a little telling about what dataset this thing learned "fantasy woman" from.
Under the hood this is Stable Diffusion, fine-tuned on your uploaded photos, and that's really the interesting bit if you dig one level down. Stable Diffusion itself was trained on a few billion images scraped from the open web, which means somewhere in that soup are the styles of a lot of working illustrators and concept artists who never agreed to have their brushwork learned and remixed into a stranger's cosplay selfie. There's been a real backlash from artists over exactly this the last couple weeks, and I don't think it's an overreaction. If I spent a decade developing a recognizable painting style and then watched a $3.99 app spit out knockoffs of it by the thousands, I'd be furious too. I don't have a tidy answer for how you fix that after the fact, but I don't think "well the images are technically transformed enough to be legal" makes it feel okay to the person whose style got vacuumed up.
Setting the ethics aside for a second (I know, convenient), the actual mechanics are worth understanding if you've never played with this stuff. What Lensa is doing is a form of fine-tuning — feeding your specific face into the model as new training data so it can generate outputs conditioned on you, on top of everything it already knows about "fantasy warrior" or "anime portrait" or whatever style prompt is baked into that preset. That's the same basic idea behind Dreambooth, which people have been running for free on their own machines for a few weeks now if you've got a decent GPU and don't mind fiddling with Google Colab notebooks. Lensa just wraps it in an app and charges you four bucks instead of you burning an afternoon on setup. I'm not against paying for convenience, generally, I just think it's worth knowing you're paying to skip a process you could technically do yourself.
One more small gripe while I'm at it: the app pesters you to start a weekly subscription for the regular photo-editing features, separate from the avatar packs, and the toggle for that is on by default when you're mid-checkout for the avatars. I almost got signed up for something I didn't want because I was tapping through too fast. Read your screens.
Anyway. I've got fifty pictures of a fantasy-novel-cover version of myself sitting in my camera roll now and I genuinely don't know what I'm going to do with them. Set one as a LinkedIn photo as a joke, probably, and see how long it takes someone to say something.