Okay so big nerd coming in here with my tech knowledge and about a year of obsessively researching AI systems especially with image generation in a quest to understand it so I could finally feel better (it's just what I do when something scares me ok). I saw this coming a mile off actually! I'm going to break down how this works and why it is like this. This one is long so buckle up, get a coffee or something.
Nightshade and Glaze are basically running on similar infrastructure to Image Generational AI. Nightshade and Glaze ARE Image Generational AI. Here's why -
The best metaphor I've heard to explain this is called 'the Chinese Room'. Imagine you are in a room full of shelves of books in Chinese (or in the case you do speak Chinese, any language you do not speak and cannot read) and all the spines of the books are labelled with numbers. On the table you have a huge reference book. No one is in the room with you and the only interaction you have is through a slot in a door, where someone passes you a note with a sentence written in Chinese. You have no idea what it says, or if it's a statement or a question. But what you do have is the reference book. When you open the book it has sets of phrases - also in Chinese - but with various numbers next to those phrases which reference one of the books on the shelves, right down to a specific passage. So to reconstruct your 'answer' to this note passed through the door, you go find each relevant book and passage, copy down the symbols you don't even understand and pass it back through the door. The person on the other side then either reacts positively or negatively, the negative reaction setting off a shift in the books that causes them to reshuffle into a better configuration. The person on the other side of the door though? They're now CONVINCED you're fluent in Chinese. This is how AI works.
An AI does not see your question, or statement, or a picture of a sunset. It sees data, and patterns within that data. It knows from 'reading' more books than anyone one ever could in their lifetime, if a sentence starts with the word pattern of 'Once upon' it's statistically likely to be followed with 'a time'. When AI in it's current form replies to your requests, no matter how convincingly - it is not sentient or sapient, it is merely producing a selection from its vast swathes of data that best replicates the patterns it has already seen according to the feedback it's been given on how accurate that data reproduction was. Image generation works in a similar fashion: it has seen more photographs of a sunset than anyone could in their lifetime, knows it usually is going to be in a specific type of composition with specific colours. It's basically looking at each pixel and following with the next statistically likely pixel to put down based on the input 'a photograph of a sunset'. The more pictures it 'sees', the better the pattern it has on file becomes.
Now naturally, this requires a lot of 'brain power' to run, especially if you don't want it to take a million years. Currently most image generative AI utilizes Graphics Cards, also known as Graphics Processing Units or GPUs for those who aren't computer savvy - since with successive generations of games being put out they've gotten really efficient at spreading the processing load of large chunks of data - this is how it shows you your nice interactive pretty pictures when you play a videogame without crashing your entire PC by hogging the CPU (Central Processing Unit; if you have anything close to a computer - including your phone - it's got one of these. This is the workhorse brain of your PC).
Another thing GPU's have in common with Computers and their CPU configuration, is RAM or Random Access Memory, called VRAM or Virtual Random Access Memory. To use a neuroscience analogy - you could consider both RAM and VRAM the 'working memory' of the machine. It needs a space BIG ENOUGH to hold its entire 'thought' or in this case, dataset - to sift through when you put in a request. Depending on how big the request is, it might have to split it into chunks, or in the case of images, quadrants - because that is the maximum size it can process at a time, then stitches them together with varying results. This is also why AI isn't very good at writing a continuous narrative like a long novel, because it literally cannot remember all the information (yet).
So this is why for the short time that we've had AI, if you wanted to run the image generation AI Stable Diffusion on your PC locally for example, the basic requirements were something like 6gb to 8gb of GPU VRAM just to produce a 512px square image. Anything bigger either had to be stitched together, with mixed results because it can only hold a 512px square in memory at time. When it starts a new square, it mostly forgets the previous square. Nor can it see the whole overall 'picture' at once, so you get a kind of patchwork of something that SORT OF works together but at the same time is limited to each patchwork piece where output data may have been slightly different depending on the limited input it had.
Stable Diffusion has released its second iteration which can now handle a 768px square but requires - you guessed it, more memory. It's now recommend to run that on at least 12gb VRAM Graphics Cards which unless you are REALLY into cutting edge gaming on PC or do a lot of video/3D rendering and have the cash to burn - you probably don't own.
Glaze and Nightshade are basically the exact same thing as Stable Diffusion - an image generative AI, with all the same requirements and pitfalls. The difference is, Glaze and Nightshade are weaponized variants of this system, where it's only task is to reproduce the visual appearance of your art while injecting data that works against the architecture of current AI Image Generators, so when fed into one of these generators to train it, it actively 'poisons' the dataset by giving it junk data.
I haven't looked too deep into Nightshade or Glaze yet, but based on the fact we know that AI cannot be trained on itself, because it results in basically the entire dataset breaking down into an unintelligible mess - Nightshade/Glaze is doing something similar; that hypothetically it is basically 'AI-ifying' your submitted artwork in such a fashion that it could be recognized by machines as AI generated (and will likely exclude it from the scrape) and if it doesn't - will actively poison any dataset when scraped, in turn breaking any training done with the dataset for future AI models. Get enough poisoned works in there, you might break a trained model so badly they'll have to scrap the entire thing and start from scratch.
So, what can you do if you don't have a big brain VRAM GPU at your disposal? Well as OP said, there are online hosted options will probably have to pay since you're essentially renting a Graphics Card, and also the server space to upload your art to process your request. This also can be effected by how many people are using the service at once, so you might have to deal with a queue before you can even put your art through the process. Even then the process itself IS long - honestly 20 to 30 minutes is pretty much the average time for processing an image and while I don't know about Glaze/Nightshade specifically, it's likely that the bigger the image? The longer the processing time. This is because the image you're submitting isn't having some kind of 'layer' applied, as the term Glaze might imply - it's essentially having to reconstruct your artwork, pixel by pixel while injecting the data it's been trained on to break other datasets AND make it look like nothing perceptible has happened. That is a lot of work! So it's going to take a lot of memory and in the case of less memory - an even longer time.
Unfortunately there is little out there in terms of non-paid options for this, basically: Glazing/Nightshade will require computational resources and those computational resources will have to be paid for in some way, either by you buying a new GPU for your PC or renting one. There's no real way around it. There is one option I can think of which is using Google Colab to host it for you much like some people do with Stable Diffusion (yeah when I said I did a deep dive on this shit, I meant it) but I'm not the coding type of tech nerd so I wouldn't be able to tell you how to do it but I'd assume the process is fairly similar to what I've linked since Glaze/Nightshade like Stable Diffusion are designed to be run locally. Google Colab is free initially, but requires subscription once your computational requirements get bigger.
Now there is one silver lining in this whole shitshow, and she's a doozy. AI scraping is about to hit a content wall. It is predicted that by 2026 90% of all content on the internet will be AI generated. While this does initially sound like a dystopic hellscape, remember what I said about AI being unable to train on itself? While releasing image generators to the public brought them big bucks initially, these AI companies have basically released the equivalent of a plague on the internet and it is infecting everything rapidly. So either they have to try and find data pre-2022 that they haven't already scraped (which probably explains the under-the-table deal with OpenAI if Tumblr/Wordpress did in fact sell data to them without telling us), or find new 'good' data to train on. That new good data? THAT'S YOU. YOU'RE THE GOOD DATA. HUMAN GENERATED CONTENT IS THE GOOD DATA. This goes for everything, from your shitty meme doodles, beginner sketches, mindless art studies, shitposts - this includes text and comments by the way for language models! - are all, extremely valuable to these AI companies and they're about to lose their fucking leverage.
With Nightshade and Glaze out there, unless these web-crawlers are specifically configured to recognize Nightshade/Glazeand exclude it - these crawlers have all the finesse of a shotgun blast when they suck up data. There's so much information in there that any living human being could not possibly, in their entire lifetime review all of the data held within the dataset. Once something gets into a trained dataset, it can be extremely hard to find and remove. And when AI companies can't improve their models and impress their investors because they've run out of Data - they're going to get desperate. They might actually start approaching people, like they should have done in the first fucking place and ask for permission with compensation! You could get paid for participation in creating these datasets. Heck you could maybe even demand royalties! If your bit of code was detected in contributing to the creation of an image - you get a kickback. I am dead serious when I am saying the balance is now shifting back in our favour and these companies have hell to pay - and it is our duty to make sure they pay for it.
So, in TL;DR summary for those who don't like text walls:
AI cannot train on itself without degrading. The internet is rapidly filling up with AI generated content. It has likely scraped anything before 2022 and to improve their models they need new 'Good Data'.
Good data is human data - and that is all data, from your text shitposting to memes. Social Media has been financially a losing game for years until the AI arms race started - and now 'content' made by humans is a goldmine of training data to sell to corpos. You need to start recognizing the value of what you produce when you post online.
Glaze and Nightshade have big requirements to run because they are structurally the same as Image Generative AI, just weaponised in the Artist's favour. It's turning your good data into not just 'bad' data, but poisoned data.
AI Image Generation is resource heavy and computational resources cost money, whether from your own pocket or rented. If they don't cost money, the cost will be time. Bigger memory resources = faster work + more expensive. Smaller memory resources = slower work + less expensive.
While retroactively running your art through Nightshade and Glaze is still a good thing to do if you can, it is likely in some way or another your data has already been scraped, and if not directly, by third parties where it has been shared elsewhere. It's now out of your control - so don't feel bad if you don't feel you can Glaze/Nightshade all your existing work.
What you DO have control over is what you put out there going forward. I'd say it is not only your imperative as an Artist but your duty to Nightshade and Glaze if you can. By doing this you put the control back in the hands of the artist and state to these companies, loudly, that if they want good data they have to pay for it like they should have in the first place. Poison anything you release publicly, if not just to protect it, but to break these systems built on a mountain of greed. Do not give them anything of yours for free.
If you cannot poison your art, and you know that a website might end up selling your data to one of these companies - assess whether your art being sold is worth the equivalent to your experience on the website. If not? HIT DA BRICKS. DON'T GIVE THEM ANYTHING OF YOURS FOR FREE.
PC GAMERS! YOU CAN HELP FIGHT AI!
Yeah I am serious - if you have a beast rig or a mini monster rig with at least a 4gb VRAM NVIDIA Graphics Card in your machine and have some time to kill doing something that isn't graphics card heavy (games, videos for e.g.) or are willing to leave your PC running while you go eat Lunch or something - if you have an Artist friend out there who doesn't have access to the same resources, please consider donating some of your time to run their artwork through Nightshade and Glaze. You'll actively be helping all Artists out and keeping money out of the hands of any money-grubbing corpos who either wanna steal people's work for their own gain or capitalize on the current panic of Artists who just want to protect their art and feed themselves.
Anyway fuck corpos and poison the machine. 🤘
