I've been following something that I hope most of you (and especially people like Bill Cipher) will appreciate quite a lot. It's called Nightshade, and it's an optimized prompt-specific data model poisoning attack. The full paper on it can be found here - Prompt-Specific Poisoning Attacks on Text-to-Image Generative Models

One of the masterminds behind this technology is Ben Y. Zhao who seems to have authored quite a few research papers fighting against the rise of these ML algorithms that have been poisoning the online art community. He's also developed a tool called Glaze whose purpose is aiding artists in the fight against products like Stable Diffusion.

The way technologies like Nightshade and Glaze work is by changing the pixels of images in subtle ways that are invisible to the human eye but manipulate machine-learning models to interpret the image as something different from what it actually shows. Artists can use these tools to parse their finished art through, and have the modifications automatically applied in a way that would disrupt the learning capabilities of these models. Attached you will find a couple of images that depict the effects of those poisoned samples on the models' performance.

Being a software engineer myself, I very much respect the fact that the companies developing these technologies are making them open-source, allowing the community to modify the code and customize it to specific edge cases, as well as develop it beyond the original creator's vision and reach. Models like LAION, the one used in Stable Diffusion, contain billions of images scraped from the internet. If we can hope for any meaningful effect, we need as many poisoned samples as possible. Unfortunately, affecting the images already used in training those models is not an option, but having those technologies available will decrease the impact of these models in a (hopefully) noticeable way in the future. At this point we will take anything we have at our disposal in what feels like ultimately a losing battle against the commercialization of one of humanity's most valuable peculiarities.

The poisoned data is very difficult to remove, as it requires tech companies to painstakingly find and delete each corrupted sample, which is just unrealistic consindering the enormous amounts of data these models work with. And implementing an automated algorithm that scrapes these databases and tries to detect entries that have been tampered with is also unrealistic. It's similar to asking a person struggling with severe psychosis and schizophrenia to describe what is real and what isn't. In the end you'll get a dataset that you can't really rely on.

Generative AI models are excellent at making connections between words, which helps the poison spread. Technologies like Nightshade infect not only the word “dog” but all similar concepts, such as “puppy,” “husky,” and “wolf.” The poison attack also works on tangentially related images. For example, if the model scraped a poisoned image for the prompt “fantasy art,” the prompts “dragon” and “a castle in The Lord of the Rings” would similarly be manipulated into something else.

Naturally, people like Vitaly Shmatikov, a professor at Cornell University who studies AI model security, are opposing this type of technologies and are urging for focus on defense against them, but that's to be expected.

And on the other side (thankfully) we have people like Junfeng Yang, a computer science professor at Columbia University, and Eva Toorenent, an illustrator and artist who uses Glaze, and Autumn Beverly, another artist that has been using these tools, all say the presence of this type of technology is a step in the right direction:


Real arists who have previously removed their work from the internet are bringing it back, but parsed through tools like Nightshade and Glaze, and say it has brought their confidence and sense of safety back, since models like LAION won't be able to unconsensually scrape their work and provide it to others for free.

We have yet to see the extent of these tools' impact on ML models that generate "art", but I believe this is a step in the right direction as well. Artists should not have to fear posting their work online. Art is a fundamental expression of human creativity, and a materialization of the human soul. Commercializing that and making it nothing more than a commodity that fuels the engine of capitalism should be considered a crime against humanity.

Let me know your thoughts on this.

Love & Light

Sounds like evolution, a plant that now grows thorns, a creature that now produces poison...
because the creature feels attacked.

Fascinating.

I wonder if this ecourages the other side of the equation to evolve too.

I feel like the issue goes way deeper though, why do we feel like we have to protect something?
The word "steal" often comes up in these topics.
A personal attack that reaches all the way down to the roots of survival and identity.

I feel like people, including myself, often associate art with the most natural and precious aspect of human experience. In that sense, an emerging technology that threatens to commercialize that aspect, thus decreasing its value by sheer volume and lower quality, is seen as a natural threat that needs to be resisted. And I have to admit I agree, only so far as actual artists are affected. If their consent is taken into account, then all is good, because these algorithms obviously allow people with no actual artistic inclinations or abilities express themselves in a way that was hitherto largely inaccessible.

It is an interesting situation indeed. The very algorithms that created this issue are, in a sense, also solving it. A sort of technological autoimmune disease. And we're witnessing the very beginning of something big that will forever change the artistic landscape.

I wonder how all of this will play out.

The problem is that Art is commercialized and reduced to a mere commodity. The problem is that the very survival of the artist depends on their ability to employ and market their creative capabilities. Freed from the yoke of capitalist exploitation, art could transform into something like a fundamental aspect of human existence, like cooking, cleaning, praying, going for a walk, watching a sunset etc. Art could thus become much more than a mere performance. It then wouldn't matter whether it's created by humans or animals or the AI, art would simply be art, irrespective of its origin.

But yeah, these software can at least help the artists in their survival... which is still a good thing even if limited in its scope.

Thank you, I like great visions! One possibility, although almost as utopian, is that we turn away from our screens. Maybe to enjoy the art of our immediate surroundings.

I am, of course, all for this and fully intend to use it once it's available (it's not yet).

I'd be even more enthusiastic if it also deposited a virus on each user's computer that destroyed their operating systems, but I'll take what I can get and hope that it becomes as destructive a force as possible.



That ain't the problem, and yes, the source of creation would absolutely still matter - precisely because art is absolutely a fundamental aspect of human existence.

Generative AI trivializes and degrades that. It's much more than a monetary concern.

I too hope it becomes as destructive as possible, but I'd prefer if that destruction remains solely on the software level. There are many, and I mean many, people that download and install this type of software without knowing the implications or impact it has. Many of them don't know that the language models behind these algorithms are trained on work whose owners have not given consent about. In that sense, destroying their hardware just because they were uneducated (as opposed to educated and holding a malicious or purposefully ignorant standpoint) would cause more harm than good.

I guess time will tell how this kind of technology develops. But it is definitely a step in the right direction. I'm still thinking about potentially malicious uses of this type of algorithm. Especially so where similar pattern-recognizing software is being used to process data for the justice systems (think fingerprints). Hmmm...

This AI thing is nutz.

I don´t think this "tainted image virus" system is going to work... a simple reprocessing of the image would allow to avoid it.
Also, I´ve been digging civit AI Loras (check image below, those are "styles" you can download, not iamges) and advanced setups in comfyui. They already have Alex Grey or Android Jones... and many others.

LINK LORAS
Psychedelic LORAS

Some tools like the IP adapter... let you copy the style of an artist, with a single image, or a combination of 4-6 images... [/img]
It doesn´t require to train a LORA, it doesn´t require to train a model/checkpoin.
If somebody wants to copy you, nothing can stop it, people can take 4 images or pictures of a certain artist... nothing to be done, sorry.

In the other hand...
I think we are moving towards and era were intelectual property will be worth poop...
Everything will be available to everybody at the same time...
Manual or service works will be revalued work done with hands, watch last southpark episode/special.

If you think about it... sounds like quantum conscience.
Artist take inspiration from entities and from higher dimentions...
The art is not yours anyway... you take it from the source... from morphogenetic fields, why do you need copyright and money?

...well, we need money for living, if you had the money issue sorted out...
will you still work in art?

if you had all the money you need... what would you do?

The answer would be yes for many.

This topic is paradigmatic, society is definetely moving towards a new model...