“I think I have a fairly decent self-image, but I looked at the images and I was like, ‘Why do I look so good?’” said James, a Twitch streamer who declined to give his last name to keep his social media presence separate from his day job. “I think it shaved off a lot of my rough edges.”
Lensa, a photo and video editing app from Prisma Labs, has been around since 2018, but its worldwide downloads skyrocketed after the launch of its “magic avatars” feature in late November, according to analytics firm Sensor Tower. The app saw 4 million installs in the first five days of December compared to 2 million in November, shooting to the top of charts in the Apple and Google app stores. Consumers spent $8.2 million in the app during that five-day period, Sensor Towers reports.
The app is subscription based and costs $35.99 a year, with an extra charge of $3 to $12 for packs of avatars. Upload eight to 10 photos of yourself with your face filling most of the frame and no one else in the shot, and Lensa will use the photos to train a machine learning model. Then, the model generates images based on your face in different artistic styles like “anime” or “fairy princess.”
Some people marveled at how flattering or accurate the portraits seemed. Others shared garbled images with distorted facial features or limbs coming out of their heads, an outcome Lensa warns about during the upload process.
The trend also raised concerns about the equity of AI-generated images, the effects on professional artists and the risk of sexual exploitation. Here’s everything you need to know before you download.
Lensa is owned by Sunnyvale, Calif.-based Prisma Labs, which also makes the Prisma app that uses AI to duplicate photos in various artistic styles. Both Prisma Labs CEO Andrey Usoltsev and co-founder Alexey Moiseenkov used to work at Russian tech giant Yandex, according to their LinkedIn profiles.
Like competitor Facetune, Lensa comes with a collection of photo and video editing tools that do everything from replacing your cluttered living room with an artsy backdrop to removing the bags under your eyes.
How does Lensa create AI avatars?
Lensa relies on a free-to-use machine learning model called Stable Diffusion, which was trained on billions of image-and-text combinations scraped from the internet. When you upload your photos, the app sends them to its cloud storage and spins up a machine learning model individualized just for you. Then that model spits out new images in your likeness.
Will the images look like me?
It depends. Some users with dark skin say they saw more glitches and distortions in their avatars than their light-skinned friends did, reinforcing long-standing concerns about equity in AI imaging. Asian people and people who wear hijabs also took to Twitter to share inaccuracies in their AI portraits.
Usoltsev didn’t address concerns about the app’s alleged tendency to Anglicize results and referred The Washington Post to an FAQ published on the Prisma Labs website.
Due to the lack of representation of dark-skinned people both in AI engineering and training images, the models tend to do worse analyzing and reproducing images of dark-skinned people, says Mutale Nkonde, founder of algorithmic justice organization AI for the People. In scenarios where facial recognition is being used for law enforcement, for example, that creates frightening opportunities for discrimination. The technology has already contributed to at least three wrongful arrests of Black men.
There’s potential for harm on Lensa, as well, Nkonde noted. From what she’s seen, the app’s results for women tend toward “generic hot white girl,” she said.
“That can be very damaging to the self-esteem of Black women and girls,” she said. “Black women are looking at this and being like, ‘Huh. Love the picture. Doesn’t look like me. What’s going on with that?’”
Because Lensa lets you choose your avatar’s gender — including an option for nonbinary — some trans people celebrated the opportunity to see a gender-affirming version of themselves.
Trans peeps: if you’re doing the Lensa thing, take a bunch of old photos from when you were a teenager & pre-transition, enter your actual gender in the prompt, and run them through the app.
You’ll get a bunch of images of young you as the real you: pic.twitter.com/5CBGRGkpfA
— Juni (@beloved_june) December 4, 2022
Should I be worried about privacy?
It also says it only uses the photos you provide to generate avatars and deletes each batch of photos, along with the machine learning model trained from your images, after the process is complete.
Prisma Labs isn’t using the photos or individualized models to train a facial recognition network, Usoltsev said. He declined to say whether Prisma Labs stores any data based on your photos but said the company keeps the “bare minimum.”
The real privacy concern with Lensa comes from a different angle. The giant collection of images used to train the AI, called LAION, was scraped from the internet without much discretion, AI experts say. That means it includes images of people who didn’t give their consent. One artist even found photos from her private medical records in the database. To check whether images associated with you have been used to train an AI system, visit HaveIBeenTrained.com. (This engine doesn’t save your image searches.)
There’s also the potential for exploitation and harassment. Users can upload photos of anyone, not just themselves, and the app’s female portraits are often nude or shown in sexual poses. This appears to also happen to images of children, although Lensa says the app is only for people 13 and older.
“The Stable Diffusion model was trained on unfiltered internet content. So it reflects the biases humans incorporate into the images they produce,” Lensa said in its FAQ.
Why has there been backlash from digital artists?
Some creators have eagerly adopted AI imaging. But as Lensa avatars took over social media feeds, many digital artists pleaded with people to think twice before giving money to the app. Lensa’s “styles” are based on real art from real people, artists say, and those professionals aren’t being compensated.
“Nobody really understands that a program taking everyone’s art and then generating concept art is already affecting our jobs, actually,” said Jon Lam, a story artist at video game company Riot Games.
Machine learning recreates patterns in images, not individual works of art, Lensa said in its FAQ.
But Lam said he’s had friends lose jobs after employers used their creations to train AI models — the artists themselves were no longer necessary in the eyes of the companies, he said. In many cases, LAION scraped images under copyright, he said, and Prisma Labs is profiting off artists’ life work without their consent. Some creators have even found what look like artists’ signatures inside images generated on Lensa.
“The details perceived as signatures are observed in styles that imitate paintings,” the Lensa FAQ reads. “This subset of images, more often than not, comes with sign-offs by the author of the artwork.”
If you want illustrations of yourself that support traditional artists, find someone local or search through a site like Etsy and commission a portrait, Lam suggested.
“I see a really bad future if we don’t rein this thing in right now,” he said. “I don’t want that to happen, not just for artists, everybody is affected by this.”