Science & Tech
I Uploaded Photos of Myself to the New Lensa A.I. Portrait Generator. The Results Were Stunning, Strange… and Super Creepy
My friend surprised me with a set of 100 A.I.-generated Lensa portraits. Here's what the app delivered.
Like everyone, I’d been seeing the images on social media: A.I.-generated portraits that had turned your photographs into glamorous, fanciful, and vaguely airbrushed likenesses. I had no intention of using Lensa—but then my husband called my name excitedly across the apartment, asking me to check out the 100 artworks the app had just created based on 20 images of my face.
Neither my husband, Nathan, nor I had downloaded the Lensa photo editing app, but a friend had a trial period 50 percent discount on image packs, which normally cost $12 for 100. He had offered to run our faces through the app, and, without consulting me, Nathan eagerly sent over our photos. (H/T Vinh Trinh.)
There was no chance for me to weigh the moral implications of using the app—was the A.I. ripping off hardworking artists and potentially even putting them out of a job?—or the possibility that I’d be turning over my likeness to some kind of sinister advertising project.
Instead, I was already scrolling through what the app calls Magic Avatars, marveling at what Lensa had come with in a mere 20 minutes.
There were 10 results each in 10 different styles. For women, the app gives options in the themes of fantasy, fairy princess, focus, pop, stylish, anime, light, kawaii, iridescent, and cosmic, while men get mystical, sci-fi, stylish, cyborg, anime, rock star, superhero, adventure, astronaut, and cosmic.
While I’m not 100 percent sure which portraits matched each style, the results were, to put it mildly, a mixed bag. Some were downright remarkable likenesses—and there were even a couple that Nathan and I swore had recreated actual outfits we own but hadn’t been wearing in the photographs we provided. With others, it would have been a stretch to say we shared even the barest resemblance.
Regardless of whether or not the Magic Avatars looked like me, many were quite beautiful.
The cosmic ones, if I am identifying the themes correctly, had a bit of a space-age Art Nouveau vibe, with ornate ornamental costumes and backgrounds that vaguely recalled Alphonse Mucha. One, possibly one of the pop avatars, even looked a bit like an Alex Katz. (I am not sure if the resolution is high enough for it to work, but I’m tempted to print it on canvas!)
Others had bizarre glitches, with distorted bodies and faces—one even had a creepy second set of eyes beneath the collarbone. And even the good ones often had extra fingers or limbs, or other strange mistakes.
My favorite, which I used as the featured image for this story, features the spitting image of my face wearing a flower crown, but with three arms. I look like an alien fairy goddess conducting some sort of Midsommar ritual with a mysterious talisman, and I love it.
Originally released in 2018 by Prisma Labs, Lensa owes its newfound popularity entirely to its new Magic Avatars feature. It runs on the free, open source image generator Stable Diffusion, using specific prompts to create each style of portrait.
The app trains the neural network anew on each user’s individual images, a process that involves an immense amount of computational power, conducting roughly 120 million billion mathematical operations every time.
Lensa promises that your photos and the model generated with them are permanently deleted from its servers after avatar delivery. (The Magic Avatar files for each user clock in at 4Gb, storage space which would add up quickly given the app’s current popularity!)
There have also been some complaints, however, that Lensa is misogynistic, spitting out sexualized female portraits that are all-too-often scantily clad and well-endowed. There’s a reason for that.
The Stable Diffusion model was trained LAION 5B, a dataset of 5.85 billion image/data pairs, and the images it produces therefore reflect the societal biases present in its training material.
I will confirm that there was a fair amount of cleavage in my generated images, some bare shoulders suggesting out-of-frame nudity, and even one full-fledged topless shot with a robotic flair which I, uh, will not be showing you. (I already get the occasional creepy Instagram DM when I post a photo from a day on the beach, thank you very much!)
Prisma told Tech Crunch that it is working to create filters that will prevent the creation of nude Magic Avatars.
Other users have questioned the A.I.’s ability to accurately render non-white subjects. I was surprised to find that several of my avatars looked distinctly Asian, especially, if perhaps unsurprisingly, in the kawaii, cyborg, and anime themes.
Lensa has also addressed concerns that some avatars contain what appear to be artists’ signatures, stating in its FAQ that Stable Diffusion is merely attempting to create images that look like paintings, and that it learned through trainings that paintings often have signatures.
What appears in a Magic Avatar, therefore, isn’t the remains of a signature from a real artist’s work, but fake letters that mimic one.
In the end, I ended up running two different sets of photographs through Lensa—the ones Nathan threw together, and a carefully curated set I compiled being sure to provide a variety of different angles of my face. The former actually had more accurate results, perhaps because most of the images were more tightly cropped to my face.
In my first glance at my Lensa results, I really did feel like I was staring at a (highly stylized) reflection with some of the images.
But now, the more I look through them, my sense of self is becoming less certain. Is that actually my mouth? Do my eyes really look like that? I feel as though I’m starting to doubt my ability to recognize my own face—a reminder that A.I. art remains strange new frontier, even as it grows more omnipresent.
Follow Artnet News on Facebook:
Want to stay ahead of the art world? Subscribe to our newsletter to get the breaking news, eye-opening interviews, and incisive critical takes that drive the conversation forward.