I’m really excited to learn that Nightshade just dropped for public use. I have been waiting since it was announced they were making it, and am curious to see what happens/what this means for LLMs (Large Language Models). Let me explain a little more (and the best I can) about what Nightshade is, and how it works. But before I do that, let’s talk about AI data scraping/LLMs, and The Glaze Project.
As you all know, “AI art” (I have such a hard time with this name because it’s not true artificial intelligence by any stretch of the imagination but I guess it’s stuck now) has been unleashed onto the world with tools like Dall-E, Midjourney, and Stable Diffusion, to name a few. The cat is out of the bag and it’s never going back in.
The problem is that these LLMs used everything available on the internet to train their models without any permission at all, with no way for the average user to determine if the end product contains copyrighted materials they should not be using. It has scraped images from hundreds of thousands (if not millions) of artists, both alive and dead, and even gamified the use of living artist’s styles (These LLMs are the subject of a large lawsuit, with damning evidence of programmers trying to keep “in the style of” prompts while removing any evidence of copyright infringement/theft… There’s a lot going on, and a spreadsheet of 5,000 artists for Midjourney alone was released on Jan 4th. Many of my friends are on this Midjourney list, and I myself have had hundreds of works scraped which I found via the haveibeentrained.com tool and opted out, but the damage is already done ).
AI generated materials themselves cannot be copyrighted (only human made items can be included in US copyrights at this point), so any AI-created artwork doesn’t belong to anyone… Or belongs to everyone, depending on how you look at it.
Still, people are using these LLMs to create AI art at an alarming speed, and models keep improving. Artists who are living and depend on their years of experience honing their craft (and simply became prompts!) are losing their art jobs to AI art and LLMs. On the other side of this, programmers, writers, and so many more are losing their jobs to LLMs as well, but we’re focusing on the art side for right now. Many of you already know this, and the exciting fad of making an instant piece of “art” via LLMs has faded.
The Glaze Project
I’ve spoke about Glaze in this blog before. Still, I’ll briefly sum it up: In traditional painting, finishing it with a coat of varnish helps protect the original pigments from UV damage, color fading, spills, discoloration, and minor scrapes. Digitally, Glaze does similar. A free program created by a small team at the University of Chicago, Glaze does this by applying a layer of pixel protection to your work that is unique to that work. These pixels help to “throw off the scent” of your work and honed style/technique, instead pushing the LLMs towards identifying it as something in the public domain category of art, like Van Gogh or Monet. Ideally, Glaze would be mostly imperceptible to the human eye viewing the work, though in practice this doesn’t always happen. Still, some form of protection against data scraping is better than none. Glaze can be run on your home computer, or be run via the Webglaze app. More information about how Glaze works is here, along with where you can download it.
Nightshade
Nightshade is an offensive program and works similar to Glaze by applying a pixel layer that should be rather indistinguishable by the human eye, but to LLMs, will actually cause model collapse if used in the right amounts. In the chart above, you can see that even if just 50 poisoned samples are included in the LLM’s data set, things start getting weird. I have heard the amount of poison samples used to start create unintended things is actually much lower than that.
For more about Nightshade and where to download it (it doesn’t have a web version like Glaze just yet), check out the official website.
What’s really interesting about this is that you can both Glaze and Nightshade your work, and it will be more or less impossible to tell if it’s going to destroy an LLM’s model. Picking out images that are causing destruction well after the fact will be a tedious if not impossible task. LLMs are already seeing levels of model collapse from AI generated content being fed back into the models.
Both of these tools are actively being worked on, and any time an AI bro claims to have thwarted or found their way around them, the team has worked swiftly to verify accuracy, and update their algorithms accordingly.
Unfortunately, the work for artists continues. Continually re-glazing and re-nightshading works for us to keep up with latest data protection standards is difficult, if not impossible to do, so these are just small stop-gap measures to try to keep our future works from being included in LLMs. Still, if it helps one artist feel more comfortable posting their works in their portfolio, it’s worth it.
