LIcense is too limiting
Hi there, I may have understood the license wrong, but this cannot be used for any kind of serious work, apart from research in university or for hobbyists, this is useless.
Will you open it to more serious / commercial work in the future?
Thanks.
Yeah, that neutered licencing guarantees that SD 1.5 and XL 1 will continue to reign supreme as king of the AI models. no matter how much "Better" the new models get it really doesn't matter if the licencing terms are so horrific no one wants to use it.
You can use it for commercial use if you pay 20$ per month. Unless you make more than a million a year using it (then it has to be negotiated based on your profits).
If your profit is too low to pay the basic commercial licensing you can ask for a grant.
https://stability.ai/license
Oh I noticed that and those are still awful licencing terms. You need to have a "Creator licence" for anything other than just messing around or Academic Research, I get having to pay that if I am using SD's Hardware, Or Hosting SD3 for other people to use as a service, but that's $20. Per Month. to use it on my own hardware. with my own electricity. There is a reason that 1.5 is still the most popular, its not censored and its got the most permissive licence.
Be Open Source. or Don't Be.
Oh I noticed that and those are still awful licencing terms. You need to have a "Creator licence" for anything other than just messing around or Academic Research, I get having to pay that if I am using SD's Hardware, Or Hosting SD3 for other people to use as a service, but that's $20. Per Month. to use it on my own hardware. with my own electricity. There is a reason that 1.5 is still the most popular, its not censored and its got the most permissive licence.
Ask Stability AI for a grant on a lower contribution, they will enjoy take some money even if it's less than their commercial base licensing, they desperately need it as they are about to go bankrupt within a few months.
Be Open Source. or Don't Be.
That's an other issue, as for sure it will be very sad to see all weights, except 1.5, disappear when Stability AI will close, just because they never provided the source code. And for sure if they claim SDXL, SD3 and other derivatives to be open source they should make the source accessible, so at least when they'll close it wont be bought by competitors in order to lock it so no one can use it at all.
But open source doesn't mean free. It costs about a hundred million to develop and train a model like SD3, this is not something doable if you don't monetize it in big proportions.
Ask Stability AI for a grant on a lower contribution, they will enjoy take some money even if it's less than their commercial base licensing, they desperately need it as they are about to go bankrupt within a few months.
The heart of Stability AI the CEO, is already gone, but I don't think they will go bankrupt, they are in a hole right now, that's for sure, which probably explains the new licencing model. my guess though they will be bought out by someone else, who might be better or worse, roll the dice on that one. Again I get why, but it feels grubby and as you said models are expensive so what is the point of spending a hundred million to develop a model no one will use if the licencing is too restrictive and designed just to squeeze those last few dollars out of you?
Also, Digging into it more, that "Creator Licence" only covers 6,000 images. (I think that's per month) which might sound a like a lot, and it is, but when your doing 10 at a time to get "the right one" you will burn through that fast.
Be Open Source. or Don't Be.
That's an other issue, as for sure it will be very sad to see all weights, except 1.5, disappear when Stability AI will close, just because they never provided the source code. And for sure if they claim SDXL, SD3 and other derivatives to be open source they should make the source accessible, so at least when they'll close it wont be bought by competitors in order to lock it so no one can use it at all.
But open source doesn't mean free. It costs about a hundred million to develop and train a model like SD3, this is not something doable if you don't monetize it in big proportions.
Yes but also no. the vast majority of Open Source projects I've ever worked with are either entirely free, or licenced in a way that if you host it yourself on your own hardware its free, if its used in the project but is NOT the project, its Free, even if that is a commercial project, now if you want a "managed" service, you pay for that because you are using someone else's hardware OR if you want to host it yourself as SAAS for other people to use as a managed service, you pay for that too. Even Meta's Llama 2 Commercial licence is less restrictive than this is.
I didn't saw this 6000 image limit, that's absurd indeed. If it was the number of commercialized images it would make sense, getting a good image to sell can take tens of inferences.
There's no way they can control that anyway...
When it comes to other open source projects, most of them do not cost nearly as much to produce. LLama 2 is a closer money burner, but it's still a different story as Meta has billions to throw at it without putting the company in jeopardy if it fails.
Stability AI's hole is the Mariana Trench... They didn't just lost Emad Mostaque : Robin Rombach, Andreas Blattmann, and Dominik Lorenz who were at the hearth of the engineering went away too, so did many other executives, and now 10% of their employees for "restructuration" (which only means they can't pay their salaries and have to cut like crazy). They also had to close some external companies they bought. Their actual loss is about 8 millions a month, they have been bleeding to death for a year, changes are too little too late and they are desperate. SD3 could have change the issue if it had provided all promises, it's impressive how light and performant it is vs SDXL, but it shows a lot of quality disparities, too much to be an instant economical success.
Time will tell but so far it smells extremely bad for Stability AI, and even though I also regret these license terms, I can understand why they push them this way. What I can't support is them not sharing the source code and by this puting in jeopardy other companies like CivitAI when they'll assume their bankrupcy, and of course our own capacity to use free AI for personal use.
Yeah, I agree Stability AI seems to have grown too far, too fast. and its cost too much on way to many projects and buying other companies. Didn't help much they've likely spent a truck load on Stability LLM which seems like a failure, and Stability Video AND now Stability Audio. if they are spending money they don't have, why are they spending so much of it NOT on Stable Diffusion?. like I would much rather them cut the dead weight (stable-LLM, Stable-Video) and knuckle down on the thing they are actually good at Stable Diffusion and Stable Audio Things not being done better by other companies until they are financially viable again.. These terms absolutely feel like one last roll of the dice to get as much money in as they can for the executives left before the fire sale. The Problem is that they very people they need to pay for SD3's "Creator licence", are the people who will stick with SD-1.5 because of the horrible licence and censorship of SD3.
And yes. if they are claiming open source, they should release the source. so that people can pick up the ashes of whatever might be left.
Also what is when I fine-tune the model? Does the license still apply? Or how?!
So silly.... Been waiting for month and.now this... :(
Also what is when I fine-tune the model? Does the license still apply? Or how?!
So silly.... Been waiting for month and.now this... :(
Apparently Yes. All Derivative (Downstream) Fine-tunes, checkpoints or LORAs. etc, all are bound by and have to use the same Licence as SD3. If you use any model that uses SD3 as a base for any "commercial" activity then you have to pay for the Creator Licence. That this Licence is being used to cover the model, derivatives of the model products of (Images Generated) by the model is insane. This feels like a death nail at a funeral. rather than the Starting Gun of a new AI model race.
I mean in our society... If you build a car... all parts are free, right? free tires, free doors, even the gas is free. So why do we have to pay for this model if we want to make money off it?
I'll just use Suno, Udio, Leonardo, Ideogram, Midjourney... Wait what? those are not trained ethically AND we have to pay?
OK, that's it. I'm back to building my own car... Wait what? Tires are not for free?
I mean in our society... If you build a car... all parts are free, right? free tires, free doors, even the gas is free. So why do we have to pay for this model if we want to make money off it?
I'll just use Suno, Udio, Leonardo, Ideogram, Midjourney... Wait what? those are not trained ethically AND we have to pay?OK, that's it. I'm back to building my own car... Wait what? Tires are not for free?
More like buying a car, paying to service that car, paying for fuel to run the car . then Toyota charging you to drive it to work and only to work.
I'm not saying completely free either as mentioned above Stability AI have dog themselves into a sink hole someone has to pay., a fairer way of doing it would be to licence it so if you are using the model as a commercial service, (charging people to use SD3) if for example you are a Suno, Udio, Leonardo, Ideogram, Midjourney etc then you have to pay to licence that. this puts the onus on those big tech companies using SD3 to pay for it. while still allowing individuals to produce images using SD3 free of charge for commercial use.
Actually that is a curious question, if somewhere like Leonardo or a Discord Bot host SD3, and you produce images with it, if they have a commercial licence, but you don't. is the licence sub-grantable or do you need your own commercial licence to use images you create using a Leonardo or Discord Bot?
- If Leonardo offers generations over SD3 then they have to pay.
- If ComfyUI or Auto1111 or other tools offer you to download and generate images then it is ok since neither comfy nor Auto1111 ever makes money from SD3 nor do they require SD3 or ship with it.
- If a tool is paid or has a paid plan but does not ship with SD3 while still offering support if the user downloads it, then again it is up to the user and not up to the tool provider.
I'm not a lawyer but it's actually not that complicated. If you or your company makes a profit from SD3 then you have to pay.
I don't really see the problem.
The only problem i see is how SAI wants to prove that a user generated over N images for a given plan at a given time. It's sort of a silly business model, but while SAI is good at providing base models, they have never been good at proving their business value, which is also clear if you look into their financial numbers or followed what happened to them in the past 6 months.
It's really best to just be fair and pay to either support SAI in what they have been doing for us or to get a commercial license for your business. If you pay 20-30$ per employee for ChatGPT then paying 20$ once for a basic commercial license shouldn't be subject for a complaint.
- If Leonardo offers generations over SD3 then they have to pay.
- If ComfyUI or Auto1111 or other tools offer you to download and generate images then it is ok since neither comfy nor Auto1111 ever makes money from SD3 nor do they require SD3 or ship with it.
- If a tool is paid or has a paid plan but does not ship with SD3 while still offering support if the user downloads it, then again it is up to the user and not up to the tool provider.
I'm not a lawyer but it's actually not that complicated. If you or your company makes a profit from SD3 then you have to pay.
I don't really see the problem.The only problem i see is how SAI wants to prove that a user generated over N images for a given plan at a given time. It's sort of a silly business model, but while SAI is good at providing base models, they have never been good at proving their business value, which is also clear if you look into their financial numbers or followed what happened to them in the past 6 months.
It's really best to just be fair and pay to either support SAI in what they have been doing for us or to get a commercial license for your business. If you pay 20-30$ per employee for ChatGPT then paying 20$ once for a basic commercial license shouldn't be subject for a complaint.
If Leonardo offers generations over SD3 then they have to pay, Yes. they have to pay. that's not the question, they have a licence to use SD3 commercial, SAI are not just claiming the use of the model here, but products OF the model. So, under that licence, Do you, as an end user of Leonardo also have to pay for a commercial licence to use the images you generate via SD3 on Leonardo?, as according to SAI's licence All Derativative products of SD3 - Models, Loras, Checkpoints, Fine Tunes, and images produced BY SD3, also require a licence.
And its NOT $20 once. its $20 Per Head. Per Month. unless they changed that to a flat fee?
i agree
update the licence
who gets licence should be able to train model and distribute model file
you can make different licence for individuals and SaaS