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

[anustart76]

I think there's a list of negative prompts to avoid things like that. Stuff like negative prompts for "extra limbs" or "extra fingers" or "long neck" or whatever is happening that you want to stop. It has to be a prompt that it recognizes, though, so check out a list.

I've tried that, it only works to some degree, and it's model-dependent too. Sometimes the model just does what it wants.

I wasn't very good at it either. If I get serious, I plan to check out sites like https://www.krea.ai/ and other compilations of those specific keywords certain models really like. Oh yeah, "trending on artstation" and "8k". Bullshit like that should not be needed in the long run, hopefully.

Making a crude drawing, collage, photo, or snapshot of a 3D model could help, for img2img. Why leave things to chance?

But allow me to rip off someone's prompt from Krea.ai, put it in Dezgo, and see how good it looks.

KREA — explore great prompts.

Explore millions of AI generated images and create collections of prompts. Featuring Stable Diffusion generations.

www.krea.ai

Epic Diffusion 1.1, Sneed: 3823911248, with a default list of negative prompts:
View attachment 4735616
It's not floating in outer space, but it looks better than the one I stole from and I only generated one image. I've never heard of Epic Diffusion. Do nothing for a few months, the AI gets twice as good.

Sounds like I should try out img2img and maybe some of these newer tools, I haven't heard of Epic Diffusion.

 

[Stephanie Bustcakes]

Stuff like negative prompts for "extra limbs" or "extra fingers" or "long neck" or whatever is happening that you want to stop. It has to be a prompt that it recognizes, though, so check out a list.

I've heard those are a placebo as well. The data the model is trained with has to have images of people with extra limbs or too many fingers or whatever negative tag you're using in order for it to work, and most sets of photos and drawings just aren't going to have stuff like that.

 

[Girl Named Sandoz]

I've heard those are a placebo as well. The data the model is trained with has to have images of people with extra limbs or too many fingers or whatever negative tag you're using in order for it to work, and most sets of photos and drawings just aren't going to have stuff like that.

note that if you're using an anime model this isn't the case because it was likely trained with booru tags which actually do include things like that

 

[Catgirls are Love]

Sounds like I should try out img2img and maybe some of these newer tools, I haven't heard of Epic Diffusion.

img2img is a very useful tool for getting images that are not dissimilar from the source material. It's also great fun.

 

[Epic Fail Man]

img2img is a very useful tool for getting images that are not dissimilar from the source material. It's also great fun.
View attachment 4743250

Do you mind posting a link to this huggingface space. I've been looking all over for this.

 

[Catgirls are Love]

Do you mind posting a link to this huggingface space. I've been looking all over for this.

That would do you absolutely no good because it's at http://localhost:7860/ and there's no way I'm turning on sharing to the wider internet.

 

[Epic Fail Man]

That would do you absolutely no good because it's at http://localhost:7860/ and there's no way I'm turning on sharing to the wider internet.
View attachment 4743366

Damn everything really did move to local. I guess I gotta clear out space on my computer.

 

[Tye Tass]

Got this from 4chan. Shame those threads are mostly filled with titties.

 

[Haint]

When I thought I had hit peak anime creepiness, it just gets worse.

 

[Irrational Exuberance]

Got this from 4chan. Shame those threads are mostly filled with titties.
View attachment 4745042

Does something about the meal disturb you, m'lady? I assume you're aware that only the finest of sofisticated gentlemen like myself has access to--wait, where are you going?! This is only an appetizer - we haven't even gotten to the main courses yet!

BITCH!

 

[Sulla]

Got this from 4chan. Shame those threads are mostly filled with titties.
View attachment 4745042

Gee thanks, I can't unsee obese null.

 

[Pocket Dragoon]

 

[Puff]

I've heard those are a placebo as well. The data the model is trained with has to have images of people with extra limbs or too many fingers or whatever negative tag you're using in order for it to work, and most sets of photos and drawings just aren't going to have stuff like that.

Spoiler: Illustration using Anything V3 and the generic Voldy suggested prompts, same seed

25 steps 2374723252 seed
Positive:Farmer holding a pitchfork, American Gothic
(no negative prompts)

Positive:Farmer holding a pitchfork, American Gothic
Negative: bad anatomy, bad hands, missing fingers, extra digit, fewer digits

Positive:Farmer holding a pitchfork, American Gothic
Negative: 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, artist name

Positive:Farmer holding a pitchfork, American Gothic, masterpiece, best quality, masterpiece
Negative: 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, artist name


For fun, swapping the positive and negative quality modifiers
Positive: Farmer holding a pitchfork, American Gothic, 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, artist name
Negative: masterpiece, best quality, masterpiece


A different modifier set
Positive: Farmer holding a pitchfork, American Gothic, 8 k ultra realistic, detailed, intricate, full of color, led lighting, trending on artstation, photo,
Negative: 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, artist name, wide shot,


And Flipped

The quality modifiers probably don't do what's on the tin, but they clearly have a positive effect.

 

[Slav Power]

Can someone give me a quick rundown of how much changed in the past 5 months? I haven't touched Stable Diffusion since then. Anything better than SD 1.4? Some major changes in the WebUI project where I'll need to make a backup of the old one? Any specific prompts?

 

[Catgirls are Love]

Can someone give me a quick rundown of how much changed in the past 5 months? I haven't touched Stable Diffusion since then. Anything better than SD 1.4? Some major changes in the WebUI project where I'll need to make a backup of the old one? Any specific prompts?

There's an absolute shitload of new models available since absolutely everyone and their dog is making models and distributing them. Most are shit or do only one thing very well, but it's a question of how much time you want to spend downloading and experimenting. Most are also for porn.

You no longer have to build xformers yourself and can just download it for better speed and memory usage.

Hi-res fix was reworked, again.

Loras. I don't use them.

If you update, create a new installation instead of updating existing just in case shit absolutely breaks.

 

[Puff]

Can someone give me a quick rundown of how much changed in the past 5 months? I haven't touched Stable Diffusion since then. Anything better than SD 1.4? Some major changes in the WebUI project where I'll need to make a backup of the old one? Any specific prompts?

Adding on to the previous poster, there's a new format for checkpoints called "safetensors" which can't execute arbitrary code. Try to use them.
Loras always act like hypernetworks to me.... I think they're easier to make or something

 

[ducktales4gameboy]

I think there's a list of negative prompts to avoid things like that. Stuff like negative prompts for "extra limbs" or "extra fingers" or "long neck" or whatever is happening that you want to stop. It has to be a prompt that it recognizes, though, so check out a list.

The easiest way to deal with this is going to civitai and finding one of the bad anatomy textual inversions. TIs are basically horrendously long prompts compressed to single words, so you can just drop 'badanatomy' or whatever the trigger is in your negative prompt and get good results.

Also note you can put any TI you want in negative if you want the network to avoid it. It's great for getting rid of gross anime stuff.

eta: May as well plug my favorite TI. Great fun for making anything you want into a pulp scifi cover.

 

[AmpleApricots]

Loras are good at getting consistent styles and motifs out of SD. They don't need to be trained on one very specific motif only. You also don't have to apply them with full strength, all the time. They're good enough if applied with a bit of strength, to just pull SD into certain directions. I prefer a drawn style (that is not anime) to photorealism, and loras helped there. Lately I've been in a robot phase and I don't think my robots should have faces. It was impossible to convince SD via text prompting to draw humanoids that are faceless otherwise, Loras helped there too.



This one might as well be some kind of scifi enviromental suit:


I'm also experimenting with a more mechanical, less sleek style but it's hard to get it down without "logical" errors, where the design doesn't make sense:



I like to imagine it's a bunch of robots just endlessly wandering the earth long after we're gone. I generate them while listening to music like this. My wife thinks I've gone off the deep end. I might have.

 

[Irrational Exuberance]

Loras are good at getting consistent styles and motifs out of SD. They don't need to be trained on one very specific motif only. You also don't have to apply them with full strength, all the time. They're good enough if applied with a bit of strength, to just pull SD into certain directions. I prefer a drawn style (that is not anime) to photorealism, and loras helped there. Lately I've been in a robot phase and I don't think my robots should have faces. It was impossible to convince SD via text prompting to draw humanoids that are faceless otherwise, Loras helped there too.

View attachment 4760094View attachment 4760116

This one might as well be some kind of scifi enviromental suit:
View attachment 4760226

I'm also experimenting with a more mechanical, less sleek style but it's hard to get it down without "logical" errors, where the design doesn't make sense:

View attachment 4760253

I like to imagine it's a bunch of robots just endlessly wandering the earth long after we're gone. I generate them while listening to music like this. My wife thinks I've gone off the deep end. I might have.

That last one looks like a female member of Daft Punk. Maybe the robots found some preserved instruments and formed a band?

 

[The Tall Man]

This silly program keeps removing the space car in my retrofuturism poster.

This stuff is infuriating, I can't ever get it to do what I really want.
Why does it remove the hovercar in the image to image I even made it a Jewish one with triple parenthesis?

I give up, back to using paint.