Here’s what you need to know about Lensa, the social media AI app

Written by Zoe Sottile, CNN

If you’ve logged into any social media app this week, you’ve probably seen photos of your friends, but reimagined as fairy princesses, animated characters, or celestial beings.

It’s all thanks to Lensa, an app that uses artificial intelligence to render digital portraits based on photos submitted by users.

Lensa’s highly stylized and eye-catching portraits have taken over the internet, but they have also been the subject of concern from privacy experts, digital artists and users who have noticed that the app makes their skin paler or their bodies thinner.

Here’s everything you need to know about Lensa:

CNN’s Zoe Sottile generated this image by submitting selfies to Lensa’s “Magic Avatars” feature. Credit: Lensa

How to get your own ‘magical avatar’

Images circulating online are products of Lensa’s “Magic Avatars” feature. To try it out, you must first download the Lensa app to your phone.

A one-year subscription to the app, which also provides photo editing services, costs $35.99. But you can use the app for a one-week free trial if you want to test it out before committing.

The generation of the magic avatars requires additional costs. As long as you have a free subscription or trial, you can get 50 avatars for $3.99, 100 for $5.99, or 200 for $7.99.

Lensa recommends users submit 10-20 selfies for best results. Images should be close-ups of your face with a variety of different backgrounds, facial expressions, and angles. Lensa also states that it should only be used by people 13 and older.

Lensa is a product of Prisma, which rose to popularity in 2016 with a feature allowing users to turn their selfies into images in the style of famous artists.

The app explains in its privacy policy that it uses TrueDepth API technology, and user-provided photos, or “facial data,” are used “to train our algorithms to work better and show you better results.” .

We tested the app to see what it looks like

To test the app, I curated 20 selfies that I thought showed a variety of expressions and angles and chose the 100 avatars option. It took Lensa about 20 minutes to return my avatars, which fell into 10 categories: fantasy, fairy princess, focus, pop, stylish, animated, light, kawaii, iridescent, and cosmic.

Overall, I felt like the app did a decent job of producing artistic images based on my selfies. I couldn’t quite recognize myself in most of the portraits, but I could see where they were coming from.

He seemed to recognize and repeat certain characteristics, like my pale skin or round nose, more than others. Some of them were in a more realistic style and were close enough that I could think they were actually photos of me if I saw them from a distance. Others were much more stylized and artistic, so they seemed less specific to me.

For some women, the app produces sexualized images

One of the challenges I encountered in the app was described by other women online. Even though all of the images I uploaded were fully clothed and mostly close-ups of my face, the app returned several images with implied or actual nudity.

In one of the most disorienting images, it looked like a version of my face was on top of a naked body. In several photos, it looked like I was naked but with a strategically placed blanket, or the image just cropped out to hide anything explicit. And many of the pictures, even when I was fully clothed, featured a sultry facial expression, prominent cleavage, and skimpy clothing that didn’t match the photos I submitted.

It was surprising, but I’m not the only woman to have experienced it. Olivia Snow, a researcher at the Center for Critical Internet Investigation and professional dominatrix at UCLA, told CNN that the app returned nude images of her image even when she submitted photos of herself when she was a child, an experience she documented for WIRED.

Snow said artificial intelligence technology like that used by Lensa could be used to generate “revenge pornography”, i.e. making nude images of someone without their consent.

For Snow, the release was a sign of the “complete lack of content moderation” on the app. She also called for greater regulation of AI apps like Lensa.

Lensa did not respond to a request from CNN to comment on the app producing nude or sexualized images.

Other users have documented different forms of bias produced in their Lensa images, such as black users being “whitewashed” and shown paler than they actually are. Similarly, writer and fat rights activist Aubrey Gordon wrote on her verified Instagram that the app produced images that made her look much thinner than she actually is.

“Lensa is really working overtime to make me a slim person,” she wrote in the caption.

CNN's Zoe Sottile generated this image by submitting selfies to Lensa "Magic avatars" function.

CNN’s Zoe Sottile generated this image by submitting selfies to Lensa’s “Magic Avatars” feature. Credit: Lensa

Digital artists say app co-opts their work

Lensa’s technology is based on a deep learning model called Stable Diffusion, in accordance with its privacy policy. Stable Diffusion uses a vast network of digital art retrieved from the Internet, from a database called LAION-5B, to train its artificial intelligence. Currently, artists cannot opt ​​in or out of having their art included in the dataset and therefore used to train the algorithm.

This has raised concerns among some artists, who say Stable Diffusion relies on their works to create their own images, but they are not credited or paid for their work. Earlier this year, CNN reported on several artists who were upset when they discovered their work had been used without their consent or payment to train the neural network for Stable Diffusion.

Several of the artists expressed concern that the apps could also threaten their livelihoods. Digital artists can’t compete with the low prices and speed of execution that artificial intelligence allows for a digital portrait, they said at the time.

Lensa owner Prisma tried to allay concerns about their technology by eliminating the work of digital artists.

“While humans and AI learn art styles in semi-similar ways, there are fundamental differences: AI is able to analyze and learn quickly from large datasets, but it doesn’t doesn’t have the same level of care and appreciation for art as a human being,” wrote the company on Twitter at December 6.

And “the releases cannot be described as exact replicas of any particular work of art”.


Leave a Reply

Your email address will not be published. Required fields are marked *

GreenLeaf Tw2sl