Stable Diffusion, NovelAI, Machine Learning Art - AI art generation discussion and image dump

[chiobu]

I've been using this repo along with a Paperspace account and it works pretty well:

GitHub - Engineer-of-Stuff/stable-diffusion-paperspace: Jupyter notebooks for Paperspace.

Jupyter notebooks for Paperspace. Contribute to Engineer-of-Stuff/stable-diffusion-paperspace development by creating an account on GitHub.

github.com

Guide here: https://proximacentaurib.notion.sit...o-get-nervou-b83c2213f17e452e8b0e37ba64fe9758 (Archive)

The people over at rdrama.net have trained a Marsey model that can be used with Stable Diffusion too!



https://rdrama.net/h/marsey/post/104962/dreambooth-marseys-marseyastronaut-added-colab-instructions (Archive)

https://rdrama.net/h/marsey/post/109951/heymoons-marseypeace-guide-to-good-ai (Archive)

 

[NoReturn]

Found this on Twatter.
View attachment 3747805

Anyfart. So I just searched stable diffusion on twitter and there's literally only two things people are doing with it. Massive boobs and minors. Disturbing.

What does the robot think of Loss?

View attachment 3747806

View attachment 3747819

It struggle.

Here's when I feed it the original comic, though.

Alright, downloaded and set this up following the Voldy guide and, using the default Huggingface model with the following text prompts:

Prompt: desert mesa canyon landscape at sunset, masterpiece, best quality
Negative Prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry

I got this:
View attachment 3747922
As somebody who lived in Arizona, this looks 100% accurate.

I wonder if there are ways I can out fill this to a widescreen resolution (or ultra-widescreen even) just to see how far I can push landscape generation...

Largest I could do

 

[Telemeter]

Color me surprised we're at this stage, but still what is produced isn't nearly as coherent as the flat image. A lot of work is going to have to be done to make the model usable in most prompts. You have it generate a decent model of the basic pumpkin here, but it's covered in bumps, it has a million eyes, and I'd hazard a guess there are quite a few reasons you can't straight up use what it generates if you want it to mesh with already existing models.

Like I said, like a year or two. Specially once they realize they can make unlimited Overwatch 3d porn.

 

[PointOfNoReturnPoint]

I can only imagine what butchered stinkditch roast beef this AI thinks is a vagina.

Keffals is salivating at the fake information he can generate on this alone.

 

[a feel]

Keffals is salivating at the fake information he can generate on this alone.

Whatever you do, don't taint this beautiful gift to humanity with photos of actual stink ditches and skin sausages.

 

[PointOfNoReturnPoint]

Ladies and guntlemen, I present to you Ethan "Gunt" Ralph, host of the Killstream:
View attachment 3747908
it even got the quad tits right

on a quick glance, that looks like the real deal, it falls apart upon closer inspection but, this is kinda scary. Anyone can now generate anything and say it's real with no way of disproving it....We really are in the timeline of manufactured truths.

 

[Estate]

I did a test for txt2image, default settings and set to 100 batch count. On a 3080Ti it finished the process in 5m27s.

"goose emperor in golden regalia"

 

[Freezard]

I miss VCR.

 

[Cuckoman]

This is going to be a gigantic shitshow. I've barely spent a day fucking around with this and I can tell you right now you'll be hearing about this for years.

• Artist outcry about being replaced by a $600 GPU.
• IP concerns in that an artist's style can be trained into a hypernetwork and reproduced instantaneously.
• Government, legislative, nonprofit, and public outcries regarding the triviality of producing near lifelike CSAM.
• Privacy concerns in being able to make passing nude / otherwise offensive or humiliating images of real people.

I'm calling it now, the Information Era is dead and we are now in the Machine Era. Computer information processing is now more important to the direction of society than the general ubiquity of information which defined the information era, just as the WWW redirected the world from the Nuclear Era and the cold war standoff to a global interconnected community.

Governments using ML to memorize faces from public IDs to track people in public (China literally already does this), civil litigation over copyright defemation and nonconsensual pornography, criminal litigation over production

It's happening.

They can learn to code.
If an artist can not create art that supersedes what an AI can create then that means that his art is not very good or elevator-music-quality of art.
In that case, meritocracy demands that that artist is replaced with an AI if the AI generates "better" or more desireable art from the consumers.

Now, lets pray we can do something similar for voices. Get rid of voice acting as a profession, but more importantly,
I want to see good quality Text-To-Speech AI to automatically convert the entire ProjectGutenberg library into audio books.

 

[IamnottheNSA]

Alright, downloaded and set this up following the Voldy guide and, using the default Huggingface model with the following text prompts:

Prompt: desert mesa canyon landscape at sunset, masterpiece, best quality
Negative Prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry

I got this:
View attachment 3747922
As somebody who lived in Arizona, this looks 100% accurate.

I wonder if there are ways I can out fill this to a widescreen resolution (or ultra-widescreen even) just to see how far I can push landscape generation...

You can set the resolution from 64x64 all the way up to 2048x2048 and you can make widescreen pictures. Just don't make it too big, or else it will take an hour to make a single batch of FullHD pictures.
For me, 1024x768 is good enough.

 

[Raven of the Burning God]

Now: No, that's a deepfake! Look! You can see where they edited the video!

Future: While the prosecution would have you believe that the grainy security video is my client, you can see from this other totes real and authentic grainy security video from a different angle that it must be someone else.

Future: While the prosecution would have you believe that the grainy security video is my client, we have this other totes real and authentic grainy security video of what is clearly my client doing something across town at the time the alleged crime occurred.

 

[Toolbox]

Like I said, like a year or two. Specially once they realize they can make unlimited Overwatch 3d porn.

Am I right on the breakdown here, that diffusion is not actually generating the model but producing multiple angles in 2d that are processed in a photogrammetry algo to dump a model out? The whole pumpkin with 100 eyes thing makes me suspicious it's not actually stable diffusion, because that kind of thing seemed to stop being a problem well before diffusion came along.

 

[naaaaiiiiillllll!!!]

Largest I could do
View attachment 3747971

I can see that being a painting in Minecraft (one of those low-res paintings you can place in your block house). Very cool.

Decided to play a bit more with people this time:

Mao Zedong wearing nurse outfit in hospital room (apparently the Mao suit IS the nurse outfit)


@JosephStalin air force pilot old man:


@Hollywood Hulk Hogan going fishing at Lake Tahoe:


Richard Nixon and John F Kennedy eating at Denny's:


I need to stop now. This is too fun and my laptop is a little slow generating these. I expect some great things from this...

 

[Caesare]

AI generated images can't be considered intelectual property of nobody

 

[Catgirls are Love]

There was shit I was going to get done today but this is too fun to not play around with! It's pretty quick even on my GTX 1080; some of the results are one part horror yet three parts hilarity.

 

[Narcotics]

Traditional physical pen & paint chads keep on winning.

 

[tehpope]

Running on a 1650 super

I tried "lost in the backrooms, 2pac". Got b&w still of some ODB lookin rapper.

 

[NoReturn]

How do I use it? Is it online?

You can do it locally from the leak, but the guys from whom it was leaked are still up and still available online: https://novelai.net

 

[Judex Meus]

EDIT: Thank you, this was the issue, I actually had 3 different older Python versions.

This is off-topic but managing python versions can be a huge pain, especially if you (knowingly or unknowingly) install multiple versions. If you or anyone else in this thread still have python 2.7 installed, uninstall it immediately. It's been deprecated and you'll save yourself massive headaches down the line.

 

[AmpleApricots]

Maybe I'm just paranoid because I mean just look at the fucking world right now, but I just can't help but feel like this technology is the beginning of the end, or maybe like the middle of the end.

Things will change, they always did. Computers changed the world, and AI will probably change it in very impactful ways again. I see a lot of potential to improve the quality of life for everyone, if used correctly.