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

[Stoneheart]

How hard is it to train your own Version?

 

[Ice Gun Ninja]

Mad at the Internet

Stable Diffusion (1.4)
Prompt:
sunset, sun behind ocean, (((clear sky))), dark sky

Negative prompt:
[[[no clouds]]], [[[beach]]], (((light sky)))

Seed: 887498
Steps: 25
Guidance: 7.5
How do I make the clouds disappear? I have tried (((clouds))) (((no clouds))), [[[clouds]]] and [[[no clouds]]] in negative prompts.

 

[Starved Artist]

Mad at the Internet
View attachment 4416440Stable Diffusion (1.4)
Prompt:
sunset, sun behind ocean, (((clear sky))), dark sky

Negative prompt:
[[[no clouds]]], [[[beach]]], (((light sky)))

Seed: 887498
Steps: 25
Guidance: 7.5
How do I make the clouds disappear? I have tried (((clouds))) (((no clouds))), [[[clouds]]] and [[[no clouds]]] in negative prompts.

I -think - I see your problem.


[[[no clouds]]], should be (((no clouds)))
And the steps should be a bit higher.
These are run at 40 steps and Euler A.
I'm gonna run it again when I get back, try DIMM and KARAS.

The prompt with Kotos what I was using at the time.


The Prompt with 1.5, as I don't have 1.4 for some reason...

 

[Eddie Riggs]

Cats

 

[Stephanie Bustcakes]

Dreamlike Photoreal 2.0 seems to fix a lot of the anatomy problems that other models have. I tested it by making some Disney princess as they would look in real life, including:

This mermaid:


This French beauty:


This Arabian lady:


This unusually long haired gal:


This very cold dame:


This Pacific Islander broad who is friends with Dwayne "The Rock" Johnson


This brave Scottish Lass:


The German girl who lives with seven dudes (kinda sus):

 

[Catgirls are Love]

Dreamlike Photoreal 2.0 seems to fix a lot of the anatomy problems that other models have. I tested it by making some Disney princess as they would look in real life, including:

How does it handle non-standing poses? There's quite a few models that can do standing poses very well, but if you get someone to sit down they still occasionally grow a third leg and let's not discuss what happens when they lie down.

 

[Stephanie Bustcakes]

How does it handle non-standing poses?

Still shit.

 

[The Mass Shooter Ron Soye]

Dreamlike Photoreal 2.0 seems to fix a lot of the anatomy problems that other models have. I tested it by making some Disney princess as they would look in real life, including:

AI seethers be like: I can't see the fingers or teeth!

 

[Alex Hogendorp]

Usually I don't like how this AI generates anime but sometimes there would be a cute artwork generated like this one. This one I liked so much I wanted to share it to the world.

 

[Monkey Pink]

View attachment 4414256
I made this and posted it on the Fediverse less than a week ago and in that time it's gone viral and spread and so far over the capital I Internet that it's made it's way back here and Kiwis have posted it in multiple image threads without even knowing where it came from. It's been interesting to see how it's been edited as it's been passed along: scaled up, saturation boosted, text altered, and translated to other languages. A guess then it fits the real definition of a "meme" then, eh?

I told you like a month ago that you were getting attention on /g/ and you were gonna get e-famous.
I always believed in you.

Can someone get me up-to-date about AI music? Has anything better than Riffusion and Jukebox came out?

 

[Girl Named Sandoz]

maybe this is common knowledge but i've never seen it mentioned: i just noticed that if you right click on the generate button in webui there's an option to generate forever. it will keep making pictures until you tell it to stop so you can leave it running overnight or whatever. pretty cool

 

[Catgirls are Love]

maybe this is common knowledge but i've never seen it mentioned: i just noticed that if you right click on the generate button in webui there's an option to generate forever. it will keep making pictures until you tell it to stop so you can leave it running overnight or whatever. pretty cool

Even though my system has adequate cooling, I think of that as the "cook meats on my GPU" button and thus never use it. I worry that should I generate very large volumes of images between reviewing sessions I'll spend even less time looking them over and picking out the interesting ones I really want to see more of in higher resolutions with different samplers.

 

[Girl Named Sandoz]

Even though my system has adequate cooling, I think of that as the "cook meats on my GPU" button and thus never use it. I worry that should I generate very large volumes of images between reviewing sessions I'll spend even less time looking them over and picking out the interesting ones I really want to see more of in higher resolutions with different samplers.

i've never really seen my gpu go over ~55 or so while generating so i assume it would be fine? it's an older gpu so it takes around a minute per image so for me this would be great

 

[AnOminous]

Even though my system has adequate cooling, I think of that as the "cook meats on my GPU" button and thus never use it. I worry that should I generate very large volumes of images between reviewing sessions I'll spend even less time looking them over and picking out the interesting ones I really want to see more of in higher resolutions with different samplers.

Modern video cards are usually smart enough not to set themselves on fire these days with a couple exceptions (*cough cough RTX 4090). Even if you use overclocking software like Afterburner, you can generally set a thermal cap to limit how much heat you get. It kind of has training wheels to prevent you from doing anything too stupid.

You really have to go out of your way to brick your machine even by overclocking, and if you don't do that, the default settings aren't going to let you mess things up.

 

[themasterlurker]

I told you like a month ago that you were getting attention on /g/ and you were gonna get e-famous.
I always believed in you.

Can someone get me up-to-date about AI music? Has anything better than Riffusion and Jukebox came out?

https://google-research.github.io/seanet/musiclm/examples/

We'll probably never get access to it but it's getting very good now

 

[Solid Hyrax]

I don't know where to put it, but just so you guys know, there are works on making a FOSS version of ChatGPT called OpenAssistant.

 

[Puff]

I'm training a hypernetwork for the first time.... I might be overplaying my hand on the graphics card... 2 seconds per iteration for at least 10k...
Anyone else done one of these? I'm following the voldy guide, but tips and tricks not covered there would be appreciated.
Edit: first question: It stops when it tries to make a sample image. No crashes, no errors, just stops and waits like it's done with it's task without any notification.

 

[themasterlurker]

I'm training a hypernetwork for the first time.... I might be overplaying my hand on the graphics card... 2 seconds per iteration for at least 10k...
Anyone else done one of these? I'm following the voldy guide, but tips and tricks not covered there would be appreciated.
Edit: first question: It stops when it tries to make a sample image. No crashes, no errors, just stops and waits like it's done with it's task without any notification.

What gpu do you have? You very likely don't have the computational power to do it, hypernetworks are said to require at least 4gb of video memory

 

[Puff]

What gpu do you have? You very likely don't have the computational power to do it, hypernetworks are said to require at least 4gb of video memory

1080ti, it's got 8gb I think... It IS training networks... I'm running the backups right now to see how it's coming along.
Either I've fucked up the settings or something is fucky with the preview feature. It gets to when it's supposed to output a preview and just stops.
The command line is left displaying the "loading weights" message, but you can carry on and generate images or start the training again, in which case it picks up right where it last saved (which is when I have it trying to generate an image). If you disable the preview generation it carries on like intended, saving backups at the specified step count.

 

[themasterlurker]

1080ti, it's got 8gb I think... It IS training networks... I'm running the backups right now to see how it's coming along.
Either I've fucked up the settings or something is fucky with the preview feature. It gets to when it's supposed to output a preview and just stops.
The command line is left displaying the "loading weights" message, but you can carry on and generate images or start the training again, in which case it picks up right where it last saved (which is when I have it trying to generate an image). If you disable the preview generation it carries on like intended, saving backups at the specified step count.

It's likely just shitting the bed, I've heard other people say the 1080ti will crash on SD