For many, AI art generators such as Misjourney and Dalle-2 are fun tools that can create impressive pictures from just a prompt. But for independent artists trying to make a living by selling their work on the internet, they represent an existential threat.
With enough training data (i.e: publicly available pictures), these tools have been shown to effectively mimic a specific artist’s style. This allows potential customers to get a knock-off picture of their choice without the artist being compensated.
Artists could avoid this by not showcasing their work online. But, without the exposure, commissions dry up anyway. It’s a difficult catch-22: either keep work private and struggle to gain publicity, or display it freely and risk an AI destroying your income.
Now researchers from the University of Chicago have come up with a third way: Glaze. It’s a free app that promises to confuse AI trained on your pictures into thinking it’s something wholly different, thus making any custom art based on your work seem completely unrelated.
As this image taken from the research paper shows, minor alterations to the images can result in wholly different results; with attempts to mimic Karla Ortiz, Nathan Fowkes and Claude Monet generating art in the style of Vincent van Gogh, Norman Bluhm and Picasso respectively.
It does this by applying “style cloaks” to images before they’re shared online: “barely perceptible perturbations”, in the study’s words. How “barely perceptible” are we talking here? Well, as the examples below show, the difference is visible, with the art taking on a slightly grainier quality. But if you’re an artist merely wanting your work in the shop window, that seems like a small price to pay for the protection Glaze provides.
“The cloak is not some brittle watermark that is either seen or not seen,” the researchers explain in an FAQ on the Glaze site. “It is a transformation of the image in a dimension that humans do not perceive, but very much in the dimensions that the deep learning model perceive these images.” In other words, it’s very hard for AI models to get around without “significant changes in the underlying architecture of AI models.”
As effective as the paper makes it seem, to many artists this may seem like shutting the barn door after the horse has bolted: if their unprotected artwork is already out there, then Glaze can’t help.
That would be true if AI art apps have completed their training. However, most are continuing to analyse new work as it’s produced, and the researchers believe that applying Glaze to future work could dilute the quality of work returned by AI generators in future.
“The more cloaked images you post online, the more your style will shift in the AI model’s feature space, shifting closer to the target style (e.g., Van Gogh’s style),” the FAQ explains. “At some point, when the shift is significant enough, the AI model will start to generate images in Van Gogh’s style when being asked for your style.”
If this becomes used widely enough to significantly disrupt AI art generators, the researchers concede that developers will attempt to sidestep the protections put in place by Glaze.
“Any technique we use to cloak artworks today might be overcome by a future countermeasure, possibly rendering previously protected art vulnerable,” the discussion section of the paper says. It adds that the researchers are under “no illusion that Glaze will remain future-proof in the long run”.
But at the very least, it’ll buy time before longer-term and slower-moving legal and regulatory protections for artists can come into force.
