AI art app Lensa and its ‘magic avatars’ are taking over the internet – why that matters and why some are worried
- An app called Lensa has taken social media by storm, allowing users to create AI-generated photos of themselves in whimsical settings from their own selfies
- These ‘magic avatars’ are wildly popular but some claim that apps of this sort are stealing from real artists, while others say they are racist and sexist
You could look like a warrior, gearing up for battle. Or maybe you would like to be surrounded by flowers and wearing a bridal gown. Better yet, how about a goddess?
These images, called “magic avatars”, are made by an app called Lensa. Owned by Prisma Labs, Lensa allows users to upload photos of themselves, pay a fee and, within minutes, download photos of themselves in all kinds of whimsical settings.
The avatars have quickly gained popularity among social media users, and celebrities including Chance the Rapper, Michaela Jaé Rodriguez and Taraji P. Henson.
But while the AI-generated photos have been a hit with some, others say apps of this sort are stealing money from the pockets of human artists. Still others say they produce racist depictions of users who upload their photos.
Here is what to know about these magic avatars and how the company has responded to warnings about them.
What are Lensa’s ‘magic avatars’? How are they created?
This latest fad started with Stability AI, the company that created a network model called stable diffusion. The model uses internet data to generate images from text.
Lensa uses a copy of the stable diffusion model, allowing users to upload their own photos and wait as the app creates personalised images, the company says on its website. The model was trained using LAION 5B, a large data set with image-text pairs.
The process takes an average of 10 minutes and involves 120 million billion mathematical operations to analyse the photos, the website says.
After avatars are created, a user’s photos are deleted permanently from their servers, Lensa says.
“Each time a user purchases a new pack of avatars, the process repeats from scratch,” Lensa says on its website. “That’s why we ask you to upload photos every time you request a new package.”
Images like the ones generated by Lensa start with something called generative AI, or a group of models, says Yaron Inger, the co-founder and chief technology officer at Lightricks, a company that produces similar AI-generated images and designed Facetune, a hit among celebrities and socialites.
The models allow users to write text prompts and keywords and create “high fidelity” images, or images that look very similar to the originals.
The models are fairly new, but good ones have popped up over the past few months, Inger says.
How was stable diffusion created?
Stable diffusion, the open source model that Lensa uses, was launched in August.
“The guys who trained [the model] basically took hundreds of millions of photos on the web,” Inger says. “[They] crawled the web, got a lot of images and showed the model these images together with text.
“The model is like a big brain that can hallucinate new images based on this ton of new data.”
Why are artists warning against using AI-generated images?
There has been serious discussion around these AI images, primarily arguments that Lensa has stolen from real artists trying to make a living.
One cautionary Facebook post from December 4, which has been shared over 50,000 times, argues the apps use copyrighted art from creators worldwide.
On its FAQ page, Lensa says the network the app uses is able to recognise patterns and connections between images and text descriptions, not individual art. The AI is then able to apply the techniques to generate new content.
“Once the training is finished, AI doesn’t refer to the original data set,” Lensa said.
“Instead, it applies the acquired principles it has developed to the process of further creation. Hence, the outputs are not replicas of any particular artist’s artwork.”
Inger, from Lightricks, says his company has taken steps to address concerns about safety among users. Stable diffusion initially came with a “pretty basic NSFW [not safe for work] filter”, so Lightricks added its own filters to protect its users.
“We monitor basically everything that happens through our systems and ensure that people create only appropriate content,” he says.
He also does not see the apps as a threat to artists and creators. In fact, he thinks it is the opposite.
“I think that these models give a lot of power to the creators,” he says. “It allows people to celebrate what other artists created and not to copy them.
“If you think about artists and their creations, a lot of their creations are inspired by other artists, right?”
Are these ‘magic avatars’ racist and misogynistic?
Another reason social media users are speaking out against Lensa and AI apps like it is because some of the images they generate are too sexualised.
United States-based activist Brandee Barker, for example, uploaded photos of her face and received at least four magic avatars depicting her partially nude or with cleavage.
In the same thread, a social media user accused the app of perpetuating “racism and sexism”.
“I ended up looking like a white woman in most of the pictures,” posted Twitter user @feministnoire.
Lensa said the app and its model use unfiltered internet content to create the magic avatars, so naturally, the AI images will contain biases humans include in their own art. “Creators acknowledge the possibility of societal biases,” Lensa said. “So do we.”
The company also stressed that Stability AI, which made the stable diffusion model, has adjusted the model to make it harder to create NSFW images.
Lensa also claims to have taken steps to reduce biases in avatars. It is taking time, though, so the company stressed that minors should steer clear of the app.
Inger says companies like Lensa and Lightricks use an open source model that is made up of tonnes of photos.
“We didn’t train the model,” he says. “It’s still early … there’s still a lot of progression in terms of what kind of data we feed to these models.”
The goal, he says, is not to include any gender or racial bias. Lightricks is trying to work internally to make sure the AI images created in its app are not offensive or biased.
And yes, he is thrilled about how things are advancing. “The options are endless,” he says. “We’re currently only scratching the surface of what is possible with these models … I think it’s only going to get better and better as time goes by.”