The Painful Launch of Stable Diffusion 3

Check out NVIDIA NIM now using this link nvda.ws/3Jn5pxb
I forgot to mention, it seems like the general consensus that people are disappointed with Stable Diffusion 3 is because there are so much potential wasted. I definitely think it's a bit unfortunate too...
Special thanks to McMonkey for fact checking this video
github.com/mcm...
[Comfy Org] www.comfy.org/
[Comfy Blog] blog.comfy.org/
more sauces:
[Initial SD3 Annoucement] stability.ai/n...
[SD3 Paper] arxiv.org/abs/...
[Robin et al news] / stability_ai_ceo_emad_...
[Emad 12 weeks tweet] x.com/EMostaqu...
[Forbes hit piece] / n_how_stability_ais_fo...
[The Register] www.theregiste...
[SD3 API] stability.ai/n...
[Creator License Agreement] stability.ai/c...
[CLA back in Dec] imgur.com/a/Lx...
[License reddit feedback] / a_plea_for_everyone_to...
[New June license] stability.ai/l...
[Olivio's video] • SD3 - Absurd new Licen...
[CivitAI temp ban] civitai.com/ar...
[CivitAI ban official] civitai.com/ar...
[Lykon's Twitter feed] x.com/lykon4072/
[McMonkey anatomy] / apparently_according_t...
[Comfy Screenshots] / the_developer_of_comfy...
[New CEO] www.reuters.co...
[New CEO announcement] stability.ai/n...
some images sources:
/ why_is_sd3_so_bad_at_g...
/ is_this_release_suppos...
x.com/moonride...
x.com/cubiq/st...
/ how_times_have_changed
/ my_son_made_a_girlfrie...
This video is supported by the kind Patrons & KZbin Members:
🙏Andrew Lescelius, alex j, Chris LeDoux, Alex Maurice, Miguilim, Deagan, FiFaŁ, Robert Zawiasa, Daddy Wen, Tony Jimenez, Panther Modern, Jake Disco, Demilson Quintao, Shuhong Chen, Hongbo Men, happi nyuu nyaa, Carol Lo, Mose Sakashita, Miguel, Bandera, Gennaro Schiano, gunwoo, Ravid Freedman, Mert Seftali, Mrityunjay, Richárd Nagyfi, Timo Steiner, Henrik G Sundt, projectAnthony, Brigham Hall, Kyle Hudson, Kalila, Jef Come, Jvari Williams, Tien Tien, BIll Mangrum, owned, Janne Kytölä, SO, Richárd Nagyfi, Hector, Drexon, Claxvii 177th, Inferencer, Michael Brenner, Akkusativ, Oleg Wock, FantomBloth
[Discord] / discord
[Twitter] / bycloudai
[Patreon] / bycloud
[Music] massobeats - daydream
[Profile & Banner Art] / pygm7
[Video Editor] Silas

[@bycloudAI]

Check out NVIDIA NIM now using this link nvda.ws/3Jn5pxb It seems like the general consensus that people are disappointed with Stable Diffusion 3 is because there are so much potential wasted. I definitely think it's a bit unfortunate too... EDIT: Update on new SD3 terms The license was changed and is now free for research and also commercial use up to $1m. - Lykon Also some elaborations by Lykon (key researcher for SD3) 1. 2B was an experiment at scaling down 8B, but it's essentially the same architecture. As Alex said on Discord, it was supposed to be released as "beta", but the label got removed last minute. That was a mistake. 2. 4B is very experimental and, as of today, has roughly the same issues as 2B. It's also using a different architecture and only one text encoder (which is probably the worst of the 3 and the heaviest). 3. 2B was released before 8B because it's easier/cheaper for us to finetune, and is also the perfect size for the community (having roughly the same resource cost of SDXL). 4. The new CEO just changed the license entirely making it free for research purpose, free for commercial use until 1M USD, free for non-commercial use. 5. We are working on a new Medium model to address the issue. 6. The issues with 2b were not intentional. The architecture was made and tested at 8B params and, turns out, MMDiT scales very well at high param count, but has attention issues on low param count. We are addressing this. 7. The 512mar SD3M that Alex showed is not the finetuned one. I just re-tested the one I have today on "that prompt" and it works much better. I'm pushing to release it (even if it's not super useful)

[@FenrirRobu]

Bycloud you really did a good job explaining the intent behind the license agreement. But that's why lawyers are so expensive - to make this stuff reasonably readable. As a developer in the AI space, I am very careful with any strange licenses. Facebook's non-commercial licenses for models are similar - they don't clearly specify what's up. (They simultaneously imply that generated works are unrestricted and that they are non-commercial to avoid music industry lawsuits) Meanwhile, if I want to create a business or a product, even an open source one I need to have the facts straight before investing resources into it. Due to the popularity I still took the risk though.

[@vladefined]

So OpenAI is actually not open and Stability AI is actually not stable...

[@cdkw2]

yes!

[@NoRightsProductions]

I mean www.dumpaday.com/wp-content/uploads/2013/09/funny-doctor-evil-9.jpg

[@edd9581]

Time to fund unstable AI and closeAI

[@jamescash4065]

This makes me so excited for Illya’s new “Safe Super Intelligence” company

[@toasteroven6761]

HMS Invincible vibes

[@cmdr_stretchedguy]

There is very little money in open source, seems some people forgot that. Having a free consumer level versus paid creator and commercial level does make sense, but most companies are not going to waste the time going through the revolving door of "call us" pricing.

[@andresreal8261]

Lol, Stable Diffusion bitching about licencing fees over technology outright born from illegal data-scraping violating a fucktillion different intellectual property rights is the funniest, most pathetic shit they've done in the last... Few days. What can I say, they got a special talent.

[@andrewostrovsky4804]

Anatomy excluded from training data for art generation model?! Exclude water from beverages next.

[@ExzaktVid]

We should exclude anything solid from foods next.

[@17th_Colossus]

Censoring AI will *always* make it worse.

[@timeTegus]

But if u watched the video u know that it has the priplems bevore savty training too

[@papakamirneron2514]

@@timeTegusyes but they were made worse by safety training, it’s a combination of problems.

[@timeTegus]

@@papakamirneron2514 no. U saw the example image that it had the same problem bevore savety.

[@mz00956]

I think the only good option is to generate it and then try to cencor whatever the ai generated after.

[@timseguine2]

@@timeTegus Removing anatomy from the training data was part of the safety strategy, and likely a big reason for the poor quality.

[@Y0UT0PIA]

Calling what is being done to models "safety tuning" is the equivalent of calling sensitivity readers "text safety consultants".

[@Happ1ness]

"You either die a hero or you live long enough to see yourself become the villain"

[@Leto2ndAtreides]

It's not villainous to try and survive after you spent millions putting out opensource stuff that other people get to build on and benefit from. The more rational position might be that they should never have tried to do it opensource and almost ruin themselves in the process. But they were optimistic... And before the current funding crunch, it might have worked out okay. But now, there's no easy money.

[@ElaraArale]

open source still the future..... but what happened to SD it's just sad....

[@float32]

I’m seeing the opposite. These models are too expensive to make. It’s clearly not sustainable. All the open source models are burning money like crazy, with no profitability in sight, with the likes of Meta doing it to keep the others from being profitable. I’m just not seeing this bright future where the gap doesn’t eventually widens further and further, as the debt collectors eventually swarm.

[@bazookaman1353]

​​@@float32The next step in AI isn't to just increase their power, but to make use cases for them. Current open source llms have the capacity to work as game characters, roguelike managers, rpg GMs, companionbots, etc. It's just that nobody is giving those things thought... yet. So to me we already won, with upgrades just being a bonus.

[@meuno101]

⁠​⁠@@bazookaman1353i was thinking about that earlier. I gave an LLM a set of lyrics and it gave me chords to go with the tone of the lyrics and they took some tuning but overall I didn't change much from the original chords. People are so hyperfocused on locking it down, censorship, 'theft,' making money fast with AI, etc. That everyday use is just thrown out a window. Shoot just recently I was looking up Japanese lyrics and I could only find the romanji so I gave it to ChatGPT and it translated it. I've not heard anyone talk of that use case. Now that everything's slowing down I do hope people are going to be looking at what to do and fine tune it from there

[@ElaraArale]

@bazookaman1353 exactly buddy, and this is exactly what I do with open sourced projects to my own project, the use cases are just great.

[@incription]

@@bazookaman1353 no they dont, even 7b arent good enough to act as game npcs convincingly, and even then most computers cant run 7b yet (especially with a game)

[@rvre]

“Safety reasons?” In a product I’m paying for? What an idiotic licensing on a bad model. Nah You’re really giving them too much credit lol

[@MrValgard]

@@rvre what I'm seeing there is some conspiracy among payment providers to exclude services 18+ and I thought USA I capitalist

[@zzzzzzz8473]

the license is disgusting your trying so hard to make it seem like its normal commercialization . the enforcement is beyond "impractical" right now , it is a poison that threatens all finetunes and merges which have been the highest quality models that people actually use ( no one uses base models ) . the open source community has been both optimizing the code to generate and train models as well as the finetuning towards quality , hoping the Open model initiative that comfy civit and laion are making will be a better representative of what that can mean for foundation model generation .

[@MrViki60]

He is a shill.

[@kgrey0582]

You can't read 😂

[@bycloudAI]

I just clarified what was being misunderstood in the community about the licensing, and other than that, I said the literally same thing as you. Did you even bother to watch the video...?

[@jameshughes3014]

making ai 'safe' is the same as making ai worthless. trying to hang on to rights too hard is as good as telling people not to use your software

[@IronFire116]

In most cases, the utility of something is proportional to its danger

[@ElaraArale]

@@IronFire116 that's actually a good quote, any reference to that? i really liked it.

[@yassinebenyamina4861]

brain dead comment

[@IronFire116]

@@ElaraArale It's my own :)

[@peromiestiloesunico]

If you are a normie reading this sabble deffusion 3 sometimes is very trigger hapoy if you want to generate an image of a girl playing soccer or training in the gym, boxing holding any type of weapon that is not fantasy cartoon or sci-fi you will get a block or get stuck on a cycle were stabble diffusion 3 will give you the same image over and over

[@TheSteveTheDragon]

Ugh. Investment groups are usually bad news for any creative IP. I can only hope with everyone leaving this will only mean new startups and competition for StabilityAI.

[@JoaoVitor-mf8iq]

No SD3 on civitai? It's okay I already have pony ( ͡° ͜ʖ ͡°)👍🍆.

[@lefourbe5596]

outside of NSFW use of pony is super bad tho. next version of pony should fix it's glaring issues. because now doing background with this dubious model is chaotic. basic older object and art style are completly absent. despite that, i like it cause it train way easier than other ...

[@mithrillis]

I never believed a second that "safety" is the main reason the model does not work well. Much more likely they just failed to train to production quality and put out a 2B model as a delay tactic. It's just "censorship" is much more convenient to rile up the online audience. Do we really have proof the model is truly "selectively" incompetent?

[@Tulsaistalking]

Think about it.. after training hoards of Human reviews doing "rlhf" take a trained model and start trying to make sure it never does things they don't like... when you consider they are just adding extra 'bias' forcing the model weights to change.. This type of finetuning "RLHF" is very similar to a digital lobotomy. They are effectively reducing the traversal space blindly. While it can be effective.. it isn't scalpel. Thus the lobotomy comparison. So yeah, the more "safe" generally the more broken

[@mithrillis]

@@Tulsaistalking but is there anything that cannot be explained by the model simply being under-trained or fried in fine-tuning? We know much of the advantages of SD3 vs older models come from the text encoder, which we can also see in other non-SD models. Apart from that, is SD3 really better than previous models EXCEPT for censored content? Or is it just not doing better across the board? I think we need some extensive testing to prove statements like "censorship caused bad model" rather than emotionally identifying with the statement...

[@sasdasu00dfsdfardo]

Image generation is nowhere near comercial usable lol, also >image generation is only usable to make porn (and this by a long strech, try selling art with deformed hands) >sd3 comes unable to make porn

[@Pawnsappsee]

Why do you guys love porn? 😂

[@marcoshalberstadt7646]

They could have competent artists doing touch ups on AI art and it is perfectly "commercial usable". The issue is that people are too obsessed with the idea of AI replacing artists entirely instead of just improving their productivity.

[@markjackson1989]

I made Dalle E 3 generate convincing images of Keanu Reeves aggressively and desperately milking cows before they censored everything.

[@Quantum_Nebula]

See i would have no problem with an ai company charging for an API for their latest and greatest model, and releasing the weights for the slightly older models with every new release.

[@Roshank231]

It really is heartbreaking though. In a perfect world, free, open source and uncensored AI access for anyone would be best. But now it feels like a dying dream. Closed source projects like midjourney are getting better, but look how far SD has fallen. Didn't want it to turn out this way, but reality is often disappointing.

[@doompoison2365]

I mean not really when it makes that off other people's stolen livelihoods.

[@hoseki4031]

​@@doompoison2365 Evolution doesn't care about livelihoods only progress.

[@clickpwn]

@@hoseki4031 I agree, and there is a reason why evolution produces only small percentage of psychopaths like you.

[@falsechord]

@@doompoison2365 eventually all jobs will be replaced with ai and we will just be free to do what we want while ai does the work for us

[@clickpwn]

@@falsechord I can’t believe there are still delusional people like this exist after the shortcomings of AI time and time again and never making actual progress toward AGI. You people must be desperate lol

[@Mythhammer]

If they had focused on making a superior product, rather than a "SAFE" product, much of this would not have happened. Once again Get Woke, go Broke.

[@Speejays2]

This is more "trying to appease investors" than "woke"

[@Mythhammer]

@@Speejays2 How do you know that the investors aren't Woke? :) Look ar Blackrock, Vanguard and State Street.

[@Speejays2]

@@Mythhammer What does woke mean to you?

[@Mythhammer]

@Speejays2 Cultural Marxism and it's fellow travelers.

[@Speejays2]

@@Mythhammer What does cultural marxism mean to you

[@differentone_p]

UNSTABLE DIFFUSION 😂😂🔥🔥

[@GhostNameless]

There's no point making AI more "safe". Pandora's box was already opened. People can create ANYTHING.

[@Milennin]

I'm feeling very safe for never wanting to use it.

[@4.0.4]

The licensing would be a very interesting conversation if the model was not lobotomized to hell and back for "safety reasons". It's trash.

[@Redranddd]

SD 1.5 + lora + controlnet is still the best

[@edd9581]

Sd xl is pretty good but they remove a lot of names due to the backlash

[@Redranddd]

@@edd9581 sdxl is better in some cases but in general is very unpredictable and more difficult to achieve masterpieces, also the there are much less LoRa's

[@ghost-user559]

SD 1.5 then run that through Sdxl, or Sdxl, then finish details in Sd1.5. Depends on what you want to do.

[@NaudVanDalen]

SD 1.5 only generates abominations for me. Even with special models.

[@ghost-user559]

@@NaudVanDalen What graphics card do you use, and what program do you use to run the models?

[@DrW1ne]

Safety. I swear I am not a kid anymore mom!

[@KEDI103]

For me SD 3 still wants to make so much NSFW but only cencorn open parts looks odd gen. But so far I did perfect female bodies very well in AUTOMATIC1111. But it fail rate crazy. Also they did wizard of coast with dnd or unity did. So gg for them if they won't step back from this. They can't be Openai or midjourney. If they try they will bankrupt in days. They are not good enought. And also they will be backstab supporters yeah.... SD only popular because its local, free, opensource and most importand can do NSFW. And now they tried to get this thing away from us so why do hell do I need to support them instead of openai or midjourney or other AI company?!? If they still do this gg for them. We saw lots of examples like this.

[@quocanhnguyen7275]

Hi thank u so much for summarizing everything. There are just overwhelming information around this

[@Froncusiek]

My guy please don't copy Fireship's thumbnail style - your content is good but I feel cheated after clicking on the video :(

[@clickpwn]

I don’t even distinguish between the two when it gets recommended anyway

[@jlljjl]

@clickpwn Fireship is to Bycloud as God is to a believer.

[@aakashchaddha6016]

So what? Don't watch it then

[@Kolesha]

​@@jlljjlYou are insane.

[@LoneWolfInsane]

Dont worry it was not stole, it was just legally created by ai

[@ashtaka]

I'm dying inside with the memes Silas put into this video

[@bycloudAI]

he's my goat

[@MrValgard]

Go woke, censor nsfw, go broke :p

[@clickpwn]

Good luck getting investments from big people with a crappy porn generator. You know it was only hype because of porn potential and those coomers don’t pay.

[@SanctuaryLife]

Maybe these companies should exclude heads from the training data, as they can be offensive and risky too.

[@OlivioSarikas]

Why are you misrepresenting what I said? - I do point out that those are 6000 commercial images. - I did point out that this is for startups, but I show in my video that they suggest that Artists use this license too. - I do specify in my video that generated images are not included - The term about the data that has to be destroyed is very vague, because it says confidential information AND ANY Derivative Works. Even ChatGPT said that this might very well mean derivative models. - The license says in clear English that you are liable for any acts or omissions by your CUSTOMER. - so if you train a model for a customer, you need to make sure they don't break any laws with it or you could be liable too. - which is probably one of the reasons Civit ai is taking a step back until things are sorted, because there is a lot of stuff on their page that goes against a lot of copyright and some other rights. I'm no lawyer either. But I do point out a lot of things you claim to correct in this video. No hard feelings. Just saying

[@bycloudAI]

Hey Olivio, I definitely should have messaged you and chat with you about this earlier, idk why that didn't cross my mind I am sorry for like what I seem to be misinterpreting or misrepresenting what you have said, I just wanted to spread some more balanced views & information on it. Since I've lurked around a lot of discord/reddit mentioning your interpretation of the license, and there seems to be quite a lot of misunderstanding coming from the people that watched your videos. I've also consulted a few people on the terms that you mentioned in the video, and do think there's a bit of a misunderstanding, at least from what we understood from your video. Sorry for any misunderstandings I've caused.

[@OlivioSarikas]

@@bycloudAI would have been fun to make the video together :)

[@maniacos9620]

The descriptive sentences are a failure to begin with already. First it requires the user to know proper English language including eloquent depiction and then it wastes a lot of tokens for fill-words like "the, is, a, an, of, to". I rather write a CSV list of things I want in the image than a novel.

[@Somebodythatoverthinks]

Stable duffusion with censorship is dogshit because dalle is better censored text to image model stable diffusion could ever be, they killed thier main selling point 😂😂😂😂😂

[@generalawareness101]

I do not like MMDIT because it is so slow to generate. Even Hunyuan LyCORIS slapped it (he is part of their dev team I believe) together with xformers and the fastest dit can be is 2.2it/s on a 4090. even if you magically doubled that due to optimizations that is still almost half of what I get from a unet/clip. Wish I could throw the mt5 onto XL and call it a day.

[@GoofyRecaps]

they updated their policy like yesterday or so

[@cubertmiso]

"OpenAI is actually not open and Stability AI is actually not stable... " LOADING.. safe superintellicence DEPLOY Y/N?

[@DefineMeAsOne]

Their terms is confusing and unless Stability themselves clarify, what you said is also moot because your source is from a biased source. I'm not saying that Olivio's video is 100% correct, but we technically just have KZbinrs and employees statements, and no clarification from the company.

[@pigeon_official]

I dont understand why theres like no other open source AI art companies there's like 10 billion LLM companies but stability AI is the only art one

[@clickpwn]

Because nobody is dumb enough to do it except Emad. It is extremely shady and he was only able to do it because it was unprecedented but it was never going to work out well anyway.

[@AllExistence]

No. First, LLMs are faster to train. Second, 98 % LLM companies use Chat GPT under the hood.

[@Arthur-jg4ji]

the other payed compagny used sd model in the dark but didn't say it ....

[@pigeon_official]

@@AllExistence I literally said OPEN SOURCE companies genius ChatGPT is not open source. also I don't care if LLMs are faster to train art models art that difficult there should be more companies

[@florntlaze810]

​@pigeon_official The early open source companies got in when scraping the internet for data and images was cheap easy. Now sites and other internet sources have anti-scraping, and the easy to get copywritten images and books (for text and sorting) are overpriced for the data they provide. In summary: The companies at the start got an easy monopoly on data, and the barrier to entry is harder for new startup and individuals who want to find data for their local usage. You'd have to be daring to become a major player these days.

[@OlivioSarikas]

I also feel like the argument that SD1.5 base model can't do a woman laying in gras either, is kind of a mute point. Since fine tuned SD1.5 models can do it very well, and stability ai had plenty of time to make and money to make their base model better. They sure lured us with perfect images in their announcement, instead of what SD3 base really does. Basically what this is saying is: Stability AI delivers a green banana, that we have to ripen - but if we do, we need to pay them to fix their model. Does that sound good to anyone?

[@Lykon]

The license was changed and is now free for research and also commercial use up to $1m.

[@cortster12]

Censorship always lobotomizes a LLM's capabilities. Always.

[@adrixshadow]

Lobotomy AI.

[@jaydeep-p]

Safety is bs

[@tioedu_]

1.5 still better

[@urgyenrigdzin3775]

well, it's always fun to fantasize about how you make a product once and whoever using it keep paying everytime thry make money out of it. Imagine if you make a screwdriver then users have to pay $0.01 for every turn you make to drive the screw (pro tier) and for every screw you drive (enterprise tier).

[@dhillaz]

I still feel the licensing terms need to be more explicit - vague licensing terms should be treated as bad licensing terms. It is not safe to assume any of our individual interpretations of the rules is correct, because when it comes down to to it, we will be laymen arguing against experienced and well funded legal teams.

[@miserablepile]

OpenAI needs as many competitors as possible

[@jonthgrutz7011]

Are you Fireship ?

[@NaudVanDalen]

Stable Diffusion is sooooooo far behind, it's not even funny. SDXL (July 26, 2023) is way worse than Midjourney v5 (March 15 2023) since the hands are still often messed up even though it came out 4 months later. Hell, I'd say it may even be worse than Midjourney v4 (November 5, 2022). Then Midjourney v6 came out in December 20, 2023 and was even better. SD3 turns out to be worse than SDXL with these monstrosities.

[@titastotas1416]

abandon clickbait, your viewerbase is not a mindless horde of dopamine junkies

[@dinogodor7210]

What's wrong? A research product is subject to product requirements. We're dealing with brain stuff here and can't have people having a say like "you mustn't use this class of image" or "you can't have this or that property on anything it creates because I consider it mine" or"I am to finicky to view the results" because science doesn't know any way to translate those requirements into actual modifications on the model. So we're denying information to a vision model and maybe filter the output. Imagine that: you are blind. technology has progressed so far that one of these models can replace your brain function or at least the processing of your eyes. The model you have to use because nothing else is available is this. Whenever you see something disturbing or erotic it garbles your sight. Very cyberpunk. With how openAI is progressing academic science might have to recreate much of their costly work now to have something workable.

[@DenisShiryaev]

Thank you for the video!

[@dot.4069]

Giggled on the "well, well, well" part Love your videos, I am too lazy to follow AI scene by myself

[@nicholash1278]

seems like AI peaked and is now getting worse. as a concept artist, this news is so awesome.

[@rhym8882]

it went from hype to complete lol meh

[@marshallodom1388]

Even if you candy coat it this is a great example of how not to train or release a model

[@pierruno]

Sounds like Fireship

[@mirek190]

When I hear "safety" I just vomit ....

[@quantuminfinity4260]

Thank you for clarifying and debunking information!

[@juanjesusligero391]

That thumbnail is... Interesting XD

[@meadbrow8479]

They should've have launched uncensored optimal 2B parameters stable diffusion with a reasonable offering of paid 4B and 8B parameters. That would fix their problem.

[@nfaza80]

The Ascendancy, Declension, and Prospective Resurgence of Stability AI: A Profound Exegesis This treatise shall elucidate the tempestuous odyssey of Stability AI, with particular emphasis on the temporal interstice encompassing the promulgation of Stable Diffusion 3. We shall expatiate upon: **1. Incipient Triumph and Hyperbolic Anticipation for Stable Diffusion 3** * **Preliminary Augury:** Stability AI tantalyzed the masses with exemplary visual specimens evincing textual coherence within generated imagery, an unprecedented feat for foundational models. They accentuated its superlative generative prowess and model magnitudes spanning from 800 million to 8 billion parameters. * **Meticulous Disquisition & Lofty Expectations:** The dissemination of an exhaustive scholarly treatise further catalyzed the fervor. It proffered profound insights into SD3's architectural intricacies, auspicious benchmarks, and captivating test images. The focal point transmuted towards the utilization of descriptive prompts in lieu of keyword-saturated methodologies. **2. The Disintegration: Pecuniary Tribulations and Internal Dissensions** * **Intimations of Fiscal Instability:** Veracious sources commenced divulgating Stability AI's pecuniary predicaments, intimating difficulties in monetizing their technological innovations and unsustainable cloud expenditures. * **Exodus of Pivotal Personnel:** Founding researchers, including seminal figures behind the Stable Diffusion technology, absconded from the company, engendering trepidation regarding its future trajectory. * **Emad Mostaque's Abdication:** In short order, CEO and founder Emad Mostaque transitioned to an investor role, further fomenting speculation about internal tumult. * **Forbes Exposé and Emad's Rejoinder:** A scathing Forbes article delineated a tableau of mismanagement, unattained objectives, and strained industry relations. While Emad repudiated certain allegations, he acknowledged challenges in monetization and navigating industry dynamics. * **Liquidity Crises & Organizational Metamorphosis:** Reports surfaced regarding Stability AI's incapacity to defray GPU rental obligations due to prodigious cloud expenditures. Retrenchments and restructuring initiatives ensued as the COO and CTO assumed interim CEO roles. **3. The Contentious Unveiling of Stable Diffusion 3** * **Protracted Release & Communal Apprehension:** Notwithstanding assurances of expeditious dissemination, the SD3 weights were delayed, engendering consternation that they might never materialize. * **Stable Diffusion 3 Medium Release and Licensing Polemic:** * The promulgation of SD3 Medium, a 2B parameter model, was met with disillusionment. Its generative quality, particularly concerning human anatomy, fell short of expectations. * The novel licensing paradigm, while ostensibly conventional, ignited significant controversy due to its commercial use restrictions and perceived limitations on individual creators. * Misinterpretations of the licensing stipulations, particularly regarding the monthly generation limit and derivative works, exacerbated discontent within the community. **4. Salient Issues with Stable Diffusion 3 and its Licensing:** * **Technical Deficiencies:** * SD3 Medium exhibited egregious shortcomings in image quality, particularly in generating human anatomy, precipitating widespread censure. * Hypotheses for the subpar performance encompassed intentional sabotage, training-inference disparity, data anomalies, and overzealous safety tuning. * Evidence intimates a confluence of factors, including a flawed base model and potential over-reliance on safety measures. * **Licensing Trepidations:** * The novel licensing paradigm, mandating paid subscriptions for commercial use and imposing constraints on monthly generations, conflicted with the established open-source ethos of the Stable Diffusion community. * Misapprehensions and ambiguity regarding specific terms, such as "derivative work," fomented anxieties among individual creators and commercial users alike. * The licensing restrictions, particularly the generation limit, appeared impracticable for platforms like CivitAI, which heavily rely on model sharing and monetization by individual creators. **5. Internal Discord & Communal Ramifications** * **Confy Anonymous's Exodus:** Confy Anonymous, progenitor of the popular ComfyUI and a close collaborator, tendered his resignation from Stability AI, citing concerns about the company's priorities and decision-making processes. * **SD3 Training Revelations:** Leaked Discord screenshots unveiled internal conflicts regarding SD3's development, with allegations of a flawed model selection process and prioritization of paid API over open-source quality. * **CivitAI Proscription and Stagnation Concerns:** CivitAI, a crucial platform for sharing and monetizing Stable Diffusion models, proscribed SD3 models due to licensing conflicts, effectively impeding the model's growth and development within the community. **6. The Future of Stability AI: Uncertainty and a Glimmer of Hope** * **New Leadership & Investment:** Stability AI appointed a new CEO and secured funding from an investor group, signaling a potential inflection point. * **Necessity for Internal Reform & Community Reconstruction:** The company faces the Herculean task of regaining community trust, addressing internal conflicts, and clarifying its commitment to open-source development. * **Focus on Innovation and Collaboration:** The success of Stable Diffusion hinges on continuous innovation and active collaboration with the open-source community. **7. Denouement:** The saga of Stability AI serves as a cautionary exemplar of rapid ascendancy, internal strife, and the complexities of balancing open-source ideals with commercial interests. While the future remains shrouded in uncertainty, the company possesses an opportunity to glean wisdom from its missteps, prioritize community engagement, and leverage its technological potential for positive impact.

[@Leto2ndAtreides]

People are being rather unreasonable expecting Emad etc. to end up poor for the sake of putting out opensource stuff. They probably should be more creative on the money making front though.

[@Lykon]

Nice video. I'll give you some of my perspectives: 1. 2B was an experiment at scaling down 8B, but it's essentially the same architecture. As Alex said on Discord, it was supposed to be released as "beta", but the label got removed last minute. That was a mistake. 2. 4B is very experimental and, as of today, has roughly the same issues as 2B. It's also using a different architecture and only one text encoder (which is probably the worst of the 3 and the heaviest). 3. 2B was released before 8B because it's easier/cheaper for us to finetune, and is also the perfect size for the community (having roughly the same resource cost of SDXL). 4. The new CEO just changed the license entirely making it free for research purpose, free for commercial use until 1M USD, free for non-commercial use. 5. We are working on a new Medium model to address the issue. 6. The issues with 2b were not intentional. The architecture was made and tested at 8B params and, turns out, MMDiT scales very well at high param count, but has attention issues on low param count. We are addressing this. 7. The 512mar SD3M that Alex showed is not the finetuned one. I just re-tested the one I have today on "that prompt" and it works much better. I'm pushing to release it (even if it's not super useful)

[@AB-wf8ek]

Really great summary. On day one, I remember reading a comment that AI tools lower the bar for entry, but the ceiling is just as high. Unfortunately, due to the free aspect of open source, it's attracted a lot of people who can't see beyond the tips of their noses. Does anybody complaining about "safety" follow the lawsuits currently in court against Stability and other AI companies? These are really big financial liabilities, of course they're going to try and shield themselves as much as possible.

[@riskyanalysis5479]

Ambiguity in any contract is and should always be read in a way that is detrimental to the the signee, user or non-corporate entity. This is why you need to hire a lawyer to read over them.

[@rvre]

Lol stability ai is a disappointment yet again. I don’t agree with your breakdown is exactly what they say is what you. Exactly, the ban is good and dumb licensing. Greed

[@BlackDragonBE]

If Olivios could read, he would be very upset.

[@NostraDavid2]

He actually posted in the comments >_>

[@Beryesa.]

My inner open source dev and artist clashes really bad here on feelings "😢","🎉" xD

[@Definesleeper]

i really hope the comfy team becomes something significant , its always been a sick UI to work with

[@freedomtownn]

This is just sad. :(

[@Warrrrrbbble]

"So I asked SAIs biggest shill to fact check what I'm about to say"

[@progamer1196]

Far from the truth

[@justahumanwithamask4089]

Instead of letting other people charge their users to generate images, open a service and do it themselves

[@happyjohn1656]

4:49 No way that's the Sentinel guy!! 11:13 PM 6/29/2024

[@dohminkonoha3200]

It’s perfect tool to make idea of enemies of cosmic horror game.

[@diamonx.661]

Thank god, I thought I was just using it wrong the whole time!

[@wikwayer]

Security vs functionalities the rest is history.

[@taco7043]

how are these business models supposed to work

[@chineduachimalo391]

This why we can't have good things

[@johelsen5776]

Spellcheck is a thing.. Has been for over 40 years.

[@NostraDavid2]

Make it 50 - Unix had spellcheck based on a dictionary.

[@thuonglongtrananh8509]

1:27 I love this community

[@swannschilling474]

Thanks a lot for this one! 😊

[@ExtraDor]

Rip stable diffusion

[@lamardoss]

thank you. this helped.

[@ItsTheWhale]

RIP StableLM

[@hakankosebas2085]

what is summary

[@JulianHarris]

What’s an “unconsistent goal”?

[@adampenbrook5751]

A goal that keeps changing, I assume.

[@hooster21]

new video LETSGOOOO

[@cdkw2]

📠

[@timtarbet4594]

There are two things I don't understand here. 1. Why people think they're entitled to the use of these models. I've seen a couple of people in this space (like Olivio Sarkas) who're up in arms about having to pay for a service simply because they got it for free, not understanding the time and money that went into training these models. These companies have to remain solvent guys. This stuff doesn't just happen out of thin air. 2. Why StabilityAI thought they could get away with such a weird and draconian contract. Again, I'm not opposed to paying a fee to these guys, but thinking that they "own" all models that are trained from their models is a little bit ridiculous.

[@joshuablaz]

It's because Stable Diffusion has built their brand and attracted users with the FOSS or Libre mindset. Lots of people who are more casually interested in generative AI use pre-packgaged and convenient services like Midjourney. SD has always been the best option for those who want to download their models, to run them locally without contacting a server, to be able to tinker tune and mix the things that THEY want the model to be. If GIMP were no longer free and saved all your pictures to a cloud that could be hacked or tampered with, who would stay with it? If Arch Linux went closed source and required a constant internet connection, it's usage would absolutely tank, instantly. I think it's mostly about values, ultimately.

[@Clementine_Serpent]

Many people who use AI: Artists deserve to starve, they ask too much for some stupid pictures Same people, when companies want to monetize a product that eats a lot of investments and requires a lot of work: >:0

[@TragicGFuel]

@@Clementine_Serpent why you here lmao

[@clovernacknime6984]

@@Clementine_Serpent Artists don't deserve to starve any more than anyone else who've lost their jobs to technology. But also no less. That's something many of them seem to struggle accepting.

[@ghost-user559]

Because if a company is going to harvest all our data, without compensation, then they should freely release the results, without compensation. It’s very simple really.

[@wsg1231]

brah yeah I feel bad for one's project not being able to make money but whom project built upon stolen images shouldn't complain or ask for money AI Image generation shouldn't be able to be sold or make profit for both parties If it from the beginning was ethically built wouldn't be a problem but it's now too late to stop the spread everyone can download one locally

[@elliotalderson7823]

That just shows CivitAI having too much power. We need a crowd-sourced open nonprofit platform. The open-source gets hindered because of CivitAI's monetary interest.

[@knoopx]

nah the culpit is SAI not CAI. there's also huggingface as alternative but the diffusion community completely ignores it.

[@ghost-user559]

CivitAi is about to make the next big open source model. They are planning on having the entire community contribute to a brand new base model. So they might be the only legitimate open source future of image generation.

[@大支爺]

The SD3 is still based on 2D which it hasn't skeleton system to generate any objects.

[@telotawa]

ai bros when they're expected to actually read more than 10 words: wtf!!! 6000 images??!!?!?!?!!

[@Clementine_Serpent]

I think FBI should check hard drives of everyone here who's crying about censorship... 👀

[@Pawnsappsee]

Someone generating kids images 😢

[@aaagaming2023]

If it cant generate women lying on the grass because of censorship thats a valid creative concern, nothing to do with kiddie fiddlers.

[@Clementine_Serpent]

​@@aaagaming2023if you read the comments, most dislike the inability to make nsfw content with the models, not the specific issue of "women lying on grass". They directly say safety is overrated, not that it is a poor excuse for poor performance.

[@aaagaming2023]

@@Clementine_Serpent Youre clearly out of the loop. Imagine thinking the issue the vast majority have with SAI's handling of 'safety' is not being able to make illegal porn. If you had a clue, youd know its because they literally gutted the model from being able to be used for human anatomy in the process.

[@Clementine_Serpent]

@aaagaming2023 I have an issue with the way some people worded their comments. And thanks for recaping the video for me, it's not like I watched it before going to read the comment section. :^)

[@JaXuun]

Plumbus? @ tbumbnail