Stable Diffusion 3 Announced! How can you get it?
HTML-код
- Опубликовано: 3 июн 2024
- Stable Diffusion 3 announced by Stability AI. stability.ai/news/stable-diff...
Sign up for SD3 here stability.ai/stablediffusion3
Get early access to videos and help me, support me on Patreon / sebastiankamph
Chat with me in our community discord: / discord
Stable Diffusion for Beginners Playlist • Stable Diffusion Begin...
My Weekly AI Art Challenges • Let's AI Paint - Weekl...
My Stable diffusion workflow to Perfect Images • Revealing my Workflow ...
ControlNet tutorial and install guide • NEW ControlNet for Sta...
Famous Scenes Remade by ControlNet AI • Famous Scenes Remade b... 
Хобби
TL:DW you can't get it right now, you can sign up for the beta and hope to get access, but for now everyone has to wait until they release it.
Saved me 7 minutes. When will they release it?
Appreciate it.
@@trebory6 Still unknown :')
Of course SD3's #1 advantage is: Open Source. We all take a quality hit to use SDXL instead of M and D3, but it's worth it b/c of the speed, infinite free gens, controlnet, ipadapter, img2img, inpaint, loras, animatediff, custom research like CADS and AgentAttention; the list never ends... A more powerful Stable Diffusion model is not directly comparable to models like M and D3, because SD is fundamentally more useful.
(And thank you for the video! Hope to get access soon. :-) )
If you have to sign up for it, is it really still "open source"?
If it’s gonna be censored which it may be then you may as well just use closed source generators
@@jimmy_kirk Yes, even if they're asking you to sign up for an early-access waiting list, SD3 is most definitely open source. When Stability releases the weights publicly (and they will), people will start building on SD3 and changing it, without checking with Stability. It will be in the wild. That's what makes an AI model open source.
@@ponternal It doesn't matter how "censored" SD3 is, as long as it's good.
Look at SD2: that was definitely worse than 1.5. Stability made SD2 specifically in order to censor it, with no improvements over SD1.5. No one cared, no one used it, and that will never change.
Then, they released SDXL. It was still very censored, a little less than SD2, but it was just a better model. *Way* better. Everyone cared about it being better, so it was worth training. People trained it with whatever content they thought was missing (mostly artists left out of the training data), and now you can use "uncensored" SDXL. That's how open source works. It makes censorship impossible.
@@jonmichaelgalindo Are you sure that they're not intending to release it under some other licence? Perhaps a CreativeML OpenRAIL++ licence, as was done with SD1 and SD2?
Having the source code available doesn't inherently mean that it has been released under an open source licence. There are very specific restrictions placed on SD1 and SD2. You can't just take the source code and do what you want with it under the CMLOR licence.
Awesome to hear, it feels like yesterday when SDXL launched and now SD3? Nice :)
I always look forward to your new videos Seb! I have very limited time and you help me catch up to new AI trends.
Thank you sir!
I'm happy to hear my content is valuable to you, Bob! 😊
Speaking of the wind :
One time, I met a "wind" that was very "wound" up in its scientific research.
I asked what it was learning about.
The answer blew me away
This joke blows
Problem is it will be censored so you might as well just use the popular image generators already available
This. If they make it so you can’t do spicy images only the artsy types will want to use it.
Even if it's fully censored there's always workarounds which you can use or wait until someone produces an uncensored model.
It's gonna be censored at least a bit, but it's for their own sakes. You can't think of being sued for every open source (free for everyone) product you release. The users are very likely in charge of fine-tuning it enough to make it fully uncensored
So is this going to be available to install locally through github like the other SD and SDXL releases?
can't wait to try
Are models of SD 1.5 compatible with this version? Or it will need totally new models? Sorry if newb question
Your question doesnt make sense. This IS a new model.
@Michael_Lochlann thanks!
Is it really accurate with text though ? We had the same claims when SDXL came out and it wasn't that great in the end, despite the improvement from SD1.5
Yeah I feel the same way.
The best thing about all these new products and updates are the competition they create between all the awesome companies and communities. Can't wait for SD3.
For sure! Faster improvements for us :)
problem with stable diffusion is commercial usage require 20$ membership ? I not getting whether i can sell AI arts on Etsy or I need to pay 20$/month to stability AI to use it ?
Nice work Sebastian!
Thank you kindly, not only for the comment but also for your support. Greatly appreciated :) 😊🌟
Thanks for covering. Text and prompt adherence "seems" really good but the image composition, detail and lighting are still behind other models. I like Dall-E outputs but generally speaking, Mid-J only seems to have only one artistic style. Variety is something SD excels in. I use different models for specific aesthetics. I really appreciate that about SD even if it takes more work.
is it avails yet?
I suspect that the sudden focus of multiple models (SD3, MJ, DALLE3) on text spelling is not out of high demand for the feature, but simply because it is easy to train. You can easily make synthetic examples, or add automatically recognised texts to your image annotations. But I wonder how many people really want this (as a priority feature) other than making memes? If you really want some good text, you can already manually overlay 100% accurate text over your image and let stable diffusion naturally blend it with the original background. Only the other hand, not destroying existing text when doing low denoising could be really desirable...
Thanks!!! Greeting!!
Hi my friend! :D
Will we be able to run SD3 on our own machines?
There are supposed to be three (at least) variants of SD3, the smallest one should be even smaller than SD1.5 but supposedly it will be worse than the largest (the largest is supposed to be way bigger than SDXL)
I signed up to get in on the testing. Got an email saying, don't call us, we'll call you. 😂
Is this stand-alone install or web based only?
I wonder if SD3 will require more vram and stuff. If so, I can't really run it :(
They said its gonna be more optimised than XL
Does anyone know if an API will be released with SD3?
There's no data about that yet but we are waiting for the white paper. Maybe that will give us a clue. But probably we'll see an API soon.
Don't expect commercial licensing anytime soon, so I don't think it will.
do we know how much VRAM this will need?
SD is open source and therefore, in my opinion, always the clear winner. Midjourney does not interest me, but when Dall-E came out, the prompt comprehension was its big advantage and has been for a while now which made it above and beyond SD. All me and my friends kept saying was if SD could get comprehension like Dall-E everything would be right with the world. Now seeing what SD3 is capable of I am more excited than I have been in quite some time.
I still remember that I actually wanted to do exactly as in the last example, last year. Positioning "elements" in a certain part of a picture. Nice that is finally coming.
How about human anatomy? Hands are just nightmare to make it right. Would SD3 finally understand that the hand has 5 fingers and no more and not less?
We don't know yet. Stable cascade did manage to get further, so maybe they took their learnings from that into SD3. Time will tell.
5:07 Interestingly, Dall-E makes pretty accurate keyboards.
That's a tough challenge! Just like hands
Have you ever seen a keyboard?
seems to be pretty fine
The biggest question - is it censored?
I do believe it will be a "safer" version than 1.5. However, most likely more free compared to MJ and Dall-e.
Not for long
@TheBagOfHolding Matt said it will be pretty hard to crack a NSFW version out of it, but only time will tell. 😅
@@undergroundxp if its open source then isnt it easy money?
@@undergroundxpnot difficulty
Does it will be open source?
Yes
I can has cheeseburger?
is it gonna be censored?
ok you had me at 0:14 the most well placed dad joke ever
noob question.. my a1111 install is SD 1.7 ... confused about the difference between it and SD3
A1111 is the UI (means: the platform you use models with) while the model in itself is numbered after another way. Currently, the latest Stable Diffusion model is SDXL 1.0 but it has nothing to do with the number of Auto1111
Sd3 will be better over time with support control net, addon, etc
Stable cascade and now sd 3 can you please explain
"How can you get it?" misleading, you'll have to wait
Thank you
so it wont be local?
who said that? lol
@@mycelia_ow my overheated graphics card whispering me
But can it do hands?....
what is mean. so we get terabyte new models now sd3 LOL or we can just use what we have even not know what we have.
there is no rush. there is no need go sign up. nt needed sign up get sd2 or what eva a1111/forge use lol..AND best thing last SD4 coming. SD2 works
Stable Dufter 33 3? Obviously some brand hate there 😂😂😂
The "AI" 0:17 felt a bit... off
they are censored and not nsfw friendly anymore tho, its a limited thing
"2 girls, one cup"
SD3 is the winner in all of the cases... including aesthetically.
Why: I prefer RAW to pre-cooked aesthetics and filters... whether in photography, film... or AI generation. Both Dall-E and MJ have decided what "look" the generation should be. As the designer, I will determine the look-&-feel myself. The most important thing is the coherence to the prompt. Period. Dall-E and MJ both failed miserably. Toys for children.
SD3 is a professional tool for professional workflows.
Slowpoke Kamph =)
Thanks for your wonderful guidance
I will upload an image1 of a person
I want to upload an image2 of an apple
I want to make the person in the image 1 eat apple in image 2 with his left hand
Later I will upload an image3 of an orange
I want to make the same person in the image 1 hold orange in image 3 in his right hand and eat apple in image 2 with his left hand
How to do that ?
My constraints : I don’t want to install softwares on my computer due to my hardware requirement not being sufficient enough and I can’t afford it. I want to achieve in some websites which freely allow me to do that
I will be highly grateful for your help. Thanks
Hah.. the weight list. :)
Hands!!!!!
If you do it now you are probably the number 200.000
Any idea how it is with NSFW graphics or making funny graphics from persons of importance? Like any good cartoonist of political satire would do with his strokes?
It's likely going to be at least a bit censored, but users will 100% find a way to finetune it and make it way less censored
where hands?
no hands= no problem 😂
SD3 shows you what AI generators really look like without unethical scraping and stealing others peoples work for training. It looks like a creepy Shutterstock.
Did they mention not scraping?
Gimme gimme gimme
Hoping that they will not butcher it like the SD2 disaster
I think they've learned from that! :D
I dont even know what SD I have. 🥴
👋
🥰
OOOOOOOOOOOOOOOOHHH HO HO HO!!
Does sd3 filter nsfw?
Probably but people will finetune
Stable Cascade doesn't, I imagine the same case here.
@mycelia_ow I couldn't get stable cascade to do anything nsfw. Do I need a Lora to fine tune it or something? Or maybe I need to figure out how to merge models
@@GES1985 depends on what you mean by nsfw, I've had cascade generate nude images, more easily than SDXL. Beyond that however I don't think any open-source model can do by itself since it's not trained on anything that "artistic". You do need Lora's or trained checkpoints for that. model merging is also an option, but the base models can only do so much.
Well def not getting Dall-E ... Needing to photoshop out naked diverse people from every picture would get annoying...
So tell it not to generate that. You can have it ignore its hidden prompts.
The only reason i dont subscribe os for stpid jokes...
DALLE 3 is still better than both midjourney and sd3 , if you eliminate text imo
That's quite a statement with no access to SD3.
@@thanksfernuthin It's not even true compared to Stable Diffusion XL/2.1, only out of the box. D3 is pre-tuned and SD is not. Comparing them isn't even fair.
Stable Diffusion by itself is not as good as Dalle-3, but what can be built from it is. There are incredible things you can do with SD.
You cannot do anything with Dalle-3 beyond what OpenAI lets you, which is just prompting and any limited preset options in the API. Dalle-3 is not better, they aren't that comparable.
Dalle-3 has the LLM advantage and is already tuned, SD can do the same and go much further. You just have to do the work yourself. SD3 will only make this more accessible and powerful.
@@mycelia_ow Well. We're in the realm of subjectivity now. I was just pointing out the illogic. I think people can disagree on value and quality of the different models regardless of state. We need access to them first though.
@@thanksfernuthin They should be judged more for what they can do. For SD it's a lot more than anything else out there.
SD3 and further models just reduce the work you need to do to achieve the same results and more.
It is at least objective to say SD and open source gen in general has a lot more potential than anything closed source.
No way you made a video about this? 😂😂😂 you hungry or what.
Because its a tech and AI channel why not make a video about something big in image gen
Shitty clickbait title
haahah so funny...
First.
❤️