STABLE DIFFUSION 3 NEW LICENSING! 

It's crazy how allergic these companies are to a simple, honest "we messed up." It should be the norm, but it's so rare than any company willing to genuinely admit fault instantly gets goodwill for it, but the vast majority still won't do it.

I mostly agree with you, but the important thing is they're trying to correct it and improve now. So let's give them a chance to.

Stable Diffusion, Adobe, Figma…

Their responses are so corporate, it's like the community is talking to Adobe. Their discord is worse but in a different way. No official responses on anything at all, just ignoring any questions from the community.

Yeah that must be it. They also lost a big client since civitai is used to run their models so they must have some kind of contract with civitai so they must have lost some revenue from it

I wouldn't say it was the only reason, the negative press and other factors would play a part but yeah, I'd definitely agree the loss of CivitAI and effectively the community with it was a major factor. Whether it's too late or not is another question.

a lot of people will "nuts" on the fine tune model.... -__-

How long you are in this bubble of SD? Many words I don’t understand and I wonder how much time it would take to speak your language ;)

@@Vanderful_life ohooho if you look at the civitai model and turn off the nsfw filtrer you Will know what I am talking about...

Flux.1 makes sd3 looks like a dumpster now..

@@kenrock2 Yup and most of the AI researchers are Stability alumni.

😅🤣

You bet! 🌟🌟

@@sebastiankamph what a mellow dude cheers brother

+1

@@TheDrumtracks how much vram do I need for it? Can my poor 6gb card handle it? Is it harder to run than sdxl?

@@DaniDani-zb4wd I tried the google colab demo. It loads two models 9GB and 10GB. So it won't work with your 6GB card.

true!

What is pony?

@@mik3lang3lo I think it's p0rn

@@mik3lang3lo hmmmmm nothing xD

@@mik3lang3lostay pure. Don’t get swayed in by the pony 😢

@@mik3lang3lo Pony is a trained checkpoint that will forever change your world view of image generation... Like... FOREVER...

16 channel vae will save everything! Hopium

No way. Too many people using them. Comfyui is too complicated for most users

@@TheCynicalNihilist Sort of. I want to see where it's heading first. But eventually, yes!

Sad if true

@@NotY0urHeroo why so? comfy is more powerfull. I was using a1111 myself and was skeptical about switch to comfy, but once i tried i wont ever go back, it's just everything in a1111 + so much more.

No true, only comfyUI users hype. ComfyUI is actually very sad

At this rate, you will die of old age before that happens. No company currently on the AI race is going to risk their necks for our entertainment. They will keep the uncensored versions for themselves. Our best shot is to fine-tune and that will have its limits because they will intentionally continue to release the worst versions of their models until they can ensure their best versions are not exploited for adult content.

Just wait for Pony SD3. SDXL was also censored and now we can prompt titties and other kinky stuff. 😏

Still optimistic, huh :) It depends on how interesting SD 3.1 will be for the community, we don't know it yet.

So your statement means that you are going to do it? A lot of big players in the customising game already stated on Reddit that they won't touch SD3 because anatomy is too broken.

@@bxl2012 Lets see what kind of fix they release..

too little too late, I guess.

What are those alternatives?

@@tomreiner8676 Currently SigmaArt but at the moment half of the team that left StabilityAI including the creator of ComfyUI are in pre-production phase of OpenDiffusion (a new model)

There are no other good open source alternatives. The only alternatives are closed source

Maybe playground ​model. @@titanmars9229but for me ,SD is the only thing I use.

@@titanmars9229 This is subjective opinion, not universal truth. There are a lot of good open source models. BTW Stable Diffusion was NEVER OPEN SOURCED, only open WEIGHTS, which is completely different :)

So, a fine tuned model is similar to modifying a vehicle. You can put bodykits on it, buy aftermarket upgrades for the engine and bigger wheels etc. But you're still driving the same car made by the same manufacturer. I think this analogy is best. Another way to look at it is like roblox or unreal engine. You can create a game within their environment, but your game cannot exist or run without the underlying engine code.

An image created via AI is not considered a derivative work as far as I understand. Selling a derivative would be like selling a fine-tuned version of the model itself.

Oh as far as i see, my linked film now seems to be outdated right? so this would be not perfect but an improvement

Tack, väldigt snällt! 😘

That's what people are saying, but I don't know if it's true.

definately. A1111 is just an interface for SD models to load into.

@@brootalbap Yes, but SD3 is currently only avail for Comfy

The base model hasn't been trained on any NSFW material, so it doesn't know how. It simply doesn't know what people look like without clothes since it's never seen anybody without clothes. When/if we get fine tunes it'll probably be ok.

@phizc thanks for the explanation man!

@@phizc That can't be true. They need billions of images to train, which means a giant dataset that wasn't manually put together, which quite simply can't completely avoid certain kinds of content reliably. I'm sure there was nudity in the initial training set.

@@peterlewis2178 maybe there was. But it wasn't told what it meant. CLIP works by associating text tokens with an image. For example, you can show it hundreds of images of dogs along with descriptions such as "A brown dog in the field" and "A Labrador dog on the sofa" and the model will learn what a dog is by what's common-ish in the image. I'm not 100% sure how the diffusion model uses the CLIP guidance exactly, but the main point here is that it needs the description during training (or at least booru style tags). So, if you don't want the model to know how to create images of naked people, celebrities, copyrighted materials, artists styles by name, etc., you simply filter out all images that contains any of the "censored" words. So while it's probable that it's seen naked people during training, they have not been tagged or described as being naked. If you ask the model to create an image of a naked person, it doesn't know how. It's like showing an artist a thousand photos of people without saying their names, and afterwards asking the artist to draw an image of one of them by name. Of course that would be impossible since it could have been any one of the people in the photos, but the artist might try to use statistics if their memory is great. If the name is female, ignore the photos of males. What's the average look of the females shown, e.g. skin color, age, clothes, etc. So they would be able to draw an image of someone, but it probably won't look anything like the correct person.

@@guvrd75ecvu6.hgvjkuhtc thank you, that's very kind of you ☺️

🤩

@@hh1162 honestly, I think it's going to be tough. They have built their product around an audience that generally likes their stuff free and open source.

Woooh! 🌟

Almost!

a1111 is just a user interface to load stable diffusion models. SD3 is better than 1.5 and SDXL in SOME areas, but it's still lacking. If the human anatomy gets fixed in next release, it will most likely surpass 1.5 and SDXL in quality, not counting 3rd party models or extensions like controlnet.

@@sebastiankamph Thanks for the answer