It seems like almost everybody jumped on the Lensa AI “Magic Avatar” train over the weekend. Magic Avatar is an artificial intelligence image generation tool that is a new feature on the original Lensa AI photo-editing mobile app. Pay a fee ($3.99 for a trial, $7.99 for one-time use), provide between 10 and 20 selfies, and sit back while the feature generates dozens of digital portraits of you across different illustration styles such as manga, sci-fi, fantasy, and the like.
The images look like they’ve been commissioned by different digital artists, prompting immediate concerns that the convenience and affordability of AI generation will prevent existing artists from making money off of their skillsets. Many have criticized the dataset Lensa AI employs to inform and train its AI neural network to generate such stylized imagery, as well as the exploitative terms and conditions section users must accept in order to access the feature.
As funny as the Popeye’s employee meme might be, it certainly hits the nail on the head that freelance artists are both threatened and disappointed with the explosion of “Magic Avatars.” Several prominent members of the arts and media industry have taken to Instagram and Twitter to voice their frustrations and warn the web about the insidious nature of Lensa AI — and, by extension, the open-source AI model on which it runs, Stable Diffusion, which generates images from text prompts.
One of the main complaints from artists is that the “Magic Avatar” feature utilizes Stable Diffusion’s open-source model for generative images, which is trained on an enormous Large-Scale Artificial Intelligence Open Network (LAION) dataset, LAION-5B, that contains URLs to billions of images across the web alongside descriptive tags. LAION’s datasets have come under fire before for hosting images that reveal private medical information and for scraping copyrighted images from millions of artists who post their work on host sites such as DeviantArt, Pinterest, and ArtStation. This means that the Lensa AI neural network is being trained on copyrighted artwork to produce the stylized images people are going gaga over — essentially stealing from previously existing artwork to throw a wrench in the market for future freelance work while failing to compensate said artists for co-opting their skillsets.
And they’re making a stack while doing it.
In the last three days, Lensa AI has soared to the number one spot on the Apple store and the Play store with a reportedly significant hike in revenue. With the affordability of $8 for a set of 50 stylized images, it’s no wonder freelance artists are fretting about their income on top of the fact that Lensa has already taken their artwork to inform its AI output for profit. Ironically, the San Francisco Ballet recently published a marketing post on Instagram using AI-generated art, so it’s already happening before our eyes.
Memes about Lensa have been bouncing around on social media feeds as well, making light of the app’s ability to enhance our normal, human faces in the name of digital art. And to be frank, many of the generated images are beautiful and have an uncannily human touch to them.
But there appear to be a few hiccups in the output so far. Some users have claimed that the “Magic Avatar” feature lacks racial nuance, producing portraits that either reflect an overarching “whiteness” or feature exaggerated racialized phenotypes in a way that definitely raises eyebrows. Now, DALL-E 2 and other AI image generators have a disclaimer noting that since their systems are trained on unfiltered web content, their output images may inadvertently convey negative stereotypes and biases. One user also pointed out that the output images lean heavily into the male gaze.
Lensa has not yet responded to Hyperallergic’s request for comment.