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

[Sage In All Fields]

I'm not familiar with the circumstances of the algorithmically produced music but for this AI image shit its possible the same copyright trigger of pressing the generate button along with the prompts and settings etc would be considered enough to create a copyright in the work. Maybe.

If one good thing comes out of this, I hope that it's copyright law finally collapsing in on itself and dying in a fire.

 

[Near]

At last, my GPU upgrade is complete.

 

[Helmut A Ebne]

I'm sorry

 

[lolcow yoghurt]

Wake me up when digital artcels try to grift commissions by saying "AI is trying to replace us, help a {POC/LGBTQ+} digital artist by buying commissions"

 

[Lord Xenu]

This poor guy has too many feet.

I don't even know what to say about this one

Here's a passable Taylor Swift.

This would be my favorite abomination so far. Downtown Toronto in the style of a "Group of Seven" painting.

Does anyone know of a FAQ for the prompts and different settings/general concepts?

 

[FILTH Tourist]

Does anyone know of a FAQ for the prompts and different settings/general concepts?

From the looks of it I am going to assume you are using NAI?
Write your prompt with tags instead of prose, don't be afraid to get really specific.

Spoiler: 1girl, sweater, snow, park, scarf, smile, closed eyes, blush,

Hope you guys are staying cozy BTW

If you are looking for tags, NAI's dataset is ripped from danbooru so their tags will work well with NAI.

also if you want to emphasize tags use {} like {orange hair} or deemphasize it with () like (snow) if you find one tag overpowering the others.

That should be the basics for NAI at least.

 

[Dammit Mandrake!]

I got it all downloaded and working.


"What does the AI dream?"

I think we've gone too far.

 

[Glowie]

View attachment 3751942
I'm sorry

I escaped from bad dragon.jpg

 

[Involuntary Celebrity]

also if you want to emphasize tags use {} like {orange hair} or deemphasize it with () like (snow) if you find one tag overpowering the others.

FYI that's just novelAI's UI implementation, if you're using the NAI model (or whatever) with webUI it's () to add emphasis and [] to reduce it.
You can also ((((stack)))) em or add specific weighting like (this:1.4) for even more precision. In webUI you can just highlight a tag and hit ctrl + up or down to tweak them, too.
But order still matters, so an unbracketed token as your first term might have more impact than a bracketed one at the end. Although putting them at the end can still be a better idea in the case of long prompts, because as I mentioned before anything after 75 tokens gets appended with an AND so you want to try to have nonspecific style tags (like lighting, backgrounds or artists--possibly, depends what you're trying to do) at the end and use brackets rather than bumping them up if it means something more important would end up after the 75 soft limit.

Also while booru-style tags definitely works well, I find that grouping them in short clauses can be very helpful if you need to keep terms related. For example "scared (confused) girl with (long braided black hair) sitting on bed, boy standing (offering bowl of eggs)" followed by all your other assorted tags will probably work better than "1boy, 1girl, bed, bowl of eggs" etc for obvious reasons. So don't be afraid to mix and match.

Edit: also if you see {curly|braces|with|piped|terms} in people's prompts that's a random prompt selector. There's a couple scripts but I use the dynamic prompts one, which gives you that plus picking random terms from __wildcard__ files, and adds a pick N syntax which lets you do stuff like {2$$one|two|three|four} to pick two random tags from your list.

...And whatever, while I'm at it there's also [::] syntax which lets you remove or swap a tag partway though processing your steps, like [foo:bar:10] which would swap foo for bar after 10 steps or [:banana:0.5] which would make bananas kick in at the 50% mark.

 

[A Humble Moose]

I now never need to go hunting for good desktop wallpapers anymore. I can just generate them to be more or less what I want. For example, my new wallpaper:



AI seems to be really good at making landscapes, as many people have already said. The oil painting prompt especially helps obscure some of the odd looking details that the AI seems to generate.

 

[FILTH Tourist]

Also while booru-style tags definitely works well, I find that grouping them in short clauses can be very helpful if you need to keep terms related. For example "scared (confused) girl with (long braided black hair) sitting on bed, boy standing (offering bowl of eggs)" followed by all your other assorted tags will probably work better than "1boy, 1girl, bed, bowl of eggs" etc for obvious reasons. So don't be afraid to mix and match.

I didn't try the clause thing, I am going to play around with it more. Thanks for letting me know! I am still playing with it but now hopefully I can specify two or more custom characters in an image.

 

[Involuntary Celebrity]

It can also be useful for avoiding redundancies. For example a lot of people write stuff like "detailed bodypart, gaping bodypart, bodypart dripping" when "detailed dripping gaping bodypart" will probably save you tokens and avoid (occasionally) generating a person possessing multiple (bodypart)s when you only intended them to have one, or just wanted to be descriptive and not stress it 3 times.
Still defer to booru terminology for most things but when you can combine them in a natural English way it usually seems to work more predictably.

Thanks for letting me know! I am still playing with it but now hopefully I can specify two or more custom characters in an image.

If you really need to separate them it might be a use case for a deliberate AND. Things on each side of an AND won't process on the same step so you can keep terms (ie characters and everything relating to them only) totally separate this way. You can have multiple ANDs as well, for example if you need to also separate background etc tags... I'd try keeping em on character or another first though because every AND is diluting your step count, since you're basically alternating.

 

[Baraadmirer]

img2img is rad as hell

View attachment 3751838

I haven't seen this posted anywhere so far, but you'll want to periodically clear out your output folder. It can be found by going to: {wherever you set up the program}\stable-diffusion-webui\outputs. Everything you render gets dumped into a folder in there. I was running several batches to find the right picture and added 10+ gbs of images over a relatively short amount of time.

You can actually prevent that in Settings → Saving images/grids → Always save all generated images and Always save all generated image grids. From then on you just have to manually save what you want.

 

[Weeb_Killer]

This ain't the end to art at all. If anything, it'll make traditional art that can be verified to have been rendered by a human hand even more lucrative.

 

[Hokuriku]

Wake me up when digital artcels try to grift commissions by saying "AI is trying to replace us, help a {POC/LGBTQ+} digital artist by buying commissions"

That's definately already a thing on the commission artist-verse.
Digital artists have been grifting commissions long before this anyway with the whole muh-living-wage thing and celebrating one another jacking up their prices.

 

[EatYourVeggies]

The best pics I've generated:

"Hidden citadel in a dark forest"

"Lucifer"

"Alien colony invading planet Earth"

"Obese overweight Hitler"

"Ogre chess piece"


"The perfect woman"

 

[Takodachi]

Not sure if this was posted, but someone is trying to make AI music now, and even using Stable Diffusion to generate MVs


Tweet
Archive

On Youtube

 

[Rungle]

One day AI will S.H.O.D.A.N us all and its gonna be all the horny peoples fault.

 

[NoReturn]

From the original we can also get a serious-looking Keffals:

"When you meet the man you could have become."

Something I find curious is how everybody cries about artists being replaced by AI, but nobody seems to care about writers being replaced by AI storytellers. Or someone taking an AI story and doing a bit of editing and releasing it as their own book.

Hasn't this already happened? I remember hearing something about this where a user had already self-published a novel that was co-written by NovelAI (or was it AI Dungeon?)

NGL the ones that actually worry me are the ones that can code a website from just a description, that alone could leave a lot of people out of work.

Yeah we're heading there right quick.
I want to find one that can compose a song from a hummed tune. I am musically-retarded but I want to hear some songs again that have been lost over the years.

Not sure if this was posted, but someone is trying to make AI music now, and even using Stable Diffusion to generate MVs
View attachment 3752493

HA! Very next page.
This is close, but I wonder how we can do "paint ins" for notes.

 

[elastic eye]

At last, my GPU upgrade is complete.
View attachment 3751926

Dang, I was happy with my 2070 Super hitting 6.2-6.4it/s.

Been having a blast generating random shit, onto horror Halloween themed things now. None of that gay anime shit