This could block text-to-image AI models from ripping off artists
- Reference: 1676550911
- News link: https://www.theregister.co.uk/2023/02/16/computer_scientists_develop_new_technique/
- Source link:
Commercial text-to-image tools that automatically produce images given a text description like DALL-E, Stable Diffusion, or Midjourney have ignited a fierce copyright debate. Some artists were dismayed to find out how shockingly easy it was for anyone to create new digital artworks mimicking their style.
Many have spent years perfecting their craft only to see other people generate images inspired by their work in seconds using these tools. Companies developing text-to-image models often scrape data used to train these systems on the internet without explicit permission.
[1]
Artists are currently embroiled in a proposed class-action [2]lawsuit against AI startups Stability AI, Midjourney, and online art platform DeviantArt, claiming they infringed on copyright laws by unlawfully stealing and ripping off their work.
[3]
[4]
Artists could protect their intellectual property from image generation tools in the future using new software developed by computer science researchers at the University of Chicago. The programme, dubbed Glaze, prevents text-to-image models from learning and mimicking the artwork styles in images.
First, the software inspects a picture and figures out what visual details define its qualities. Traditional oil paintings, for example, will contain fine brushstrokes, whilst cartoon drawings will have more exaggerated shapes and colour palettes. Next, these features are altered by applying an invisible "cloak" over the image.
[5]
We don't need to change all the information in the picture to protect artists, we only need to change the style features," Shawn Shan, a graduate student and co-author of the [6]study , [7]said in a statement. "So we had to devise a way where you basically separate out the stylistic features from the image from the object, and only try to disrupt the style feature using the cloak."
The cloak is, in fact, a style transfer algorithm that applies another image's likeness onto the features extracted from the programme. Glaze basically remixes the original appearance of an image with another style so that an AI model trained on the image fails to capture its essence effectively.
Here's an example of artwork from three artists, Karla Ortiz, Nathan Fowkes, and Claude Monet that have been cloaked with different styles from Van Gogh, Norman Bluhm, and Picasso.
[8]
The columns on the right show how much an image can change using the Glaze programme, with the left column altered less than the right column
"We're letting the model teach us which portions of an image pertain the most to style, and then we're using that information to come back to attack the model and mislead it into recognizing a different style from what the art actually uses," Ben Zhao, co-author of the research and a computer science professor, said.
The changes made by Glaze don't affect the appearance of the original image much, but are interpreted differently by computers. The researchers are planning to release the software for free so artists can download and cloak their own images before they upload them to the internet, where they could be scraped by developers training text-to-image models.
[9]Midjourney, DeviantArt face lawsuit over AI-made art
[10]So you want to replace workers with AI? Watch out for retraining fees, they're a killer
[11]Adobe will use your work to train its AI algorithms unless you opt out
[12]Adobe to sell AI-generated images on its stock photo platform
They warned, however, that their programme doesn't solve AI copyright concerns. "Unfortunately, Glaze is not a permanent solution against AI mimicry," they [13]said . "AI evolves quickly, and systems like Glaze face an inherent challenge of being future-proof. Techniques we use to cloak artworks today might be overcome by a future countermeasure, possibly rendering previously protected art vulnerable."
"It is important to note that Glaze is not a panacea, but a necessary first step towards artist-centric protection tools to resist AI mimicry. We hope that Glaze and followup projects will provide some protection to artists while longer term (legal, regulatory) efforts take hold," they concluded. ®
Get our [14]Tech Resources
[1] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2Y@5hNp7jJtRBcSgxwS0XqAAAABg&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[2] https://www.theregister.com/2023/01/16/stability_diffusion_lawsuit/
[3] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Y@5hNp7jJtRBcSgxwS0XqAAAABg&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33Y@5hNp7jJtRBcSgxwS0XqAAAABg&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[5] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Y@5hNp7jJtRBcSgxwS0XqAAAABg&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[6] https://arxiv.org/abs/2302.04222
[7] https://news.uchicago.edu/story/uchicago-scientists-develop-new-tool-protect-artists-ai-mimicry
[8] https://regmedia.co.uk/2023/02/15/glaze_ai_demo.jpg
[9] https://www.theregister.com/2023/01/16/stability_diffusion_lawsuit/
[10] https://www.theregister.com/2023/01/29/ai_retraining_costs/
[11] https://www.theregister.com/2023/01/07/adobe_ai_training/
[12] https://www.theregister.com/2022/12/06/adobe_ai_images/
[13] http://glaze.cs.uchicago.edu/index.html#whatis
[14] https://whitepapers.theregister.com/
Because neural networks don't "take inspiration", they match noise patterns. It is, by definition, derivative. A more appropriate comparison would be kitbashing, using pieces of existing work as brush textures in digital art.
The brain does something similar.
It can do pattern matching, but that's not all it does. It also has pattern formation for instance, seeing patterns that just aren't there, e.g. constellations.
Machine learning can also see things that aren't there, such as dogs where there are actually buns.
Copying has long been an established part of artistic tuition. Imitation is the sincerest form of flattery . But style isn't, and shouldn't, be copyrightable.
It's doing that by mistake.
AI is not being 'influenced.' AI cannot bring the experiences that humans bring.
Surely selling the art as an NFT would stop this...
Which Craft?
I'm not in the art business. Some of this stuff I see doesn't look like the artist spent a lot of time perfecting their skill. Could just be me who does not know what to look for among the paint stipes or bent metal.
Copyright Immunity?
Maybe the AI generated image file should include a detailed AI version number and the descriptive text used to create it. If the description references a single specific artist or specific art work then copyright maybe infringed. If not, it should be immune from copyright claims.
This assumes the version number and description can be used to reproduce the disputed work and that the training set can be show to include a variety of artists.
One line
The one line artists hate:
image = decloak(image)
Re: One line
IMO, this one-liner image has more artistic flair to it: f c ( z )= z 2 + c
Re: One line
> image = decloak(image)
Ah, So that is how they finally saw through the Romulan defences.
That ship has sailed, adapt or die
Those who want to, will be able work around this. Basically, it's time to admit, as even Picasso did*, that the fakes have won. Style isn't copyrightable.
In "F for Fake" Picasso famously claims to be able to fake his own pictures.
What are those samples meant to prove?
Seems to me they show that training on one image isn't as good as on two. They also use different artists as examples.
I love AI
I've been enjoying AI for years now with my girlfriend but these days she just tells me to go one hole higher ... this has been the end of Anal Intercourse for me.
The thing many people seem to forget...
....is that image generative AIs don't have to be trained on a certain artists images to mimick that artists style.
The end goal for such ML models is to generalize. Given the gigantic pile of imagery that humanity produces on a daily basis, there is simply enough training data to build models that will eventually generalize to solutions that encompass all art styles, past, present and future.
The only difficulty is then for the prompt engineering people to figure out how to describe a certain style if they can't do it by simply putting "in the style of {artist name}" into the prompt. Once the prompts are found that describe a style to the model, it's simply a matter of fine tuning.
What's missing from those example images
is a version of the original image with cloaking applied, so we can verify that it really does not change the image much for a human viewer.
Derived versions
It has long been the case that one can publish a summary or a paraphrase of a written document. Copyright applies only to the exact original text. Newspapers will thereby reproduce the scoops of their rivals.
The same will apply to pictures and to music. Anyone can sing their own version of a pop hit. Artists are on to a loser if they try to stop that.
If AI cannot study things and be "influenced" by it, shouldnt the same be done for the human brain.
Every singer who ever said Elvis inspired them should have their songs deleted. Every 'tribute' film etc....
why can AI not be inspired by stuff but wetware can? just because AI is better at it?