Stable Diffusion 3 is HERE! MASSIVE Improvements, Turbo, 3D, Can Stability AI Survive?
Vložit
- čas přidán 16. 04. 2024
- It's here! Stability AI's powerhouse image generator, Stable Diffusion 3, is finally out. This new model boasts stunning photorealism, crushes prompts, and even offers different sizes to fit your needs. But is it worth the extra juice? With Emad stepping down and unpaid AWS bills looming, Stability AI might need a hit. Dive in and see if Stable Diffusion 3 is the gamechanger we've all been waiting for!
Let us know what you think in the comments below!
Twitter announcement: / 1780599024707596508
Stability blog: stability.ai/news/stable-diff... - Věda a technologie
"In keeping our commitment to open generative AI, we're closing our next generative AI"
Wonder if it's because Emad left
Yeah, that was pretty humorous (and a bit sad). Reminds me of them describing how OpenAI is now closed source.
well, open isn't useful if they are no longer around. Hopefully they'll survive
@@NowayJose14he left cause of that
is this some kind of orwe!!ian new speak?
Sometimes I let my PC just sit there and render 1000 SDXL images on a day just for fun. That would be $65 with stable diffusion 3.
So, it’s a no go.
As soon as the weights are released it won't matter who would ever pay for API call if they can use a fine tuned model instead. They will try to milk this release for as long as they can because once you release the weights you eventually lose control of it because it hard to prove where AI image came from. Guess all critics were right GPU cluster cost to much to subsidize at the moment. Maybe by June we will see the first fine tuned model.
GPUs are expensive - but at least the debt they're in to AWS and Coreweave gave us SD3 haha
@aifluxchannel you can use commodity gpu for this... thats the definition of not expensive.
I want the weights locally so I can finetune it. It's not about the cost, it's that requiring payment takes away peoples' freedom to use it.
I agree - the principle of "paid" open source just doesn't make sense. What kind of finetunes do you plan on using SD3 for?
@@aifluxchannel "Paid open source" only makes sense in a donation context. I mean, I could see this working out where you can get the models early for a price and then they release them for free later. That would also be fine. Let's get real, if Stability goes through with this, then people will start sharing the models on Civit, HuggingFace, or via torrents.
To me it can be paid, and still be open to the general public, as long as the price is reasonable, as in doesn't price most people out.
@@armadasinterceptor2955 That's a fair point to be honest. I suppose people shouldn't be completely unreasonable and expect free lunch. 🤔
@@TheSniperWho It would be fair if we could expect a fair price ;) but...
I hope it is released soon. Other image generation ai's are getting really good and Sora is looming in the background...
Soon TM
This is what Emad said on reddit:
"To clarify weights will be made available soon (always API first, then a few weeks later weights).
They will be downloadable on hugging face for anyone.
To use them you need a membership which is free for personal and non-commercial, costs a bit for commercial use."
From this, I believe even if you use your own GPU, you will need to pay them if you want to use the images you generate commercially. For up to $1m you can use professional license ($20 USD/ month) and for more than $1m you will need the Enterprise license. That's my interpretation and I see no official announcement clarifying things but it appears this will be a "paid" open source model.
The question is, the previous models were good but have question marks on copyright usage, does this new model mean they have licenses for all the training data now?
I agree the most interesting details they left out is the "enterprise" pricing or what they mean by "enterprise' use
Sounds to me like an infrastructure change. Like people will be able to use their own GPU but the program that they utilize to create the images will require a membership. In that case we might not be paying for the weights themselves. This could turn into an Adobe Photoshop kind of situation... Well regardless, there will still be "piracy."
@@TheSniperWho I think it will be like Unity and Unreal Game Engine model. Well, closer to old Unity where you get the whole product (weights) but if you start making money, you have to pay.
And most people who think, "I'm not gonna pay", are not the target audience as they will likely not generate significant revenues. The moment you make decent money, it will show in tax returns and company accounts and by the time you get that big you will likely have lawyer/ accountants who will pretty much tell you it will be cheaper in the long run to pay rather than amass large legal bills.
The big issue is the legality of training data and whether Stability have addressed this, because don't forget, the output of the AI images can't be copyrighted. So most people who use it commercially will use it for non IP stuff like backgrounds, ads, client mock ups etc.
This also means for people thinking, I'll just use the free SD 1.5 might not realise its very possible that you won't be able to use it commercially ever if Stability and co lose their court case. This is why stores like Steam have limited games which use AI because they want to be covered legally - hence the need for Stability and similar tools to address this which I believe will be by getting the copyrights to the training data - Adobe Firefly I believe trained on their proprietary data set.
@@phoenixfire6559 Very interesting! Honestly, this is a great take. I've been conceptually considering exactly this case. You vividly described quite a compelling point. If you'll allow me to steer in another direction... From what I gather, there's a solution to the legal problems SD users might face. Consider someone using Bitcoin and the lightning network over the traditional financial system to avoid these kinds of legal problems. See, the question of the commercial viability of this situation lies in the adoption of Bitcoin and decentralized social media. It's a part of the open discussion we're having culturally regarding AI art. If people find AI art valuable they might have to adopt Bitcoin in order to financially support SD users and artists. There are open social media networks like NOSTR that allow people to post and share AI generated content. NOSTR clients also have TOR and lightning compatibility for payments. Of course none of this particularly matters for big companies like Valve but for others it might be a solution.
Politically speaking, I'm quite in favor of people using AI to generate art for profit. I find the concept of Bitcoin and decentralized social media very compelling for SD users. Of course it's very stiff competition in this kind of market environment. You need to really create something that requires a fair amount of human input that also looks coherent. If the AI art doesn't look real, coherent, or "professional" then there's no reason for someone to tip or pay someone with Bitcoin. And believe me, people who would engage in this stuff know what good and bad AI art looks like.
Now this might be a little insensitive of me but to me it seems like those who are against AI art tend to prefer centralized or government based forms of power as a way of stopping the proliferation of AI art for profit. So basically if you want to support AI artists you might have to adopt these technologies. Which might inherently mean less revenue for AI artists due to the scrutiny of the market and the necessity to adopt technology that is arguably still niche. This certainly excludes you from the traditional financial system in some sense.
It's wild that something that appears to be so benign like AI art is being pushed towards an alternative financial system. In the past we used to think all that Bitcoin was ever good for was the sale of illicit goods like recreational drugs. But now people are using it for AI art. Basically pixels on a screen that were generated using a computer. Which sure, that's perhaps over simplified and reductionist to say that it's "just pixels on a screen." There could be some pretty messed up stuff on screen. Regardless, it's a strictly visual thing over something that really could ruin someone's life like in the case of "drugs." Which then again, it could be used for blackmail or to lie about someone. So perhaps I'm wrong about it being benign. We're at that point where people use it for those reasons but I don't think it's being done on a mass scale and being done by a significant amount of "common people." So eh, I guess I'll backtrack again and say that I'm a little bit right and a little bit wrong in regards to that...
Anyways, thoughts?
If stability AI wants to make money, they should do a crowdfunding random for the weights.
There's no way to have people be able to run it locally while also charging them. Once it's out it's out.
they could have created a civitai style website and made money from memberships
@@asmrgamingOz sheesh u a genius
Crowdfunding isnt big enough.
They should release it open source free. But charge any companies that want to host models. Users can take it for free but businesses that want to charge memberships have to pay a fee. That would bring in millions. And keep the ecosystem free and open.
Stabilityai had a good run, but it seems that after Emad stepped down, they're gonna be charging a subscription for this.
Damn emails, why did email have to go away,
that damn email moustache
Emailing is always stepping down, when are people going to learn?
It's always email stepping down, when will people ever learn!
@@ryzikxI was drinking raspberry tea and f’n sprayed it everywhere when I read your comment holy $h1t lmao. 🤣 💀
Who needs safety?😅
It always annoys me they still call themselves Open AI.
they dont
Thanks for your timely update on this.
Any time
A bit strange that there are no samples with hands or human human or human tool interactions, it can text ok, nice-that about a ball in a cup on a tv that shows friends? How consistent follows it, or better, understand it prompts? Grok should be good on that type off understanding, understanding physical function and space should help with hands to..but not local is a killer.
When are they releasing weights? No way I'd pay $20/mo for an "open source" gen ai model 😂
We'll have to wait and see. I also agree their approach to this new way of monetizing open source models seems half-baked.
@@aifluxchannelI'm not a fan but we'll have to wait and see if I can really run it on a 4080
I gotta disagree a bit. It'll still find its way into the public, people will quickly share around the weights, but I'd rather have sd be kept alive and be paid then die all together
@@betterthantrash111yeah - I also think the weights will end up on HF soon enough. Even if a few people have to pay to get the base model 😂
maybe its cuz emad left?
Community driven fine tuned models I can run offline and local mean everything to me. I doubt I'll ever give that up even for a better version if SD3 remains online.
2:37 Sooooo are they fixing hand/feet (small detail) generation? LOL. The text thing is cool though.
I wonder if with all they have done, the name they made themselves and the importance they had for open source weights, if some crowd funding strategy could help them out. It'll be a major loss if they go as open as Open AI...
No one can go as open as OpenAI. That would be Universe destroying openness.
Where can you use the API currently?
If you can't fine-tune it, it probably won't survive the other state of the art models. It's not OpenAI if it's not open. It's fine if they offer an API service and charge for the service itself to run it but this seems way above that, as if they're charging for the model.
Open source era for SD has come to an end. Now they'll charge money for it.
Wait I will be using my own GPU and still pay for use??
That's what it looks like unless they change course :(
I mean, that's how it works with any software. But I get what you're saying, I'm disappointed too
@@isatche No, that's how it works with any software *with a subscription model.* Plenty of software is either open source, closed source but free, or pay-once-use-forever. The idea of having to keep paying to keep using a piece of software on my own machine, and to use the images I generate myself freely, is the greediest of the possible business models. (I guess there's one level of greed higher, if they didn't offer a free tier at all, but that's a small difference and it would of course lead to them having basically zero customers.)
That model won't be working without an internet connection?
I remember their pinkfloyd model was using your local GPU but wasn't working without an internet connection to check your account.
@IceMetalPunk I'm afraid there is some misunderstanding here. Adobe has a subscription model and uses your own hardware. So is Maxon software like Cinema 4D and Zbrush. So is any other locally installed software with a subscription model. There are online ones, but not every sub based software is online.
Having said that, I disagree with their direction. I prefer a true open source local model I can use and not worry about subscriptions, plans, usage rights or any other restriction.
Aside from the cost, what really worries me is the level of freedom we lost without the weight. This would inevitably slow down the development of gen ai.
will stable diffusion 3 be available to download locally (safetensors file)?
The real problem I find with this new SD models is that they don't share the 2 main features of SD1.5 (And that's the reason I think SD3 will also flop), SD1.5 main features are:
1. Is big enough to do essentially anything.
2. Is small enough to do essentially anything.
I mean you may argue that SDXL is better than SD1.5, but then you go to civitai (and similar) and there are not that many fine-tunes, loras, etc. since training them is more expensive and the final result is not that better.
There is an entire ecosystem around expanding SD to do other things or certain things better. SD3 taking that away is shooting itself in the foot imo.
I'd be more willing to support a model where they charge for commercialization of their product.
Lots of services basically sell access to SD generation and that's where it makes sense for them to tuck some royalty fees in.
That keeps it more affordable for me if I just want to roll my own server and generate content that I have commercial rights to. And still a lot more people are willing to pay for a service that is priced competitively. If SD throws up a firewall allowing only themselves to sell/generate the result content, well people are just going to move to other models.
I get pretty annoyed when people say that "oh the results of the new base model are not that much better than the old one." People say that due to a reverse recency bias and it's very annoying considering it's plainly obvious that new official base models from Stability have had MASSIVE visual improvements in almost every case. SD 1.5 was way better than 1.4. SDXL was also a massive improvement from 1.5 and yet people were still saying that 1.5 base model outputs looked better when that was clearly not the case. Even worse if they said that SDXL community models would be negligible improvements on 1.5 community models. People only say these things because they are afraid of being left behind due to a possible increase in VRAM requirements. Meaning that you might need to pay for a better GPU, computer, or pay some membership to be able to generate images on someone's server. This has hardly been the case before. There have always been optimizations for SD and SD3 will be no different. Perhaps the only difference this time around is that SD3 will be able to scale such that higher end GPU's might be required to use a heftier model which could lead to better results. Even in that department we might still see optimizations for less demanding variations of the SD3 base model. And please don't give me the subjectivity argument. It never works out in practice, especially relative to time and adoption which is the only thing that matters in the subjectivity argument. There will be a majority that switches to SD3 because it's clearly better than SDXL. We saw how people adopted SDXL and we will see it again. This is also certainly not an SD 2.0 situation.
7:02 Simply they inserted the payment of the training model. In inference, it's faster cheaper safer, but the model training was slow and costly.
you didnt even test the model? GG bro was a good run
Honestly their models weren't that impressive ,it's what community did with them what is impressive. But if you can't dl free fine tuned models from civitai and everyone gonna start asking money for their models they won't last long. I doubt many ppls wanna pay for average censored image generation. Being opensourced is what separated them from rest.
I will become a member to support Stability AI. We need this more than ever. Also, I'm going to be honest, uncensored is the ONLY way to go. We can filter it for kids or whatever as we do anything else. But generating an AI character from scratch and doing things like reproducing Roman marbles, or life drawing, is necessary for this to be considered a professional tool. Yeah civit is full of teenage hentai and social-media-bot-girlfriends are the scare of the day.... we can do this better but it needs support.
Community involvement will drop like a brick.
It's certainly an odd move by the new leaders of their org.
SD1.5 is good enough for my purposes and runs on a 12gb 3060. Hell, SDXL runs local if you want to wait a minute. That said, I'm sure there are people out there that either don't have the hardware or don't know how to install something like easydiffusion and are willing to pay for the latest and greatest, like maybe marketing teams that want to automate social media presence or something.
Forge and comfyUI is the answer. If you know how to edit commands for GPU,even it's just 10gb vram.
The problem with all the SD models so far is the legality of the training data which is going through the courts right now. Even if they win, new legislation could be introduced requiring appropriate licenses for the training data. Why does this matter? If you can't use SD commercially, it will only be used by hobbyists which means little to no income which will eventually mean the model is no longer updated.
@@phoenixfire6559 Don't limit yourself, life has no SOP. Even creativity. You can reproduce from what SD has now. Train your own models. Generate beautiful artwork, tease your audience want more on the masterpieces you created,make them forget to question you whether it created by AI or not. AI art won't reveal itself where there were came from, unless you reveal it yourself to your audience.
as long as they don't release the weight i don't care lol.
Take a drink every time he says interesting
hahaha
@@aifluxchannel Glad you take it light hearted, I find it really hard to speak without a lot of ahs & ums myself
Appreciate you taking the time to cover the news
The weights are gonna be released in 3 - 6 weeks, don't panic my guys, it's always been this way
If I make the image on a free individual account, there should be some way to later transfer the image to commercial use - perhaps buying a license afterwards? The generated image can hardly be reproduced exactly as it is if you acquire a license and try to make the image again.
If you are talking about using it locally with your own GPU then if you use same seed, model and exact image settings (prompt etc) you should output the exact same image.
Nope ... even then won't be the same ... must be also the same schedular and implementation of your work flow. @@phoenixfire6559
The membership thing is not new. Its been on their site for months.
You're citing a tweet from almost a year ago that SAI can't pay their bills?
I'd gladly pay. What other company in this space is even interested in you having access to the weights?
I mean... every previous version of Stable Diffusion? The guy who created ControlNet?
open source ≠ free source
Crowd funding seems better.
I agree! I'm shocked this isn't a feature offered by Hugging Face yet - dare I say something we could crowd-fund with crypto haha.
Please correct your video. The model weights are going to be available in ~3-6 weeks, I have confirmation from Emad himself. The API access has always come first, then self-hosting via membership. It's just that we haven't had major SD releases in quite a while. The API, the membership self-hosting, and the open release are all separate. THIS IS HOW THEY HAVE ALWAYS DONE IT!!
think the math comes out to $.25 an image on turbo? video game arcade style pricing. I'll wait for the weights.
I tried SD3 and created some images. I didn't liked it at all. They say it competes with Midjourney and Dall E 3 but it is no where near.
So SD is no more free anymore .. interesting
well, they need to make money somehow, otherwise we won't be able to see a potential Stable Diffusion 4 in the future
It was fun, while it lasted.
Their professional membership offer for developers is dog sh*t. It states: "For creators and developers with less than $1M in annual revenue, $1M in institutional funding, and 1M monthly active users (all three must apply)". So basically you need 1 million active users and $1 million in funding to commercially use stability's models in your apps. No wonder they're struggling to make a nickle.
Hmmm.. better than midjourney v6??
not even close
@@zSion Depends on the usecases
SD will always live or die based on its finetunes and Lora. We'll have to wait and see if/when it's properly released to see how it responds to training. Them dragging their asses to get the model out so they can sell API access should give everyone pause as to whether we'll see a proper SD3 release.
I would argue Midjourney v6 and SD3 are very different tools - one offers far more control / tuning than the other. But Midjourney will always be a standard.
The sad part about Opersource AI. indeed. and AMD and Intel still behind aswell. xd
The bleak future is that they will make their stuff more like MJ, if they can't make a decent money out of it. and we can't hold the model whatsoever.
when can we hope to see it on Automatic1111 and Comfy, then?
Can Stable Diffusion exist without the company?
The people write the code on their own and make branches
Make Stable Diffusion like Blender, and any company should be Non-profit
Boo!
Its nothing new this time, If you wanted to use any of the prior models for commercial use, you had to buy a membership, as far i know, its just that no one paid attention to it before
People thinking they can just get weights and stuff for free for ever 😂😂😂😂
Hopefully they go broke!
Can you please give context and explain why you keep saying “interesting”, “curious”, and “wild”. You aren’t explaining why you think that!!!
Subscriptions models have thoroughly tainted the internet. If you wanna sell a product, sell a product. If you want to make more money, make more products, sell Addons, specializations, license commercial use, and so on. This "lets turn all buyers into renters" crap has been played out since photoshop became a subscription.
Not really a better product than what exists already from the images they posted. There's gotta be a better way than these garbage subscription models.
I agree - interesting time to be reaching performance saturation and Stability is clearly desperate to find "revenue" anywhere possible.
Bad hands fixed? bad body fixed? bad face fixed???????????????????????
Which is most important for you?
Hands not fixed completely from the preview images but it has way better context and good text. Until we see a finetuned model I would reserve judgement base SDXL isn't that great but finetuned models haven't gotten really good.
there are good loras/embeddings that fix these already
Bad-hands v3, badhands v4 😂😂
bye bye open source as always, you suc*
A guy on Reddit has open access to SD3 on the Stable Diffusion subreddit and asked people for prompts to try and so far the results he’s posting are not even as good as MJv6 (I’ve used all the T2i models) or even Ideogram or fine tuned SDXL models. The prompt adherence is literal A$$, it still can’t do objects and interaction well with other people or objects and stovetops still look like lovecraftian art etc. and they have the audacity to try and *charge* people?! It still gets characters wrong too Lmao This is gonna flop harder than a woke Hollywood movie. Not the least excited at all. It’s way overhyped and for not much “improvement” at all to anyone who actually uses all the different generative platforms.
With such a positive video like this, you make my day.
We already have Midjourney and DAL-E. A closed SD3 is no different than those models.
Very very disappointed by the pricing. Thought this was going to make my project a lot more viable than it would be with MJ or Dall-E but looks like I'll have to go with Ideogram
It has been implemented into ComfyUI already, sample images are here
czcams.com/video/3tI6eg4pWiU/video.htmlsi=duGpLZrePyb9f_6v
Been on 1.5 since its release. Don't want cloud versions. That old one has the best ecosystem with auto1111. Batch processing, controlnet, in and outpainting, custom loras, custom everything and running on just about anything as old as a 980. You can get a dumpster dive computer and do AI on it, and with a modern GPU you'll be flying. Can tweak everything. Cloud subscription? Screw that! They're the reason I quit Dall-E and Midjourney. I don't think I'll be moving on from 1.5. If you know how to work it, it can already do everything, short of writing text on things, but who cares about that. It's a miniscule aspect. Can't compete with "free". I didn't even like 2.0. Scrubbed models aren't worth it. 1.5 was as good as it ever got.
making money by not open source your models sounds like a good idea until your fired from the company!