Glaze tool messes with digital art so AIs can’t steal your style

By this stage, you’ve probably played with an AI image generator. Services like DALL-E and Midjourney can create extraordinary images in response to simple text prompts. They’re a lot of fun, and like many AI tools, they’re improving and evolving at an incredible rate.

They are particularly good at imitating styles and incorporate more or less the entire history of popular art as part of their training regime. Feel free to order your own Picasso, Van Gogh or Monet. It’s okay to some extent. These creators, like Pablo, demonstrated their talent while they were alive and did very well, or died penniless like Vincent and Claude, leaving others to profit from their groundbreaking work. is dead for a long time.

Of course, there are many artists who are not dead and have spent their lives creating original works and styles. AI does not discriminate. They collect tons of images and scrape them off the net. The more popular a work becomes, the more likely it is to be eaten by AI, digested and excreted thousands of times, with zero credit or copyright to the original creator.

Admittedly, these AIs are just an accelerated version of what young artists have always been doing, trying to adapt the styles and techniques of established artists as they strive to find their own voice. There is an argument that they are trying to imitate and develop. But when it’s done at the speed of AI, it’s possible to generate derivative works on an industrial scale.

Legality is beginning to be tested. A class action complaint was filed against Github, Microsoft, and several OpenAI entities in November, claiming that many of Github’s open-source AI training libraries contain stolen images and that attribution, copyright, and It claimed that the license terms had been stripped and these were being used. Generates significant commercial value.

And while individual creators are armed, stronger hitters may soon take action and attack on the other front. Have fun creating images of 6 Minions® partying with Buzz Lightyear®, Batman® and Bender® in the . In the near future, AI will be able to create animated footage, reconstruct and overdub captured audio, and have familiar characters say and do whatever they want in new, fully customized videos. You will be able to Large copyright holders must act to stop such things.

In the meantime, a team at the University of Chicago has created something of an interim solution for digital artists. The Glaze app allows artists to upload and make subtle changes to their work before publishing it, as well as AI training her repository to steal her collectors.

Nearly invisible to the human eye, these “style cloaking” changes are designed to fool the pattern detection capabilities of AI image tools. They make subtle changes to the pixel layout of each image to make the AI ​​think the new artist’s style is very close to the established artist’s style. For example, come up with something like Van Gogh or, worst of all, hybrids of old and new styles to ruin the result and prevent style theft.

It’s done pretty cleverly. First, Glaze takes images of new artists and uses style transfer AI to recreate them in the style of famous past artists. Then, using that style-transferred image as part of a computation, we “perturb” the original image towards the style-transferred image, maximizing spurious patterns that the AI ​​might pick up. while minimizing the visual impact of changes. human eye.

The team behind Glaze doesn’t expect this technique to work long term.

“Unfortunately, Glaze is not a permanent solution to AI mimicry,” the project website says. “AI is evolving rapidly, and systems like Glaze face unique challenges of being future-proof. could be desecrated, making previously protected art vulnerable.”

Glaze certainly has its counterarguments. Some lament the fact that tools like this can confuse algorithms and make AI image generators work worse for everyone. Others point out that Picasso himself is widely quoted as saying, “Good artists borrow, great artists steal.” sheds light on how similar in many ways are to these transformative algorithms, collecting visual, auditory and conceptual information from birth and continually reconstructing it to produce ‘own’ thoughts and outputs. is creating

There is an irony here, as yesterday’s news about Stanford University’s Alpaca AI showed. There is another group that should be worried about AI stealing jobs, and that is the AI ​​companies themselves, the Alpaca team uses OpenAI’s GPT language model to generate thousands of question and answer prompts. , and used it to tweak the open-source language model. The result is a new AI that can perform just like ChatGPT on certain tasks. Made for hundreds of dollars instead of millions.

Whatever the reaction from the public and AI companies, the Glaze team believes the app is a “necessary first step, pending a long-term (legal and regulatory) effort to take hold.” I believe there is. Beta versions of Glaze can be downloaded from the project website for Windows and MacOS.

Source: glaze