How AI avatars are created and why some users aren't too happy about them
All you need are an app, artificial intelligence and a bunch of your selfies and you can look like a warrior princess, fairy or an extra on Star Wars.
Have you noticed that many of your friends are suddenly fairy princesses or space travellers? Is your Instagram feed overrun with Renaissance-style paintings of people who were definitely born in the 1990s?
If so, you are entitled to an explanation of what exactly is going on here (and it’s not time travel).
In the past week, users have flocked to Lensa AI, an app that uses your selfies and artificial intelligence to create portraits in a variety of styles. Created by the company Prisma Labs, the app is generating images – and controversy.
WHAT IS LENSA AI?
Even if you haven’t heard of Lensa AI, you’ve possibly seen its work this week. As of Dec 14, it was the most popular iPhone app in the United States in Apple’s app store.
Lensa takes your selfies, studies them and churns out original, computer-generated portraits of you – or anyone whose photos you feed it.
DO I HAVE TO PAY FOR IT?
You do. Right now, you can get 50 avatars – 10 images in five styles – for US$3.99 during a one-week trial period. (For US$35.99, you can subscribe to Lensa AI for the year, which gets you a 51-per-cent discount on future avatars.)
“Magic Avatars consume tremendous computation power to create amazing avatars for you,” according to Lensa’s checkout page. “It’s expensive, but we made it as affordable as possible.”
Fair warning: Prices have been fluctuating as the app has gotten more and more popular and may have changed since this article was published.
HOW DOES IT WORK?
After downloading the app, you’ll upload a bunch of selfies. (Do yourself a favour and don’t include any where your hands are touching your face, unless you want to pay money to get back a mess of images with phantom phalanges hanging from your mouth.)
Select a gender – male, female or other – and walk away from your phone for around half an hour, and when you return, presto.
Your face, or something like it, has been stretched and squeezed across a suite of 50 to 200 – depending on what package you purchase – AI-generated images with themes including “cosmic”, “fairy princess” and “anime”.
OK, BUT HOW DOES IT REALLY WORK?
Lensa uses Stable Diffusion, which is when all the horses in a barn spread out to give each other a little space. Just kidding: Stable Diffusion is a “really powerful” AI-based image generator, said Subbarao Kambhampati, a professor in the School of Computing and Augmented Intelligence at Arizona State University.
Similar to DALL-E 2 and Midjourney, Stable Diffusion uses image prompts (like your selfies) and text prompts (like “fantasy”, one of Lensa AI’s categories) to generate high-quality images that sometimes get “trippy”, Prof Kambhampati said.
“It’s showing you pictures that nobody took; it’s just able to stitch them up from all the other images that it has seen.”
It is cool! It’s fun to see yourself rendered as a painting, an anime character or a little woodland elf with eyes of two very different sizes and only one hand, even if you have two IRL. (Not every image is going to be perfect.)
Some users, particularly transgender and gender non-conforming people, are even finding that using Lensa to create can offer a sense of gender euphoria.
BUT SOME PEOPLE ARE UPSET ABOUT IT
There are a couple of reasons that some users are bristling at the images Lensa AI is spitting out.
The first is that many users are reporting that the art is sexualising them. When we tested the app, several of the images we received after uploading selfies and selecting “female”, included full-body renderings, despite the fact that users are specifically instructed to upload only close-up selfies.
One image featured our avatar in a metallic bikini a la Princess Leia in Return Of The Jedi. Another included only half a face atop a scantily clad body.
Prisma Labs states on its FAQ page that “occasional sexualisation is observed across all gender categories”.
Thus, it’s fairly simple to use the app to create lewd images of anyone you want. This week, Tech Crunch was able to create topless avatars of celebrities using images of an actor’s head edited onto topless bodies.
“It turns out the AI takes those Photoshopped images as permission to go wild, and it appears it disables an NSFW filter,” Tech Crunch reported.
Now imagine that same experiment, but through the lens of someone who is angling to make revenge porn. It gets murky in a hurry.
WHAT ARE USERS SAYING ABOUT THE APP?
Yasemin Anders, 29, decided to use Lensa AI after seeing it on social media and was particularly excited to see herself recreated into an ethereal fairy.
She was, however, ultimately disappointed to see that Lensa had given her a thinner body and a slimmer face and neck.
“If it’s smart enough to turn you into either a fairy or a manga figure, you would think there was some sort of smart enough software behind it to detect fat people, too,” said Anders, who lives in Berlin and works in marketing.
“Even if I imagine myself as a fairy, as like an idealised fantasy version of myself, I would still want to look like myself,” she added.
Other users are reporting that the app is making their skin appear lighter or whiter and is markedly altering their facial features, which Prof Kambhampati said was an ongoing issue with generative AI tools, including Snapchat’s face lenses.
“If most of the images you fed the system were of white faces, then it’s not surprising that when it tries to make an image that looks supposedly ‘better’, it just makes it whiter,” he said.
HOW DO ARTISTS FEEL ABOUT LENSA?
Stable Diffusion was trained on the creations of many artists, who did not explicitly consent to the use of their work for Prisma Labs’s profit, Prof Kambhampati said.
“If van Gogh was alive today, you’d probably need to pay van Gogh some licensing fee to make your pictures be in van Gogh’s style,” he said. That is not what has happened here.
“To a lot of people, having our art stolen, they don’t view it as anything personal – like, ‘Oh, well, you know, it’s just a style; you can’t copyright a style’,” said Jonathan Lam, a storyboard artist, who works in video games and animation. “But I would argue that for us, our style is actually our identity. It’s what sets us apart from each other. It’s what makes us marketable to clients.”
Lam, who lives in Vancouver, British Columbia, is one of many artists, who have since spoken out on social media about the app.
He also noted that some of the works being generated by Lensa include what appear to be renderings of artists’ signatures. (Several of our own Lensa-generated images had scribbles, where a signature might go on an IRL painting.)
“All these tech enthusiasts are saying that these generators are creating something new, but if the artist’s signature is still there, it’s not something new,” Lam said. “It’s just generating something based on the data it was fed.”
HAS PRISM LABS RESPONDED TO THESE CONCERNS?
Prisma Labs wrote in a Twitter thread that AI “will not replace digital artists” and pushed back against the characterisation that Lensa was ripping off artists’ work.
“The AI learns to recognise the connections between the images and their descriptions, not the artworks,” the company wrote.
“As cinema didn’t kill theatre and accounting software hasn’t eradicated the profession, AI won’t replace artists but can become a great assisting tool.”
ARE THERE ANY PRIVACY CONCERNS?
“I doubt that the whole business model is, ‘Give us $10 or $15 and we’ll send you back an AI glam shot’,” said Jen King, the privacy and data policy fellow at the Institute for Human-Centered Artificial Intelligence at Stanford University.
Would King use Lensa AI? “No.”
By Madison Malone Kircher and Callie Holtermann © 2022 The New York Times
This article originally appeared in The New York Times.