If you're talking about training from scratch and not fine tuning, that won't be cheap or easy to do. You need thousands upon thousands of dollars of GPU compute [1] and a gigantic data set.
I trained something nowhere near the scale of Stable Diffusion on Lambda Labs, and my bill was $14,000.
[1] Assuming you rent GPUs hourly, because buying the hardware outright will be prohibitively expensive.
That argument also makes little sense when you consider that the model is a couple gigabytes itself, it can't memorize 240TB of data, so it "learned".
But if you want to create custom versions of SD, you can always try out dreambooth: https://github.com/XavierXiao/Dreambooth-Stable-Diffusion, that one is actually feasible without spending millions of dollars on GPUs.
1. https://www.youtube.com/watch?v=cdiD-9MMpb0 Lex Fridman podcast with Andrej Karpathy
P.S. To counteract that (unintentionally actually, likely because of a simple optimization of instruments' duty cycle) in astronomy people come up with a concept of "observatory" (Like Hubble, JWST) instead of "experiment" (like LHC, HESS telescopes) where outside people can submit their proposals, and if selected get observational time. Along with raw data authors of the proposals get required expertise from the collaboration to process and analyze that data.
This is just lossy compression with a large and well-tuned (to the expected problem domain) dictionary.
Video compression codecs can achieve a 500x compression ratio, and they are general-purpose.
Uncompressed, LAION-5B would be 4PB, for a compression ratio into SD of ~780kx, or one byte per picture.
The only practical limit is the amount of information entropy in the source material, and if you're going to claim that internet pictures are particularly information-dense I'd need some evidence, because I don't believe you.
Both are just simple statistical relationships between parameters and random variables.
Most human behavior is easy to describe with only a few underlying parameters, but there are outlier behaviors where the number of parameters grows unboundedly.
("AI" hasn't even come close to modeling these outliers.)
Internet pictures squarely falls into the "few underlying parameters" bucket.
We can speculate they apply to certain models of slices of human behaviour based on our vague understanding of how we work, but not nearly to the same degree.
When a human looks at a picture and then creates a duplicate, even from memory, we consider that a copyright violation. But when a human looks at a picture and then paints something in the style of that picture, we don't consider that a copyright violation. However we don't know how the brain does it in either case.
How is this different to Stable Diffusion imitating artists?
> That argument also makes little sense when you consider that the model is a couple gigabytes itself, it can't memorize 240TB of data, so it "learned".
The matter is really very nuanced and trivialising it that way is unhelpful.
If I recompress 240TB as super low quality jpgs and manage to zip them up as single file that is significantly smaller than 240TB (because you can), does the fact they are not pixel perfect matches for the original images mean you’re not violating copyright?
If an AI model can generate statistically significantly similar images from the training data, with a trivial guessable prompt (“a picture by xxx” or whatever) then it’s entirely arguable that the model is similarly infringing.
The exact compression algorithm, be it model or jpg or zip is irrelevant to that point.
It’s entirely reasonable to say, if this is so good at learning, why don’t you train it without the art station dataset.
…because if it’s just learning techniques, generic public domain art should be fine right? Can’t you just engineer the prompting better so that it generates “by Greg Rutkowski“ images without being trained on actual images by Greg?
If not, then it’s not just learning technique, it’s copying.
So; tldr: there’s plenty of scope for trying to train a model on an ethically sourced dataset, and investigation of techniques vs copying in generative models.
It is 100% not something we can just brush off.
If you compress them down to two or three bytes each, which is what the process effectively does, then yes, I would argue that we stand to lose a LOT as a technological society by enforcing existing copyright laws on IP that has undergone such an extreme transformation.
Does that mean it’s worthless to try to train an ethical art model?
Is it not helpful to show that you can train a model that can generate art without training it on copyrighted material?
Maybe it’s good. Maybe not. Who cares if people waste their money doing it? Why do you care?
It certainly feels awfully convenient for that there are no ethically trained models because it means no one can say “you should be using these; you have a choice to do the right thing, if you want to”.
I’m not judging; but what I will say is that there’s only one benefit in trying to avoid and discourage people training ethical models:
…and that is the benefit of people currently making and using unethically trained models.
You couldn't teach a human to do that without them having seen Greg's art. There are elements of stroke, palette, lightning and composition that can't be fully captured by natural language (short of encoding a ML model, which defeats the point).
However, copyright doesn't prevent someone to look at the work and study it. Even study it by heart. Infringement comes only if that someone would make a reproduction of that work. Also, there are provision for fair use, etc.
Is it fair to hold it to a higher standard than humans though? To some degree it's the whole "xxx..... on a computer!" thing all over again if we go that way
Can you please rewrite this in the writing style of Socrates?
Harping about copyrights in the Age of Diffusion Models is unhelpful (for artists) like protesting against a tsunami. It's time to move up the ladder.
ML engineers have a similar predicament - GPT-3 like models can solve at first try, without specialised training, tasks that took a whole team a few years of work. Who dares still use LSTMs now like it's 2017? Moving up the ladder, learning to prompt and fine-tune ready made models is the only solution for ML eng.
The reckoning is coming for programmers and for writers as well. Even scientific papers can be generated by LLMs now - see the Galactica scandal where some detractors said it will empower people to write fake papers. It also has the best ability to generate appropriate citations.
The conclusion is that we need to give up some of the human-only tasks and hop on the new train.
Oh and also I second the fast.ai suggestion, part 2 is 100% focused on implementing stable diffusion from scratch in the python standard library and it's amazing all around. The course is still actively coming out but the first few lessons are freely available already and the rest sounds like it will be made freely available soon.
I trained from scratch with 4x3090 and while it’s not as good as SD it’s surprisingly better with hands.
I particularly interested in the image generation part (the DDPM/SGM)
There’s code on my GitHub (glid3)
edit: The architecture is identical to SD except I trained on 256px images with cosine noise schedule instead of linear. Using the cosine schedule makes the unet converge faster but can overfit if overtrained.
edit 2: Just tried it again and my model is also pretty bad at hands actually. It does get lucky once in a while though.
I use an open air rig like the ones used for crypto mining. 4x3090 would normally trip the breakers without mods but if you under volt the cards the power draw is just under the limit for a home AC outlet.