Stable Diffusion XL weights were just LEAKED by an unknown researcher!
Follow us on Twitter!
---
Was it leaked?
Yes, and this happened faster
than I think any of us thought it would
because initially when I saw this tweet,
I figured, oh, okay, interesting.
I guess Emad officially wanted to release this.
Yeah, here, like they don't mention the word leak anywhere.
It just says Stable Diffusion XL 09 is available
with two links to Hugging Face and a Colab notebook
and also saying you could run it locally.
So I figured, oh yeah, of course,
the weights must be available.
And we knew from what Emad said
Curiously, both of these pages on Hugging Face
have now been fixed.
And that's true both for the base model
and the XL base model.
I don't see the safe tensors weights,
So basically this is someone who applied
for the research permissions to use the weights,
which as Emad said, the idea was to give researchers
the ability to use this so they can get some feedback
from professionals first.
One of those people decided to be a little goofy
and similar to Lama, I don't think this was planned,
this is clearly not planned,
just released them on the internet.
And they're now available in a link somewhere,
which I am not going to post on YouTube,
but if you happen to dig in Reddit,
I don't think it would be that hard to find.
And yeah, this all emanates from a user
called MysteryGuitarm.
And they've done a lot of fine tuning work
on the stable diffusion subreddit before,
you've probably seen them before, they're an engineer.
But before we go too far into this,
I do wanna say their post here is very important.
So if you wanna mess around with this,
generally speaking, you wanna avoid doing this
with those checkpoint files or those .ckpt files,
which is kind of a weird way of compressing stuff in Python,
and they're technically executable.
Now that said, what's cool is people are using this
and the results are pretty incredible.
So some people are saying,
oh, I'll just wait for the real link.
These are all links that are definitely not working.
So right, it doesn't work in automatic 11.11 now.
Obviously, there's some tweaks you have to do.
The key here is having PyTorch 2.0 and no Xformers,
which is kind of interesting.
The other curious thing is there are,
the reason there are two models
I've gotten it working,
but I'm gonna be careful sharing this
because I'm not sure how YouTube is going to treat that.
We'll put it up soon.
But in the meantime, we're gonna be a little careful here.
Now, what's interesting is Comfy UI also works.
So there are other non-automatic 11.11 UIs
that work with this.
The coolest thing I've found digging through this
is that it looks like you can do animations with it.
So similar to Pica and a few other models,
animation just works out of the box, which is kind of cool.
And what I think is cool is there are people
who have tuned the FB16 safe tensors base.
And it looks like even with a meager RTX 3060,
within about 30 seconds,
you can render 1024 by 1024 pixel images
right out the bat, which previously was not possible.
Clearly, this is a, not a quantum leap,
but it's a big step forward
in terms of the performance of stable diffusion.
And where previously you had to have a bunch of goofy steps,
now it's a little bit easier.
And so you guys know,
I was just running this through the terminal.
I wasn't actually using a UI.
I've been curious to try Comfy UI,
so I might actually look into this.
And yeah, so the key is, yeah,
you need the base model and the refiner.
Both models are necessary.
They're not really used separately.
But when you prompt this,
which it is more complex to prompt, I should say.
Technically, this is the research version of the model.
So you can still disable the safe work filter if you want.
Whatever, if you're into that.
This is what's going on here.
There also curiously is a Google collab
that's available right now,
but you do need an official access token though,
and you have to be approved on a hugging phase
for the research weights.
So this is not gonna work necessarily
if you do not have the access.
Now that said, technically with RunPod,
you could put in the unofficial weights
that may or may not exist,
but there are some incredible images that have come out.
This is coming from the artificial guy.
And what's crazy is these are single shot.
So these are coming straight out of stable diffusion, Excel.
And I think these models are crazy.
What is important to remember here
is it's not exactly surprising
that a model that just creates larger images
out of the bat would have more detail.
That said, the coherence is great.
It's approaching a general set of capabilities
that is very close to mid journey.
I will say it is harder to prompt
in terms of just with bigger images,
there's so much more availability for different details
and different attributes.
So obviously longer prompts
generally will get you better results.
Информация по комментариям в разработке