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

[purpleflurp]

So a question: What's the minimal amount of images that one needs to build a coherent model? If I were to, say, train the AI with Harold McNeill's art portfolio (~100 images), would that be enough for the AI to work with?

Jesus. What's taking up so much space? It's gotta be the pictures it pulls data from.

From my understanding, the model was trained using the danbooru image database. Given the site hosts shit for almost* every fetish imaginable, makes sense that the model would be ginormous.

*Using Danbooru as a source would probably explain why it doesn't really "get" gay prompts, as a few others have pointed out.

 

[ZazietheBeast]

Looks like the starving artist shall remain starving for longer.

In all honesty, this is a great tool to have even for artists. As one could learn much form the machine on where to improve upon.

Also, does that mean anyone can literally continue sonichu if you have the AI read Chris' comics?

 

[Male Idiot]

I can definately think that this help out many, many indie rpg games with many portraits and isometric view. On the other hand, crypto is deflated and they could make a nightmare out of this.

 

[Mr.Miyagi]

FYI:
The novelAI torrent is ~56GB.

I haven't setup a local version yet, but based on the Voldy Guide, you only need the "animefull-final-pruned" folder and "animevae.pt" file, if you want to use the NAI model for image generation. Should cut it under 5GB.

Jesus. What's taking up so much space? It's gotta be the pictures it pulls data from.

The models are 5 or 6GB each, and there are a few in there that are previous versions, SFW versions, modules, github stuff, etc. But you really only need the 1.

 

[Male Idiot]

I bring you Evil Sonichu! With that, I'm off to rest. Work tomorrow, pay to bill, all that stuff.

Special thanks to Somethingproblematic and Noreturn for helping me dwelve into the abyss of the TISM.

 

[NoReturn]

Jesus. What's taking up so much space? It's gotta be the pictures it pulls data from.

I haven't setup a local version yet, but based on the Voldy Guide, you only need the "animefull-final-pruned" folder and "animevae.pt" file, if you want to use the NAI model for image generation. Should cut it under 5GB.

The models are 5 or 6GB each, and there are a few in there that are previous versions, SFW versions, modules, github stuff, etc. But you really only need the 1.

 

[LikeRivenIRL]

It's pretty addictive. Any decent artist will be able to fix hands and weird anatomy to perfect them.

 

[Puff]

I wanted to join in the Roon posting and actually generate her in a nazi uniform, but both waifu diffusion and the novelAI anime model seem to have little to no knowledge of shipgirls.

So anybody crying "this will replace bottom-tier deviantart/twitter/waifu artists", you need not fear. Just pivot to drawing shipgirls, or flavor of the month gacha waifus.

Anything new will always be a problem unless communities take off around AI and keep up to date models.

 

[Baraadmirer]

I wanted to see how well the software did with celebrities, so I made a few prompts with the celebrity's name, realistic, and eating a burger. It's usually the jaw that ends up looking really uncanny.

Spoiler: Results

Marcia Gay Harden:


Whoopi Goldberg:


Dwayne Johnson:


Stephen Amell:


Brian Tee:


Stana Katic:


And last but not least, Nathan Fillion (who looks the most realistic):

 

[Hokuriku]

I can definately think that this help out many, many indie rpg games with many portraits and isometric view.

There's already someone either in the /g/ or /trash/ general thread using it to illustrate a VN

 

[bigasscigar]

Anyone on linux and AMD able to get the gpu working? Followed the arch guide, not in docker, followed the troubleshooting guide and I still can't get it to launch without adding the --skip-torch-cuda-test
Which puts me on cpu making it slow as hell, still fun but a pain to wait so long for the result. I have a gtx 1080 lying around with a mobo and cpu, debating saying 'fuck it' and throwing a build together real quick just to fool around with this.

 

[Gaunt King Ithan Rilph]

Stupid sexy man-made horrors beyond my comprehension.

Spoiler: Positive prompts

((masterpiece)), ((best quality)), ((art by Giger)), horror

 

[Socrates]

The last thing your health insurance plan ever sees

 

[NoReturn]

Looks like the starving artist shall remain starving for longer.

In all honesty, this is a great tool to have even for artists. As one could learn much form the machine on where to improve upon.

Also, does that mean anyone can literally continue sonichu if you have the AI read Chris' comics?

Fun fact. NovelAI is a writing program originally. You could indeed feed it Sonichu comics and see what comes out.

View attachment 3751080
The last thing your health insurance plan ever sees

RETURN THE SLAB.

 

[seri0us]

I made these a while ago and I tricked somewhat
View attachment 3749176

View attachment 3749179

What prompts did you use for these?

 

[A Hot Potato]

Oh yeah, I did another picture of that brunette woman again.

 

[LikeRivenIRL]

Thought it was cute lol.

 

[Hopkinsville Goblin]

I'm trying to use Docker to install NovelAI in Windows 10. I'm following the guide and it looks like the pull worked. Everything eventually extracted. Then I got to this, and I don't understand:

3. Pull the latest image, and run "the master test".
4. If it works, great! Follow the tard guide inside the running image.

I seem to be able to get the images running but I don't know what "the master test" means. And I can't find a tard guide either.

 

[Puff]

I'm trying to use Docker to install NovelAI in Windows 10. I'm following the guide and it looks like the pull worked. Everything eventually extracted. Then I got to this, and I don't understand:

3. Pull the latest image, and run "the master test".
4. If it works, great! Follow the tard guide inside the running image.

I seem to be able to get the images running but I don't know what "the master test" means. And I can't find a tard guide either.

Link is in the OP, but it's this
https://rentry.org/voldy
And the master test is probably the Auska image the guide uses as an example. Basically if you enter what the writer did and get a highly similar output, it's been installed correctly. If you're getting reasonable outputs I'd say you're fine.

 

[DickMain]

I can definately think that this help out many, many indie rpg games with many portraits and isometric view. On the other hand, crypto is deflated and they could make a nightmare out of this.

That's what I was thinking. Quick background assets and tiles could be done so much more easily now.