Stable Diffusion 3 API Released.
Вставка
- Опубліковано 18 тра 2024
- stability.ai/news/stable-diff...
x.com/StabilityAI/status/1780...
Prompt styles for Stable diffusion Automatic1111, Forge, ComfyUI & Vlad/SD.Next: / sebs-hilis-79649068
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... - Навчання та стиль
A few precision:
This is NOT the "real" SD3 model, the API one is a much older model that is not gonna be representative of the final model because...well...the SD3 model is STILL in training and will be released when it's ready. The API one was probably released because of a contract with Fireworks AI who made the workflow for that version of the model.
So YES, for those asking the SD3 model will be free and open-source, it will be much better than what you see here and it will be released to the public when it's ready, so be patient yall.
thnks
While it is correct that it is not the final form of the SD3 model (which was addressed in the video and in Stability's news post), it is in fact very real and not "much older". There are different internal versions of SD3 currently as the training progresses.
You are also right that the final version will be free and open-source. With free comes licensing limitations however.
Source: Stability AI
@@sebastiankamph Sure licensing limitations that will be vigorously ignored by the vast majority of the community.
Besides. No one really wants SD3.
We all want the Fine Tunes and Loras based on SD3.
@@vi6ddarkking straight to the point. I really wonder how how hard it’s gonna be for developers to finetune this model or to create loras. This is why it took so long for sdxl to get “good” many people still use 1.5 simply because they don’t wanna give up on all the loras… and still even in present I feel like there are more loras released for sd1.5 than sdxl due to training issues..
The music is very distracting, maybe have the music volume lower next time.
or just don't add music : ) thanks!
Thank you for the feedback!
@@sebastiankamph Please don't add music with lyrics when you are talking. Thanks.
I am a native Spanish speaker, and I had to go back to the video when reading the comment to check how loud the music volume was because I didn't even realize there was music hahaha. Maybe it's just each person's experience. Great content as always! Looking forward to the launch of the weights. Great content as always Sebastian! Cheers from Uruguay!
Agreed. Music is added by default to lots of things now for no reason what so ever. Even my local radio news has a constant drumbeat in the background and it's annoying to the point I can't listen to it any more. What most people fail to realise is that it reduces accessibility. Background music cases havoc for people who are hard of hearing and who need hearing aids. The devices often amplify the wrong things so the voice gets drowned out completely.
Please no music. It sounds like another tab is open.
get rid of the background music
Look up the word “please”… please
Nah its insane
The problem is not the music but the fact that there are lyrics.
Good feedback. Was testing an AI generated song instead of the usual background music.
@@sebastiankamph I watched that video. I will have to try that program. The music is fun. I am hoping Bandlab will eventually upgrade its AI music features.
Gunna be hard to get me off 1.5 with my 50 extensions, 100 trained models, and 3000 Loras.
Was hoping you would show mainly images you created on sd3, especially if you’ve had prior access
They didn't really want us showing those, as they improved on the model before releasing it publically ;(
do you have a tutorial about how to train stable diffusion to generate similar videos to the video you give it as a source? TY
In fact, apart from a further advance announcement, of which there have already been several, there is still nothing.
I'm curious to see when the models will finally come out for download so that you can really see what you can do with them. Until then, I find other developments more exciting at the moment.
Now that we appear to have a functional text generation, I'm curious about the implications of copyright for the fonts in the training data. AI companies are already being sued by creators of text and images used in the training data… are foundries the next to jump into the fray?
I am honestly salivating for the next few months.
Once the Community has had the time to Fine Tune SD3 And Develop the best practices to train the New Models and Loras.
Things are about to get really fun.
In the sofa picture one of the dots disappears from 2:00 to 5:15. I wonder how good the text really is, how often they tweaked the pictures before releasing them.
When they do the open release, will it be uncensored? I’ve just tried and for example “woman in bikini” fails because of content moderation
haha great info, and the joke in the middle made my day!👏
Happy to hear it :)
Why don't we have different open source models like LLM? For instance there are many architecture derived from transformers, like mistral, llama ... But for stable diffusion there are a lot of finetuned models but not new architectures. Do you know any other kind of architecture used to generate image with high quality like stable diffusion?
I have been enjoying SD2 w/a1111. Will something like that every be created for SD3 so i can run locally, for free, as much as I want?
ComfyUI. The way it’s designed should work out of box (or with minimal update) for SD3.0
SD3 will be available for all user interfaces as soon as the weights are released. Currently it's api only.
@@sebastiankamph Thank God. I appreciate the positive news!
ELLA did such a great job in prompt comprehension to the point that it’s difficult to see what SD3.0 is adding, if quality of example is what to go by.
So how do you use it?
I think text implementation will be better overall. The initial base model will always be the “start” and people will often overreact at the quality. As time goes on and with improvements and fine tuning, the forthcoming forms of SD3 will be better.
Yes! 100% agree. I'm sure the improvements they make in the coming weeks will get it even further, and then the custom trained finetunes will take it all the way.
Is Automatic1111 still the thing at this stage? or is time for me to move on do you think? I use all kinds of plugins etc.
I still use a1111 primarily. Sometimes I use Comfy, sometimes Fooocus, sometimes Forge.
@@sebastiankamph Ahh ok thank you. Have you heard word of a model being around soon?
5:20 Your dad jokes give me life XD
they are not gonna paywall everything are they?
interesting. the first time I watched this video the music wasn't distracting to me at all. I didn't even notice it that much. but liked that there was something in the background to listen
Different strokes for different folks I guess. Thanks for the feedback :)
Do you have and idea how to use it? it looks complicated, I don't understand anything 😅
Wait for someone to implement it on their website so you don't have to use the API yourself.
I don't like the focus on the simple, instant result. While nice and impressive to the majority of people, the prompt to image is only the first step imo.
The options to fix and improve upon images, things like controlnet and comfyui, that is where the magic happens!
it wll be localy or not ?
Yes, in few weeks or so hopefully, free for personal uses. not commercial tho.
@@LuckyPed10 where did you get this info?
I think it either won't be, or they are waiting to bake censorship into the weights before releasing. The politics are different in the company now that Emad resigned.
Music too loud but ty for the vid.
Wouldve been nice to see more examples & how we can use it. Also its good to mention like half the leadership of Stability AI left during the last month or so due to their not actually being open.
Its semi open sourced...not fully.
hi there, is there a way of booking a call with you for consultation?
Yes, pm me on Discord.
"Thank you S. for a great video... It's great that you got to try SD3. When it comes to pricing for commercial use, which payment model are they talking about then, and how much? It would also be good to get a simplified explanation of what type of product use they envision is acceptable. Obviously, many have used SD because that model has been free and maintained a relatively good standard regardless of blood, boobs, and other personal nuances, which as you know are hard to even come close to with models like MJ."
Thank you! You can read more about that here: stability.ai/membership#select_membership
Have tested it SD3 and i am disappointed, prompt coherence in not even close to Dalle or ideogram... very sad...
What's the point of sd3 if it as censored as mid journey and dalle 3, atleast Mj and dalle 3 has better image quality than sd3, i hope that a less censored version of sd3 will be available in the future.
SD could add an option for HUMAN FEEDBACK to continually improve, as with MJ ?
It is a possibility for sure. It will also skew results towards what people "like" instead of what might actually be correct. MJ had that problem a very long time, everything was just looking beautiful and artsy, for a time it was almost impossible to achieve simple realism.
I don't like it that SD3 currently looks bad compared to lykon's examples on his twitter page. I think it was a bad move to release a half-baked workflow version of SD3 that doesn't meet the aesthetic quality of MJ6. Looks like i'll be sticking with SD1.5 models for a while longer
Emad says SD3 can generate video as good as Sora. Did anyone try generating videos?
❤
❤️
So... you said that you have access to SD3 for few weeks and all you show are images I can find on net myself. Clickbait?
They wouldn't let us show images from the closed testing. And if I did, I wouldn't get invited to closed tests like that again.
@@sebastiankamph So my apologies and thanks for the quick reply. This explains everything.
please cover Hyper SD
the best thing is: uncensored
Stability: we MUST stop smut at all costs!
The music was nice IMO, particularly when there weren't singers though. It's the additional "talking" I think that makes it particularly problematic, (plus it's in a similar register for a double whammy) that makes it tricky to understand/catch your voice alongside it. Do the Spanish uphold their ideals? Si'Bastion!
Thanks for the feedback. I don't know what bastion means, but it sounds like a great dad joke
@@sebastiankamph it's English for a kind of 'last stand' 😉
Bastion: an institution, place, or person strongly defending or upholding particular principles, attitudes, or activities.
Why and how is all this free?
What are the technical differences between SD3, SD3 Turbo and Cascade?
Interesting video topic ?
via Pi
.
Great question! Stable Diffusion 3 and Stable Cascade are two distinct models developed by Stability AI, and they differ in their architecture and capabilities.
* **Stable Diffusion 3:** This model uses a spatial compression factor of 8, encoding a 1024 x 1024 image into a 128 x 128 representation. This enables efficient processing of high-resolution images.
* **Stable Cascade:** This model employs a unique, three-stage architecture, achieving a much higher compression factor of 42. Stage C transforms user inputs into compact 24x24 latents, while Stages A and B act as a Latent Decoder, similar to the role of a VAE in Stable Diffusion. This architecture allows for additional training and finetuning on Stage C, including ControlNets and LoRAs.
In summary, the main difference between Stable Diffusion 3 and Stable Cascade lies in their architectures and compression capabilities, with Stable Cascade offering a more efficient compression factor for handling high-resolution images.
*Maybe it will become open source in future*
Yes, you will be able to download the models (weights).
If it doesn't it belongs to the trash.
Good joke
SD3 for ComfyUI node just available today
Is that serious?
Still free or not?
I preferred Stable Diffusion 1 over Stable Diffusion 2. In part because of bikini's in Rembrandt style, but also because it had more freedom in creativity. Stable Cascade could not do artist styles any more. Reading this "Bad Actors" text here, I expect this one will be even worse, although can maybe draw hands and text 😢
I couldn't watch the whole video because of the volume to the background music with lyrics. Too fatiguing.
Thank you for the feedback.
"this is what turtles, uh, would've looked like if, uh, was kinda, half, semi, real"
turtles are the new pidgeons?
I just wanna tell everybody in the comments that we have enough anime porn to last us a lifetime so if you guys wanna focus on some other type of art, that’d be great
The music didn't bother me.
Thanks for the feedback :)
so, there is a high possibility of being "open source" as it was, I thought they are now going to stop being open and start get paid for their models
It will be open source and available to download. They will have a pricing model for licensing.
Open source doesn't mean you can't make money. Red Hat, SUSE, Canonical all exist as companies that make money.
I hope my 6gb card can somehow run it, just to try it once.
Porn. That's the main difference. SD can do Porn. The other closed source models cannot.
correct the power of boners is stronger. and if V3 can't be used for adult content adoption will be low.
The background music is very loud and distracting.
I just hope they stay open source. That would be a real downfall if this goes closed source and censored.
As of right now, their plans are still to keep it open source.
Noone tell him that turtles are real.
😅
donatello
He was my favourite! Which one was yours?
@@sebastiankamph Don and Leo for sure!
@@sebastiankamph Michelangelo had the best sense of humor and was the least burdened by pointless stuff. Also nunchucks are cooler than sticks, sharp sticks, or pointy sticks.
At this point i think its best, for professionals, to use midjourney because the details and being so on prompt is looking unreachable anytime soon by any other source BUT to use SD for inpainting what you cant get in MJ.
Sucks, i wish SD in automatic1111 could get on that level.
I mean if you're just using a prompt and then being happy with that image, sure, MJ has got a lot of them beat. But with client briefs and demands, MJ has no place in my workflow where images and videos have to look exactly as described, with particular poses, colours, fabrics, faces etc.
It's hard to call something "professional" if you have no creative control over your work.
The bg music is too loud... either you speak louder or just remove it totally.
Not free
Dude, you really need to articulate better, I'm rewinding constantly just to make out what you said.
half of the video and already i am shocked... is this a joke? stable diff 3 is this? even with instagram edits u can u do more... has to be a joke right?
So sad the dream of the community based models comes to a quick end... money making $$$
this mediocre loud music is unnecessary and annoying. just give us the info. We're not partying here.
annoying background music. put an instrumental next time 😊
Music? Man! You are not 13 years old boy!
Boy? Sir, I am a 13 year old man!
Now the waiting game begins, so people train their models.
I hope the training process is faster than SDXL, and to be honest SDXL was a disappointment.
SDXL models and the turbo variants are great, not sure what are you on about.
Not free. Costs credits. F-all that garbage.
That's because you're using someone else's service through an api. When the weights are released, it will be free to use with your own machine.
@@sebastiankamph Got'cha.. Thanks.
Honestly disappointed, saw a lot of pics generated on Reddit and it is not really mind-blowing. The hands are still shit. I am sure SDXL is more than enough.
I think you have to consider that it's a base model. The base models of 1.5 and SDXL are not great, far from it. With prompt understanding like this and then custom fine-tuning them for quality, I have hopes we'll see stuff that is similar to, or surpasses, previous models. But that's my opinion.
The "mind-blowing" part seems already solved, even by SD 1.5 if you master some of the custom models. The problem is in text generation, composition, AI-spew etc. This means the AI part takes up only a small slice of my overall workflow. Lots of 3D modelling in Daz and Blender->pre-composition in Pixelmator->AI wrestling->recomp and enhancements in Pixelmator... With big chunks of that workflow requiring a LOT of iterations and backtracking.
@@sebastiankamphit is mind blowing for a base model. The base models for the others can't make a good picture at all from what I have seen.
get rid of the background music
Yeah, I can't hear the hentai in the other screen