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

[bifftango]

Still getting some good nonsense out of this.

Spoiler: Boobs ruined

"Nuclear holocaust in Boise on Valentine's Day 1993, bright, vivid, horrific"


"Cenobites bikini carwash, sunny, colorful, party time"


"The chatterer reading fairy tales at the Duluth YMCA, somber, contrast"


"Invalid elected to the senate, horrifying, depth of field, contrast"


"Angel divebombing a picnic in Central Park in 1957, grainy, contrast, black and white"


"Uschi Digard being a prude at a pool party, Kodachrome, vivid, retro"


"Catwoman working the night shift at Orange Julius in 1985"

 

[SomeBirds]

My wife has been running prompts and has been using it more than I have and essentially, stable diffusion is her realm now. I only exist to update, merge and troubleshoot.
I'll spare you guys the story of her 40K+ renders of everything that pops into her head.

Has anyone mastered clothing/dressing two separate characters without in-painting corrupting the fuck out of it?

View attachment 3846015

The objective is to create a female knight holding a princess, and gently kissing each other. This she has managed to do.

The issue is the skin color, clothing and height. Whenever one of those is and I quote "JUST FUCKING PERFECT WHY CAN'T IT DO THE REST RIGHT", the other is absolutely wrong. Even when negative prompting men, it randomly turns one of the women into a feminine dude which absolutely rustles her jimmies. Or gives them both dresses or it gives them both armor.
The Princess is supposed to be fair skinned/ tanned/ lighter than chocolate black, short with wavy hair.
The Knight is supposed to be Pale and tall, with red wavy hair and wearing knights armor.
View attachment 3845910View attachment 3845967View attachment 3845988


In experimenting on how to do this, I've tried the novel Ai specification route and have gotten nowhere.

View attachment 3846009.

Anyone can figure out how to specify exactly what to change?

This seemed like an interesting challenge, but I actually thought you were trying to get a male/female pairing and it was giving you lesbians. Here's some results I got testing your prompt idea. The first batch are m/f, but there's a bunch of f/f. None of them are technically perfect, but they're weird in ways that are normal to AI art. These are all original gen images. I didn't run variations although one or two might have gotten the 'enhance' treatment (which I recommend if you are trying to coax out minor changes).

Prompt: knight holding a princess and tenderly kissing her, realistic, detailed, digital painting, fantasy, trending on artstation



This one was just funny 'cause of the random cat. Where did it come from? IDK.



The f/f were a collection of prompt tests based off the above prompt. I only changed the gender stuff at the beginning. 'Woman in armor' 'holding a princess, etc' seemed to get the best results.



And finally, here's one using an image your wife generated as the seed. Noise .5, strength .25


I might revisit this later as we have plans to train our AI on images of people kissing. It's currently being trained on human faces, but the training process is... involved. So one thing at a time. I hope this helped?

 

[A Hot Potato]

Doing some bullshit with places instead of Waifu's now.

This is NovelAi so it's probably not quite what this is meant for, but eh.

None of this is meant to be amazing, I was just trying some stuff out..

 

[Catgirls are Love]

Doing some bullshit with places instead of Waifu's now.

This is NovelAi so it's probably not quite what this is meant for, but eh.

None of this is meant to be amazing, I was just trying some stuff out..

I mostly use NovelAi's model and I'm usually generating landscapes or buildings. It's not bad at that sort of thing so long as long as you don't mind it being a little wobbily with lines and perspective.

Spoiler: A few from the first days of experimenting

 

[Baraadmirer]

I mostly use NovelAi's model and I'm usually generating landscapes or buildings. It's not bad at that sort of thing so long as long as you don't mind it being a little wobbily with lines and perspective.

Spoiler: A few from the first days of experimenting

View attachment 3858882View attachment 3858888View attachment 3858891View attachment 3858897View attachment 3858900View attachment 3858903View attachment 3858915

It definitely feels more picturesque than Stable Diffusion, which seems to go for photorealism more by default.

 

[The Mass Shooter Ron Soye]

Doing some bullshit with places instead of Waifu's now.

This is NovelAi so it's probably not quite what this is meant for, but eh.

None of this is meant to be amazing, I was just trying some stuff out..

I mostly use NovelAi's model and I'm usually generating landscapes or buildings. It's not bad at that sort of thing so long as long as you don't mind it being a little wobbily with lines and perspective.

The perspective on those shots make them look like visual novel screenshots without the waifu. As if they are intended to have characters overlaid in the foreground.

 

[Catgirls are Love]

The perspective on those shots make them look like visual novel screenshots without the waifu. As if they are intended to have characters overlaid in the foreground.

Often, yes. I tend to put something like "human, people, girl" in the negative prompt for landscapes/cityscapes otherwise I occasionally get images with some random girl centered in the frame which wasn't what I was aiming for.

Right now I'm messing around with the realistic/photorealistic prompts to see what NovelAI puts out if I don't want waifus in an anime style. First impressions are that it does a better job with men with those prompts, but maybe that was because I put in "lord of the rings" just to see what it'd output. Much more experimenting required.

 

[Baraadmirer]

Has anyone had issues with some of their base images for image prompts churning out pitch-black results? I tried tweaking all my settings but nothing seems to work. I've noticed it on NovelAI, though it might also extend to Stable Diffusion.

I also found Lexica today, which serves as a repository for AI-generated images and what text prompt, seed, and CFG scale were used to get it, so if you find an art style that you particularly like, you can find out if the user referenced an artist in the prompt.

 

[Soda Can]

I wonder what hellscape would be created if you trained a model on livejournal or tumblr. Or even specific fanfiction communities. Hell at that point the guy still writing that SSB:B fanfic could train a model on just his single work alone and have it forever spit out new chapters.

This already happens. I know one guy who created a model on just his own work going all the way back to highschool to act as a kind of proofreader and it comes up with twists on ideas he already had. And he got the idea to do this from a discord community of nothing but writers who do the same. There are porn writers on DA, FA, FF.net and a couple other places who have admitted to doing this as well. Hell, I tried it for a thing I'm writing and it's helped me sort out a few pacing issues.

 

[Catgirls are Love]

Has anyone had issues with some of their base images for image prompts churning out pitch-black results? I tried tweaking all my settings but nothing seems to work. I've noticed it on NovelAI, though it might also extend to Stable Diffusion.

I also found Lexica today, which serves as a repository for AI-generated images and what text prompt, seed, and CFG scale were used to get it, so if you find an art style that you particularly like, you can find out if the user referenced an artist in the prompt.

That happens sometimes for me too, and I feel like it occurs more often when using a horizontal aspect ratio though that may be entirely unrelated. I haven't really done any sort of troubleshooting as it occurs roughly once per twenty batches, so I just shrug and assume an algorithm fucked up somewhere. If you find one of the prompts/seeds where it does that and get it to output results step by step, might be able to see exactly when it's screwing up. Not sure if it's something you'd be able to fix, but at least you might be able to see what's going wrong.

 

[Potatoherder]

Has anyone had issues with some of their base images for image prompts churning out pitch-black results? I tried tweaking all my settings but nothing seems to work. I've noticed it on NovelAI, though it might also extend to Stable Diffusion.

From what i have heard this is caused by running in fp16 mode instead fp32. The automatic webui defaults all loaded models to fp16 mode and you need to pass "--no-half" and"--no-half-vae" to disable. The downside is it'll greatly increase the vram needed. If you are using a paid service it's likely to be a nsfw filter. I personally have this happen so infrequently it's not a concern at all running in fp16 mode.

 

[Baraadmirer]

From what i have heard this is caused by running in fp16 mode instead fp32. The automatic webui defaults all loaded models to fp16 mode and you need to pass "--no-half" and"--no-half-vae" to disable. The downside is it'll greatly increase the vram needed. If you are using a paid service it's likely to be a nsfw filter. I personally have this happen so infrequently it's not a concern at all running in fp16 mode.

Hmm, interesting. Sometimes it goes away when I readjust the dimensions, but other times it seems specific seeds are impacted but others aren't when I generate a batch of them.

Any chance of getting some resources added to the OP? This page ( Archive ) looks like a good primer to anyone who's new to the software. I learned some things from it, like using style and artists in the prompt.

 

[Involuntary Celebrity]

Spoiler

 

[Catgirls are Love]

I decided that my laptop could use a new desktop background.

It was a learning experience and while I could work on that eye a bit more I figure it's good enough for something I'm really only going to look at when logging in. Definitely had quite a bit of fun experimenting and I have a better idea of what to do next time.

I'm still experiencing black outputs with certain prompts and/or higher resolutions with the arguments mentioned above, so unfortunately that didn't seem to help much for me.

 

[ToroidalBoat]

DeviantArt, forgetting who their entire userbase is, has announced their own AI art tool DreamUp.

So the dA userbase is still full of rabid pro-copyright fanatics who REEE at "art theft"?

(I remember when it was that way in the '00s and early '10s.)

 

[Neurotypical Mantis]

here's a request that i think would be cool:

a backlit cloud in the sky in the shape of a Renaissance angel

tried with dall-e2 and dawn ai and didn't get much. i want a cloud that looks like an angel dammit.

 

[Starved Artist]

I added the porn packages to my custom package.
Not safe for existence.

This seemed like an interesting challenge, but I actually thought you were trying to get a male/female pairing and it was giving you lesbians. Here's some results I got testing your prompt idea. The first batch are m/f, but there's a bunch of f/f. None of them are technically perfect, but they're weird in ways that are normal to AI art. These are all original gen images. I didn't run variations although one or two might have gotten the 'enhance' treatment (which I recommend if you are trying to coax out minor changes).

Prompt: knight holding a princess and tenderly kissing her, realistic, detailed, digital painting, fantasy, trending on artstation

View attachment 3856356View attachment 3856365

This one was just funny 'cause of the random cat. Where did it come from? IDK.

View attachment 3856377

The f/f were a collection of prompt tests based off the above prompt. I only changed the gender stuff at the beginning. 'Woman in armor' 'holding a princess, etc' seemed to get the best results.

View attachment 3856395View attachment 3856404View attachment 3856410View attachment 3856416

And finally, here's one using an image your wife generated as the seed. Noise .5, strength .25
View attachment 3856419

I might revisit this later as we have plans to train our AI on images of people kissing. It's currently being trained on human faces, but the training process is... involved. So one thing at a time. I hope this helped?

She thanks you a lot for this. Really appreciates this .She's been trying to pull out mixed prompts, while I have been prompting every weird idea I've had for the last couple of days.
She wanted to know, that if you're training a model, would you release it and would it have pants suited women in it?

----

Yesterday, I finally seized control of Stable Diffusion and I had no idea what to prompt, so I started googling random shit and came across the top 30 ugliest celebrity list.

HEY MODS - What's the rule on models and soft-core boudoir shitposting? This isn't hardcore stuff,


Removed: Awaiting admin response before ruining your day.

Notes.
Jack Black is fucking Randy Marsh.
.

Charlyne Yi causes it to flip out. I couldn't delete these faster. Half the generated images couldn't decide if she was a boy or a girl.


There is no Young Tilda Swinton. There has only been old Tilda Swinton

.

Out of respect for Miss Hunt, there will be no Bikini Pictures. But Slav Squatting is on the Menu.


Last but not least.

Adolph Hitler playing Yugioh brings joy to my blackened heart. I haven't played Yugioh in 12 fucking years and this is the EXACT face I made when my nephews hit me with that syncrobullshit this weekend.

 

[DeadwastePrime]

still waiting for some dude to train a model off of guns and people holding guns. i say this as the ai is not very good at generating them and i wanna make cool character concept arts with it

 

[bifftango]

still waiting for some dude to train a model off of guns and people holding guns. i say this as the ai is not very good at generating them and i wanna make cool character concept arts with it

What's the problem? Everything looks perfectly fine!

 

[Woe-B-Gon (TM)]

What's the problem? Everything looks perfectly fine!
View attachment 3875811

The thing can't make a person eat food but give it a gun and it goes right in the mouth. (Though it's more like halfway between eating and playing you the song of his people.)