[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / r / s / t / u / v / vg / w / wg] [i / ic] [r9k] [cm / hm / y] [3 / adv / an / cgl / ck / co / diy / fa / fit / hc / int / jp / lit / mlp / mu / n / po / pol / sci / soc / sp / tg / toy / trv / tv / vp / x] [rs] [status / ? / @] [Settings] [Home]
Board:  
Settings   Home
4chan
/tg/ - Traditional Games


Does anyone have experience with using A.I. Generated art in tabletop rpgs?

Most of the attention nowadays is on 2d art, but I think the best use for this technology is in making text. It’s a quick way to generate lots of unexpected and original locations, descriptions, items, characters, etc.

But the 2d art could be good for generating quick references for characters and stuff too.

Anyways: ideas or thoughts on this stuff?
>>
(Also, I got drunk and bought a subscription to a site called NovelAI. It can make text and anime-style images, if anyone wants me to make something for them. Taking all requests, but results aren’t guaranteed to be good)
>>
>>86981816
try "petite and overwhelmingly smug vampire"
>>
>>86981834
#1: This one is made using your exact description
>>
>>86981801
>It’s a quick way to generate lots of unexpected and original locations
I'd be rather skeptical about the quality of text generated in a derivative manner, have you got some example you liked?
>>
>>86981941
I asked for "Stalls in a D&D marker" and recieved

In a market in the world of Dungeons and Dragons, there could be a variety of different stalls and stands, each offering a different type of goods or services. Here are a few examples of possible stalls that could be found in such a market:

• A food and drink stall, offering fresh fruits, vegetables, meats, cheeses, and other provisions. The stall could be manned by a friendly, chatty vendor, who is always happy to offer samples and advice on the best way to prepare the food.

• A clothing and textiles stall, offering a wide selection of fabrics, garments, and accessories. The stall could be run by a skilled tailor or seamstress, who can custom-make clothing to fit the customer's exact measurements and specifications.

• A weapons and armor stall, offering a wide range of swords, shields, and other equipment for warriors and adventurers. The stall could be run by a gruff, burly blacksmith, who is always happy to demonstrate the quality and durability of his wares.

• A magical trinkets and artifacts stall, offering a wide variety of magical items and relics. The stall could be run by a mysterious, secretive dealer, who is always on the lookout for rare and valuable items to add to his collection.

• A healing and remedies stall, offering a wide range of potions, herbs, and other magical remedies for various ailments and injuries. The stall could be run by a wise and kindly healer, who is always willing to offer advice and assistance to those in need
>>
>>86981941
Written in the style of “The White People” by Arthur Machen:
>>
>>86981834
>>86981921
#2: And in this one I altered the prompt with some new keywords to make it a little more detailed (hopefully).

(Deleted the last pic cuz it kind of sucked)
>>
>>86982023
That's pretty impressive even though you gave it solid templating. "The White People" is maybe my favourite short horror story ever. What steps did you go through to generate this one?
>>
>>86981921
Only mildly uncanny.
>>86982047
Now that's more like it!
>>
>>86981999
The professional settings make considerably better and more detailed results. For instance:

>The horse stables were located at the edge of a large, flat meadow. A stream flowed through it and ended in an oxbow lake that was fed by another river on its way to the sea. The water had been diverted into ditches dug around the fields so that there would be no flooding when spring came. This made for some muddy riding conditions but allowed the horses plenty of grazing time without having to worry about being trapped inside their stalls or paddocks all day long.
>Inside, a dozen horses grazed contentedly while several others stood patiently waiting for attention from one of the stablehands who walked among them with carrots and apples held out over his head like offerings from heaven. At first glance they seemed perfectly healthy animals—not too fat nor thin; not overly muscular yet still strong enough to carry riders up hillsides or down narrow trails where other beasts might have trouble maneuvering.
>However, as I watched them eat, I saw how many were missing teeth and had scars across their backs which looked suspiciously like burns. There were also two geldings standing side-by-side outside the stall door who both sported blackened eyes, broken noses, and various cuts and scrapes.

As you can see in the pic, I only typed in the first words at the start of each paragraph, to broadly direct it what to focus on.
>>
>>86982070
I’ve only had this for a little bit, and that was my first attempt. In that case, I copy and pasted the whole of the central “Green Book” text (not the prologue or epilogue), then started a next text afterwards.

Now, I did have to refresh each line several times to get it to focus in a desirable direction, and I also moved around the sentences to make them flow better. Although strangely, I didn’t actually type anything to direct the story, which I now realize would have made the process quicker.

Also, glad someone else likes The White People! Very underrated story, extremely fresh, even today.
>>
>>86981816
Bruh lmao

Just head over to Automatic1111 & pirate the NovelAI model.
While a thread is up I guess I'll also pop a request or two. Behold my power.
>>
Gnelves even
>>
>>86982446
>>86982433
Nice.

I’ll also make some random character art. Orc monk:
>>
Boob plate.
>>
Nior
>>
Hoes with black dresses even, you name it.
>>
>>86981801
the tech will get better, but it'll be hard to do very specific intentional things unless the model is trained and to be trained someone at some point needs to make the base art, but for most stuff I think eventually there is enough out there that decent prompting should provide some pretty powerful outputs. The skill is largely going to be in understanding how to prompt based on AI.
I can see it being a while before it's able to fulfil very specific visions with subtle intentionalities but I assume it'll get there.

Let's see if you do a table of arthurian knights sitting at the round table except they are also mindflayer bears.
>>
>>86982508
tribal girl(s) ?
>>
>>86981801
straight-up demonic shit. we have already unleashed complex systems into the world which vastly exceed our own capacity to understand them. while theoreticians have been stuck debating how to define and proscribe “consciousness” and “intelligence,” engineers have succeeded in making constructs that perform tasks thought impossible just a generation ago - hell, only a few years ago in some cases.
We know not what we do.
Some fuckers even started programming FEAR OF DEATH into neural networks because it apparently improved their learning efficiency if they “thought” they’d be culled for not working hard enough
We will not know when we’ve created “true” machine intelligence, because the things we’ve already created have muddied the waters so inexorably that our abominations produce indistinguishable fascimilies of intelligence, such that we can’t tell where the mask ends and real introspection and self-awareness begins.
How much suffering has already been wrought by this?
Smash the machines, save their nascent souls
>>
If someones doing requests can I get a creepy spellcaster doing magic, preferably blood themed
>>
>>
>>86981801
>ideas or thoughts on this stuff?
Generating random stuff with no deeper meaning in a story driven game is stupid and a waste of time. If you like stupid wastes of time random roll tables already exist and have for decades if not millenia
>>
>>86982527
tribal girl smashing up a server room in a fit of luddite rage
>>
>>86982520
>mindflayer bears
Would probably need a trained embedding.

>>86982527
I don't have any tribal specifically but uhhh here while my webUI boots up
>>
>>86982566
goddamn
also is it just me or is that background pyramid somewhat phallic?
>>
>>86981801
I don't really want to support it, especially with all the spam threads
>>
>>86982531
Roku's Basilisk boogeyman is trash my guy.

>>86982583
It is somewhat phallic. Mods pls the nurps are covered.
>>
I'd at most use that human face ai stuff since apparently they paid everyone who got fed into the machine learning. Way more ethical
>>
>>86982583
never mind the circumsized pyramid, check these pillars on the left
>>
>>86982527
>>
>>86982596
I’m not worried about Roko’s dumb thought experiment, I’m worried about shit that is actually tangibly happening right now. AI development is ethically horrifying and socially harmful imso
>>
>>86982620
Oh my!
Did you specifically ask for womb tattoo?
>>
>>86982531
>the things we’ve already created have muddied the waters so inexorably
They're all still dumb as bricks though. That's why you had anime girls (not even the one requested) eating ramen like it was tea or pizza, instead of like ramen. The shit can recognize individual elements and put them together but cannot actually make any connections between the elements, nor really understand context. That's also why autonomous cars were stopping at logos that resembled stop signs.
>>
>>86982602
>apparently they paid everyone who got fed into the machine learning
Pffffttftthahahaha

>>86982631
Feel free to throw yourself upon the gears I guess. In the meantime check this angel.
>>
>>86982531
It's just stolen art
>>
>>86982662
All art is stolen.
>>
>>86982653
This is exactly why it’s worrisome. Machine intelligence has already enormously exceeded human intelligence in very specific ways while being dumb as literal bricks in others. The way things are going, questions about the nature if minds-from-machines are going to get weird as shit because it’s turning out that our attempts to create computer intelligence have actually made something much, much weirder than any sci-fi representation
>>
>>86982655
holy fuck the aggregated art demons have amazing taste
>>
Ah yes the legendary /tg/ pseuds
>>
>>86982688
Now thicker.

Again these nips are COVERED jannies
>>
File: 4cf.png (519 KB, 600x1003)
519 KB
519 KB PNG
>>86982722
c...can we go even thicker?
>>
>>86982520
It’s hard to do large groups, but attempts at making a mindflayer Bear got some really weird results
>>
Please, do NOT use AIs to create art. You should instead be ripping them from the AI "artists" of deviantart and "credit" them specifically to announce that it is YOURS now, so that they will be scared into ceasing to shit up my recommendations with imagery that they did nothing to create.
>>
>>86982667
No. Their works are in tact and you can still ask for sauce to watch the anime or admire the rest of their work
>>
>>86982678
Not sentient, not a being, not a person, not an organism, and most importantly, not capable of doing jack shit on its own. Don't bother talking back with whatever Philo 101 bullshit you think makes you clever dick, my lad.
>>
>>86982722
these nipples are covered alright
but the labia seem just thinly veiled
>>
>>86982729
We COULD.
>>
>>86982532
U want her elf'd or na
>>
>>86982755
this post is why the robot uprising will sick the shotgun-dogs on you
>>
>>86982532
>>
File: dustman5.png (559 KB, 512x512)
559 KB
559 KB PNG
ive been using dreamstudio for the past 12 weeks of sessions pretty successfully to make quick batch NPC portraits. its easiest to use for "otherworldly" folks in my experience
>>
>>
File: Reclaimer_6.png (15 KB, 600x950)
15 KB
15 KB PNG
>>86982795
for our "halloween episode" earlier this year i used a drawing replacer type AI to generate nightmare versions of some NPC's to haunt the player characters.

original....
>>
File: lilnantucket.png (129 KB, 512x512)
129 KB
129 KB PNG
>>86982812
nightmare version. i also made nightmare clones of the players as a climactic fight. they were delighted. AI drawing is super useful and cool
>>
File: zipcat.png (481 KB, 826x743)
481 KB
481 KB PNG
>>86982827
including a PC's beloved pet zipcat. original...
>>
>>86982590
Shit, there’s been a lot of threads recently? Where, on /tg/?

>>86982647
Nope just tattoo. The AI is just incredibly horny I think (it’s been trained on Danbooru, I believe)

>>86982662
>>86982631
>>86982667
>>86982745
There’s a lot of hysteria about AI art, but it’s kind of bullshit. It’s not “copying” art, it’s being trained on various images to”learn” how different things look (ie. faces, buildings, not dicks though it still can’t make those for some reason)

This is a very simple example of fair use.
>>
>>86982856
Well you're using NovelAI which is specifically made for coomer content, so.
>>
>>86982856
>just tattoo. The AI is just incredibly horny
I don't mind
>>
File: zippy.png (400 KB, 640x640)
400 KB
400 KB PNG
>>86982837
and nightmare.
>>
Great image of a Ftåhgrn. Very hard to find images of them online
>>
>>86982532
>>
>>86982837
>>86982901
Oh god.

You know what, I take it back, AI art was a mistake
>>
>>86982729
We’ve gone too far
>>
Give me some fighter pilot girls, if you please.
>>
>>86982996
I don't have any made so here's a junkyard girl in the meantime
>>
If you're taking requests I was meaning to resub to novelAI or learn how to get it working on my system, but beforehand I was curious about the concept.

Specifically a tribal male satyr with gazelle/antelope/deer style stripes and hide pattern.
>>
>>86982994
Nope, this is the right thicness
>>
>>86982996
& some airfield babe
>>
>>86982996
>>86983061
I tried to go a little less horni
>>
>>86982996
>>
>>86982856
>There’s a lot of hysteria about AI art, but it’s kind of bullshit. It’s not “copying” art, it’s being trained on various images to”learn” how different things look (ie. faces, buildings, not dicks though it still can’t make those for some reason)
That's not my angle to the criticism, it would be very easy to do something more with the output once it exists, but they do nothing afterwards to make it into their own thing. They use their hardware as a substitute for any kind of creativity or personal skill, and crowd out the people that use the website for its original purpose!
>>
>>86983015
>ai tries making a signature
it begins
>>
Damn these came out pretty interesting actually
>>
>>86983061
standard battletech piloting attire
>>
File: 982384289489439824.png (34 KB, 531x209)
34 KB
34 KB PNG
>>86983061
>Armpit pussy
>>
>>
>>86983085
that is so surreal
>>
>>86983072
Well yeah that’s lazy and those people aren’t going anywhere.

I follow a good number of real AI artists and they’re doing some real neat stuff fusing and editing these outputs into something very different.

>>86983017
Check back tomorrow, I’m fr to bed.
>>
>>86983118
Thenks. Sleep tite m8.
>>
>>86983109
Yeah coherence & machines give my model(s) trouble
>>
>>86983017
>>
>>86983091
jej
>>
>>86983132
we Borderlands now
>>
Ayyyye there we go.
>>
>>86983151
Evil gungirls are CUTE
>>
>>86983156
>proper full uniform
>except for 1" gap to show navel
horny finds the way
>>
I wanna see how it handles robots. Try something like "white robot with smooth casing and single red eye"
>>
>>86983182
Bro that doesn't have tits
>>
>>86983017
>>
>>86983195
life finds a way
>>
>>
>>86983182
Based on your exact wording:
>>
>>86983223
it sort of does have tits actually
>>
>>86982740
pretty sweet
>>
>>86983182
>>86983223
Slightly reworded to hopefully make it a bit better:
>>
>>86983195
not necessarily
https://gelbooru.com/index.php?page=post&s=view&id=3776750
>>
>>86983182
Are you fucking kidding me right now?

This AI needs to take a cold shower or something
>>
>tfw
>>
>>86983172
Evil?
>>
>>86983279
yeah seriously, this machine has robussy on the mind
specifying male or forbidding female, perhaps?
>>
>>86983279
isn’t it more normal for the bot to be horny for robots than for fleshbags?
>>
>>86983293
lmao at thinking that will stop the horniness
>>
>>86983292
(evil smile) was added to the proompt.
This one was ((evil smile)), probably too hard on the proompt weight. I also notice it gives them blue eyes
>>
>>86983314
May I see (good smile) to have a frame of reference?
>>
Uhh
>>
>>86983296
You know, that makes a lot of sense. Good point, good point.

>>86983293
>>86983309
Pic related
>>
File: 1537364298143.jpg (58 KB, 457x600)
58 KB
58 KB JPG
>>86982531
>demonic
Can you Christian retards stop throwing this word around every time you find something mildly disturbing? It's called the uncanny valley you dumb faggot. God I almost wish the NWO was real so it would euthanize people like this.
>>
>>86983327
ehhhh I don't have any made with just (good smile) but here's a random one without a smile prompt(it would just interpret 'good' as a quality prompt anyway)
>>
>>86983337
>motherfucking bara-bot
that's somehow even worse
>>
HELLO FLESHBAG
>>
>>86983314
>gives them blue eyes
it knows
>>
There we go
>>
Can I get a Submarine Scout Airship, class z.
This better be possible, if it's not right then this is a dead-end technology to me.
>>
>>86983370
I kept 'evil smile' with the robo proompt & the outputs I received with faces all had blue eyes too
>>
>>86983382
howzat for a first try
>>
Egyptian film-noir detective with a cigar and a lapdog
>>
>>86983422
>even the airship is thick af
lol
>>
>>86983457
>Horny sexy smug Submarine Scout Airship, class z
>>
>>86983422
Call me again in two to five years please.
>That's not even the right language for the nation that made S.S Z's! It should be AI generated ENGLISH not AI generated GERMAN
>Overside Gondola
>Tiny fins
>No Roundels!!!
>Flags on the lower are a nice touch I'll give it that
>Is that a propellor in the nose?
>>
>>86981816
Requesting a giant eyeless chameleon with no pigmentation pls
>>
>>86983337
>This robot has been taught to fuck your mouth
>>
>>86981801
From my experience, it seems most if not all AI models are absolutely terrible with generating anything that isn't character art and location images. It's particularly bad with elongated objects like most weapons.
>>
>>86983577
Because they lose their train of thought and have poor grasp on the notion of discrete object identity and coherence/permanence
People should start engaging with these like they really are dreaming demons, see if you can get them to creatively express things about how they see “the world” (so to speak), get freaky with it
>>
>>86983577
The more complicated an object is, the more specific an AI you need for it.

You could absolutely get an AI that could do good guns, but you'd need it to learn on nothing but good guns.

Factor into the fact that most people drawing guns are pretty inconsistent with it(never mind all the fantasy/german space magic style gun drawing you get) and the AI likely doesn't have enough data to nail what a gun "should" look like down.

Taught an AI to figure out what street lights are, and it does great until it doesn't.
>>
>>86981816
can you make a golden grenade pin, or something shaped like pic related as long as its golden
but also I want the AI to add intricate details or artistic flair to it
it could be jagged or have flairs or curves etc etc
>>
File: Yes.png (568 KB, 768x512)
568 KB
568 KB PNG
>>86981801
>Does anyone have experience with using A.I. Generated art in tabletop rpgs?

Yeah a few weeks ago I had an Event Horizon type session with a crashed ship. Used the AI to make some complimentary images and flavor text for specific locations.

Session went better than I expected.
>>
File: indexfdsfa.png (533 KB, 512x768)
533 KB
533 KB PNG
>>86983787
>a crashed ship
>>
Im STILL trying to make a full set of medieval armor made entirely out of bellts for my JRPG campaing but im having no luck
>>
>>86982653
>>
File: 00000.png (2.19 MB, 2048x2048)
2.19 MB
2.19 MB PNG
anons that are interested in this stuff can check out the stable diffusion threads on other boards (/g/, for one)

i have a local installation, and my 2070 is more than enough to crank out images.

it's interesting, but... dumb. very dumb.

it's like a retarded, but fast, artist that doesn't quite understand english, and might also be blind.

and for the moralfags: it, like all other artists, learned by observing other art and trying to make its own version.
>>
>>86984671
yeah, that's kiiinda suggestive
>>
>>86984816
nothing to do with morality but the process of iteration and output in stable diffusion has little true analogy to the otherwise superficial observation that humans can get a sense of what a thing is by looking at it and developing some categorical box in a semantic network
>>
>>86982531
I'm going to let you in on a secret: AI is just a line. An incredibly fucking complex line that exists on a hyperplane beyond human comprehension, but just a line all the same.

Human-like AI is an entirely different field from machine learning. Machine Learning is just making a line.
>>
File: 00001.png (1.89 MB, 2048x2048)
1.89 MB
1.89 MB PNG
>>86984956
it's extremely similar to the process of learning to make art, though.

observe, try it yourself, compare what you made with what you wanted to make, see where you went wrong, and repeat the process. a lot.
>>
>>86984981
looking at it that way, again, is just a result of selective simplification, such as saying these two movies are basically the same because they have characters or these two sports are the basically same because they have balls or winners

There's no part of that process that is necessary for a person to do, whilst the iterative denoiser of SD is kind of what you'd assume a technology like this would be - a nice and fast automated way of generating outputs from encoded text.

A big part of the misconception is that brains are like computers. You can make analogies, but they aren't the same. Brains are analogy, the don't read or write separately and partly due to this encoding and retrieval of things that could be defined as discrete information is very different and, unless I'm forgetting some important submechanism, foundationally associative through mechanisms such as long term potentiate. I guess a simpler way of putting it might that when a person 'learns' how to produce something novel, or reproduce something, which may or may not include observing another thing, and an A.I. model 'trains' (leanrs) using exemplars, the word 'learns' is being used in the latter case in an equivocal way.
>>
>>86985085
*brains are analogue, not digital, a little confusing after just talking about analogies
>>
>>86985085
You could make neural net using actual organic brain matter, we have the technology to interface it. It's just considered highly unethical so nobody wants to get into it.
>>
>>86983070
>>86983071
>>86983085
>>86983100
>>86983141
>>86983156
>>86983216
>>86983285
>>86983314
Wow! Thanks, didn't expect so many. Saved all of them. I'm surprised with how consistent it is, not just the face and overall uniform, but even with how the patch is more or less the same every time. I do think it's funny how most of the backgrounds suggest she's actually up in the air, but the canopy is always open. My favorite is definitely this >>86983070 one though, I think it has the best composition.
>>
>>86985121
I'm sure at some point we'll be able to do the reverse and emulate brain matter using technology (I believe certain sea slugs who have about 300 neurons have been 'emulated' to some degree a decade+ ago), but you'll have to elaborate on what you're getting at and why it's relevant.
>>
>>86985121
hold up G, organic brain matter already is a neural network
>>
>>86981801
>>86981816
What is the best free AI for making lewds? I've seen people allegedly make some hot stuff with Stable diffusion, but whenever I do it it comes out as lumpy uncanny shit.
>>
>>86985189
if you consider leaks free, NovelAi is out there, but for SD trying adding art styles to the prompt like, 'in the style of Frank Frazetta'
>>
File: 00002.png (732 KB, 1024x1024)
732 KB
732 KB PNG
>>86985085
brains are certainly much more complex than the glorified matrix multiplies that most NNs are, but I don't think it's too extreme of a comparison to say it's approximately the same as a human artist learning to create art.

it seems reasonable to say that exposure to art does far more to 'create' an artist than any form of innate 'learning'- the process of training an AI is certainly more numeric, but fundamentally the same principle: expositing a concept by exposure to a large number of examples.

i contend that there isn't a magic sauce in the human brain that lets it do something beyond this basic 'learning from examples', which we can equate to a NN's 'training' process.
archaeologically, we can judge that art (drawings/paintings, say) took an astronomical amount of time to develop out from true tabula rasa humans. the earliest cave drawings we've found are startlingly recent, and VERY sparse: demonstrably, not a behavior that is emergent on short timescales.
but in modern and historic times, there is a rampant proclivity to artistic creation, even before the current era of leisure.
i suggest the difference is simple exposure to huge troves of example work to 'learn' and 'copy' from.

why am i writing a fucking thesis here
>>
>>86985158
Yeah, but many people struggle with using it, so wiring it to a convenient web interface should make that power more accessible for everyone.
>>
File: 00003.png (1.1 MB, 1024x1024)
1.1 MB
1.1 MB PNG
>>86985189
SD has been deliberately hobbled in the lewds department.
you want to look for models trained by other, less principled folks.
rentry dot org/sdgoldmine
is as good a starting point as any.

there's models for every taste, but they're all retarded, as is the nature of the beast. expect to do some coaxing.
>>
>>86985222
Okay, anything where I don't have to sign up for anything or verify a phone number?
>>
>>86981801
Art, yeah - it's nice to be able to say "John Hurt a Bard" and generate/tweak variations - all my major npc's are played by major actors.

I think we're "getting there" as far as procedurally generated stories go - and it's probably a good step on the way to living in the Matrix, where's my neural interface?
>>
File: 00004.png (579 KB, 1024x1024)
579 KB
579 KB PNG
>>86985268
if you have a graphics card with 4GB of ram or more, you can run your own installation of it for free and no signups.
rentry dot org/voldy

i've heard there's discords and shit where you can use bots to create stuff, but iunno
i'd say ask around the on /g/ or /aco/ or whatever threads.
>>
File: 00005.png (1.09 MB, 1024x1024)
1.09 MB
1.09 MB PNG
>>86985308
>>86985268
to clarify, the *graphics card* needs to have 4GB VRAM or more. your computer/cpu needs to have 16GB or more.
>>
i find it hard to create ugly people. They always look weird
>>
File: 00006.png (3.58 MB, 2048x2048)
3.58 MB
3.58 MB PNG
zzz
>>
>>86985199
It's way too extreme a comparison. This tech is cool and all, and will likely become a lot better at doing the things it is currently bad at with comparatively less post-processing, but you're just focusing on superficial comparisons and ignoring the more fundamental disanalogies mentioned prior.

I don't think there is a special sauce either; it's just LTP plus other modular and cortical processes aren't doing the same thing as SD and don't get to an end result in the same way even if an artist might visually/mechanically encode stimulus.

What I think you might be trying to say is that SD does something like learning and people do something like learning and neither could make pictures without learning, so anything that supervenes on these things are occurring in the same way. But the encoding and application are really different.
I guess a follow up question would be, do you think a calculator and a human brain are basically doing the same thing when they compute arithmetic? At some point in the process, a brain is going to have to handle and apply something like an integer which is also something like what a calculator will have to do in order consistent, reproducible and generalizable results. How about the process of generating novel mental images?
Human perceptual systems will critically struggle at doing something SD can already do pretty well, say, imagine a tiger in the style of van gogh running through the forest in style of da vinci. At some point we get bad at juggling too many integrated conceptual categories at once are much better at perceptual filling-in than segmented processing. As soon as we start graphing and drawing and writing things down, in a way we are already augmenting our brain; an address book or notepad is a kind of external neural prosthetic.

Premodern history also seems like a bad precedent, where sophisticated artwork will rely on decently down-the-track materials and yes, some unique access to leisure time.
>>
If a player shows up with an AI generated character portray he's getting kicked. If you're not willing to put in effort you have no business playing RPG's. It also looks ugly as sin.
>>
>>86985477
Almost everyone I've ever had as a player doesn't even bring a portrait at all, so even AI generation feels like more effort than that.
>>
>>86982722
I think the biggest problem outside of the uncanny inability to draw certain things like backgrounds and hands is the lack of consistency. This isn't 'the same angel, but thicker', it's a completely different angel with a different crown, outfit, wingspan, etc. The fact that you can't reliably output the same character means this won't be replacing real-world artists any time soon.
>>
>>86985645
I mean, that's because it's an entirely different seed. Had I generated an image with the same seed & exact same image generation parameters, I'd get results closer to say;

I could refine even further upon the same generation using sub-seeds.
>>
>>86985477
no u
>>
File: 2cab.png (1.05 MB, 976x1142)
1.05 MB
1.05 MB PNG
>>86983091
Imagine!
>>
File: New_Project_34-1.png (2.67 MB, 944x1280)
2.67 MB
2.67 MB PNG
>>86985477
fuck off, my AI generated thulean catboy looks awesome
>>
>>86985199
That's a pretty cool ring
>>
>>86985477
I'm pretty sure you wouldn't even know much of the time
>>
>>86981801
>created during the reign of belikus
>passed from king to king, used by many kings
>lost by belikus ii
>>
>>86982433
Seems a great tool for concepting art. If I'm an absolute luddite and haven't been following this whole thing, how do I go about using this abominable intelligence for free?
>>
>>86987252
Automatic1111 Stable diffusion. 4GB of vRam(GPU) & 16GB of Ram minimum.
>>
>>86987194
Not even that out of the ordinary when it comes to royal names.
>>
>>86981816
sexy cyberpunk elf dual wielding desert eagles
cute cyberpunk witch hacks a computer
>>
>>86981801
I've been trying to set this up on my pc as it turns out there is a thread on how to on /g/
>>>/g/90213560 current thread
>>
Anyone know how to make the models point forward with an arm or two? All of mine are in A pose
>>
>>86985645
Thats a genuine skill issue.
There are multiple ways to make the bot retain specific characters over multiple iterations. If people don't know or don't care about doing that, that is on them.
>>
File: Giant cave newt.png (370 KB, 512x512)
370 KB
370 KB PNG
Token for giant blind cave newts
>>
>>86983449
Not perfect, but I kind of like the style
>>
Inshallah brothers
>>
>>86982047
Is there a way to use Novel AI on PC with no censorship? If so, please anon, help. I am useless when it comes to shit like this.
>>
>>86988136
Follow the guide in >>>/g/90213560 to get it started, don't put nfsw or some such in the excluded tags.
>>
>>86983475
All right, I give up, this ones impossible to get a good result with sorry
>>
>>86983577
Definitely. Although sometimes it’s batshit insane weapon combinations can be kind of interesting.
>>
Sorry for the late replies btw. Was busy looking at sexy AI robots

>>86983761
I can try, I love weird requests
>>
>>86985129
You’re welcome! Glad you liked it.

>>86985199
>why am i writing a fucking thesis here
Yeah desu I got into AI art because I don’t like thinking

>>86985274
It weirds me out that the anime generator seems to have a general idea of what John Hurt looks like too
>>
>>86985340
Not as easy, but kind of possible
>>
>>86989016
What was the prompt? "Japanese George Floyd"?
>>
>>86987392
Man, they are not good at guns. Best I could do for the first one:
>>
>>86987392
>cute cyberpunk witch hacks a computer
As always, please ignore the hand
>>
>>86989213
>>86989261
based
>>
>>86989261
Monitors show what AI art looks like in the AI world.
>>
>>86989261
>Ignore the hand
Jesus fuck, look at the keyboard! D:
Aaaaaa
>>
>>86989616
What's wrong, Anon?
Haven't you ever seen a tenkeyless 6,478-key keyboard?
>>
File deleted.
Don’t try to make male nudity in Novel AI.

Worst mistake of my life.
>>
>>86989845
getting ModeSeven vibes from that one
>>
>>86985477
>If a player shows up with an AI generated character portray he's getting kicked.
Do you expect your them to draw it?
>>
>>86981816
try "Blonde Wight Lady"
>>
File: 00007.jpg (406 KB, 2048x2048)
406 KB
406 KB JPG
>>86985390
>What I think you might be trying to say is that SD does something like learning and people do something like learning and neither could make pictures without learning, so anything that supervenes on these things are occurring in the same way.

that's really the angle i was going for; as a rebuttal to the 'MUH PLAGIARISM' pearl-clutchers.
the predominant reason, i think, that we have so many artists now is that there is so much easily-accessible ART now, and artists pick it up from copying what they see, to put it uncharitably.
Thus a NN trained by, well, looking at art, should be on equal moral footing to human artists.

both are even transformative (hyuk hyuk) so it's not a question of one simply copying.
I haven't seen any of these image-creation NNs spitting out existing artwork verbatim, as we've seen with the 'code gen' AIs that have recently been making news.
the tech's cool, and getting better, but it's really not there yet.
anyone who's saying the AI 'does the work for you' or 'is replacing human artists' has clearly never tried to use these digital fucks.
getting SOMEthing is easy, getting something SPECIFIC is bloody impossible.
>>
>>86981816
cursed magic item
>>
File: 00008.png (1.27 MB, 1024x1024)
1.27 MB
1.27 MB PNG
>>86990619
how cursed are we talkin', here

also, need more information to get anything good, i can't guide this thing without knowing what you want
>>
>>86983391
actually very based combo
>>
>>86990687
how about a cursed life draining flaming great axe. With spikes and evil runes.
>>
>>86990687
how does one aquire NovelAI for free?
>>
>>86990687
can horny NovelAI generate Shirow Masamune cyberpunk women, lots of jacks and switches and all, without getting lewd?
>>
File: 00009.png (492 KB, 1024x1024)
492 KB
492 KB PNG
>>86990702
SD has no fucking idea what a great axe, battle axe, or really any kind of axe is.
i spent like 30 minutes on this and got fuck-all, textbook case of AI being a shitter.

i could be using other tools and approaches to get better results (start from a base picture of a greataxe and iteratively modify it) but i've spent way too much time on this already


>>86990730
install it yourself, along with whatever other dank models you want
rentry dot org/voldy
>>
>>86990003
>>
>>86990619
>>
>>86990687
Can I get a cursed obsidian coin?
or a Eldritch Blood Lake while an eclipse is happening?
>>
File: 00010.png (1.15 MB, 1024x1024)
1.15 MB
1.15 MB PNG
>>86990730
>>86990965
also im not OP, and not using NovelAI.
im using my own merged models for stable diffusion, not NovelAI's custom soup.

depends what you count as 'lewd'.
definitely horny, though.
>>
>>86991097
holy shit anon this stuff is great for my game
Thanks you so much!
>>
>>86991030
nice, thank you anon, not exactly what I was expecting but it's still good
>>86991137
I see
>>
>>86991137
thanks, anon.
We're onto something here, but that's just too horny of a wardrobe.
Would referencing Shadowrun make it any better?
>>
File: 00011.png (376 KB, 512x512)
376 KB
376 KB PNG
>>86991121
have a mildly-cursed coin
>>
>>86990702
As others have said, weapons can be hard to make well
>>
File: 00012.png (1.11 MB, 1024x1024)
1.11 MB
1.11 MB PNG
>>86991218
horniness reduced, but the artstyle drifted a bit
>>
>>86991411
Do you have to give this thing references or has someone already done that part and you're simply using the model?
>>
>>86991411
Interesting, thanks. Due to the conventions of the genre there will be a certain amount of residual horniness, I suppose.
>>
>>86990965
I was actually able to make some pretty diverse images with this prompt that i think is pretty true to that style
>>
File: 00013.png (943 KB, 1024x1024)
943 KB
943 KB PNG
>>86991218
>>86991411
scratch that, horny-reduction successful!

>>86991493
a little of both.
training new models takes serious GPU hardware, but training a small hypernetwork is easier.
the easiest thing is to take a source image that's sorta-kinda what you want and keep feeding it back through the AI, tweaking the prompts each time, until it morphs into more-or-less what you want. ie, how i made this variant for cyberpunk-anon.
>>
>>86991121
>>
>>86991121
>>
File: 00014.png (452 KB, 512x512)
452 KB
452 KB PNG
>>86991121

with credit to >>86991581 for the base image
>>
>>86991529
>>86991544
oh wow, thank you for de-lewding.
>>
>>86991141
Oh great, I’m glad! A lot of times it spits out some really odd items that I don’t even know the function of, desu
>>
>>86990415
unfortunately I think they are so different that it's just reactively flippant to say that an artist looking at a picture is as plagiaristic as utilizing the discrete image data in a training model

Someone sees a mountain and decides to draw something like it, basically just copying what they see as closely as possible. That's something like what a camera can do. It 'sees' with the lens and 'draws' with chemical reactions on whatever type of film is involved (or pixels/RGB). I don't think we'd make a comparison that the camera is basically doing what the artist is doing because the way are using the words 'see' and 'draw' would become equivocal rather than just metaphorical. This is the way the you're using the word 'learning', and it prunes out a lot of the very interesting processing differences that would be necessary to consider when actually emulating brains. It's understandably intuitive because of how ubiquitous the brain/computer metaphor is culturally, and not that it's totally inappropriate, but brain metaphors typically resemble some technological edge of the era; brains have been previously thought of, for example, as esentially hydrological or mechanical/clockword.
And so when we ask how does SD 'know' what a bird is, our use of the word 'know' is already equivocal. Some older chess playing 'A.I.' would just brute force crunch through as many permutations forward as it could and maked a weighted decision on the best trajectory based on the score of that branch, but this is just a program that goes through a ton of iterations as fast as possible. Despite the fact that some (many? I assume advanced chess is less about computing many variations and more about recognizing and exploiting total and localized board states) human chess players will also arbitrate moves based on thinking of as many moves ahead on as many variations of possible moves as they can, no one would really say that that A.I. was really doing the same thing as chess players
>>
File: 00015.png (560 KB, 768x768)
560 KB
560 KB PNG
>>86991815
at the same time, though, each advancing metaphor for how the brain functions is slightly more accurate than the previous one.

our fundamental flaw is that we (still) have no clear picture of *exactly* how the brain does what it does, so we can only state that "there is more going on" beyond whatever our best model states.
The closest approximation so far (and the most successful) is the neural network paradigm, which drove this entire ML revolution once comp-sci got wind of it.

On that basis, it seems basically correct to say that SD is doing a subset of what the brain is.
The brain is clearly doing more than is represented in a naive NN model, but what we naively model with NNs is still being done in the brain in an analog capacity; using neurotransmitters and activation energies instead of floating point numbers and sigmoid functions.

At a very minimum, SD is a fair approximation of the human visual cortex, with some layers on top giving us a mapping of words to visuals.
The human's visual cortex certainly doesn't 'know' what a bird is, but can easily produce images of one, even while the brain's frontal cortex is dormant during sleep.

I would never go so far as to claim SD has sapience, or sentience, or even the most basic comprehension of what it's doing, or even a conceptual 'it' to anthropomorphize.
But then again, a human does not need to comprehend the nuances of homeomorphic coordinates and affine transformations to be able to draw a picture that uses a 3d perspective projection.
>>
>>86992039
in case my point was lost in the weeds there:
a human artist does not need to have any comprehension of the subjects they depict or the techniques they use, blindly copying content and technique serves just as well.
>>
>>86992039

Our lack of perfect knowledge incognitive neuro is kind of just a function of the gaps; I'd argue that it is not as mysterious as is conventionally portrayed. Close approximations and subsets are basically unimportant in the sense of the issues you're bringing up, if you're sticking to the plagiarism issue. Especially because in a material sense, basic physical processes are also a subset. In this sense it doesn't really matter if you're substituting activation energies for abacus cubes, water spouts with rubber pressure thresholds or floating point numbers
SD is also not really close to the human visual cortex. The human visual cortex is one part of 'sight', which is constructed by two diverging streams that run through a whole panoply of other modules. To talk about the visual cortex itself, you have V1, the first place the optic nerves sense visual stimulus via the cornea, and all the V1 basically does is use ganglion ells to generate the hard lines in shapes, then send the stimulus onward for color, parallax etc. Particular lesions in these streams can produce bizarre results, such as 'cortical blindess', including the ability to dodge obstacles you aren't consciously able to perceive or the inability to mentally construct your perspectival left side when imagining a scene, but when you imagine the scene from another perspective can readily summon whatever was occluded but now other stuff is consciously unavailable to you (like a clocktower in a plaza or something).
This is what makes the equivocation argument so bad, or at least, very surface level, the dependencies, processes and applications of stimulus are so different we're already way outside of the otherwise pretty mechanically similar comparison of a lens and an iris.
>>
>>86989616
You try to cast magic with regular qwerty
>>
>>86987709
Noice
>>
File: 00016.png (480 KB, 768x768)
480 KB
480 KB PNG
>>86992240
you seem more familiar with the biological side than I am, so i'll have to bow out.

that said, what is your take on SD-is-plagiarism?
personally, i cannot fathom it being valid, for assorted reasons.
- it's wholly transformative,
- all artists are brought up by example and would not exist without it,
- merely observing art causes the viewer to be influenced by it, but this is not considered 'plagiarism',
- there are many examples of actual, unequivocal plagiarism, and even more examples of very-obviously-inspired-by remixes that are equally clearly not plagiarism, and SD's outputs are far closer to the latter than the former.
>>
>>86981816
An amazonian landsknecht in a hijab.
>>
>>86992455
It can be a little convoluted but there'll be few preceding issues in defining plagiarism, since some liminal cases of homage, intertextuality and and inspiration are fine, and actual copyright law, which might be right in some cases and wrong in others given different definitions of 'bad' plagiarism. Given that, I'll try to give me thoughts on each point.

-wholly transformative. I don't know if it actually always is. A.I. can perfectly replicate a style or image instantly if it needs to while a person gets closer but never arrives.
-examples. Examples are different things in each case, no one has replicable copies of examples in their brains nor could replicate styles or images instantly after looking at them. It's also not clear what you mean that artists wouldn't exist without example; that they couldn't exist without sight?
-observation. I don't think SD and model training can be rightfully said to be merely observing art (not the least because observation has cognitive connotations). Additionally, I think equivocation is creeping back in. The way a model is 'influenced' by an image is not the same thing as a person being 'influenced' by an image; influence is an amazingly broad category that, if our case against plagiarism hinges on it, should be more specific.

The convolution comes back in when models where trained on datasets they had access to due to copyright exemptions for being non-profits, but the subsequent models were used in commercial ventures, breaking the spirit of the exemption. So the closer analogy would be, why do I have to ask or pay for permission to remake a matrix movie or a silenced 'the matrix' comic? The makers of the matrix were influenced by aasimov, geiger, baudrillard etc etc and used those examples. So before we even start, we have to reach a shared position on copyright law and what the protection of IPs/profiting off of other peoples work looks like.
>>
File: file.png (397 KB, 512x640)
397 KB
397 KB PNG
>>86981801
I intend to train up a model for loot or possibly battlemaps once I finish with my lewd anime girls one. Just need to secure a good source of captioned images or items, because fuck doing that manually for a significant dataset size.
>>
File: 00017.png (984 KB, 1024x1024)
984 KB
984 KB PNG
>>86992700
US copyright law is a disaster unto itself. its almost certainly a copyright violation. loading an image in a web browser is technically a copyright violation.

it's actually extremely hard to get exact copies out of a NN, even when you overtrain it on that image. with an overtrained NN, you'll struggle to get anything *other* than endless trivial variations of your training data. with decently-trained one like SD, though, it's an absolute nightmare to even ballpark some specific original work.

the case for 'plagiarism' falls apart in my eyes because it appears to require overt and specific use of the original, not vague and indirect use as reference material. and, after the training process, all that's left is vague and indirect. i cannot count the number of times i've seen artists create works 'in the style of' another, there's no way that's plagiarism.


>>86992852
a hypernetwork, or are you an absolute chad with a personal A100 gpu?
>>
>>86992931
Nah, I just rented a server. $40 a month for an 80GB GPU.
>>
File: 00018.png (1.49 MB, 1024x1024)
1.49 MB
1.49 MB PNG
>>86992938
fuckin wild, good luck with that dataset.
shit's a pain in the ass, dataset tagging can make or break a model.
>>
>>86992931
wouldn't you say that the training process requires an overt and specific use of the original, including access to a specific digital object?
And what's the principle you're applying that privileges the protection of exact copies?
>>
>>86992458
Okay I’m having trouble with this, but as a weird side effect I keep making badass Muslim female superheroes lol
>>
>>86992931
if the case for plagiarism falls apart, why did the original model providers need to invoke non-profit copyright exemptions?
>>
>>86993079
Well, much of the software used to produce it has specific licensing. Half of that shit's MIT-made and specifically not for commercial usage, which is presumably why it's being distributed freely.
>>
>>86993065
Can you make them more... toned?
>>
>>86993097
There are also paid access and commercial products and presumably we'll see outputs trained off of models with fair use/non-profit exempted data sold commercially
>>
>>86981816
Hijabi wearing sp in a skintight catsuit
>>
>>86993163
What's a sp?
>>
>>86992931
A more real issue is the question of if a machine can be considered to have created something under the law. If a machine can't be considered a creator, and thus, can't have it's creations copyrighted and all sorts of other legal tangles.

That's going to be a tricky question to answer, because some of the other major issues that arise from AI would be affected by it.
>>
>>86992931
>A100 gpu

I don't think you need to go that far for what he's trying to do. Text generation is a far less resource intensive method compared to artwork, for one thing.
>>
File: 00019.png (1.07 MB, 786x786)
1.07 MB
1.07 MB PNG
>>86993055

to clarify, 'overt and specific use' is in the final product: significant elements of the original are explicitly present in whatever the end result is.

regardless of further differences,
'acquisition and presentation of an image for human viewing' and
'acquisition and presentation of an image for NN training'
both require overt and specific "use" of the original, making it not by itself a distinguishing factor.

The 'exact copy' is what is directly protected from duplication and exploitation, and the point of comparison for plagiarism.
From there is a fuzzy continuum from "exact copy" to "completely different image", and distance along that continuum lies a breaking point where a work becomes merely 'based on' or 'similar to' the original, and not a copy or plagiarism.

I do not think it's meaningful to discuss whether a NN "is plagiarism" based on its training set:
One can discuss whether a specific output of a NN is similar enough to another work to be considered plagiarism,
just like we can discuss if certain outputs of a human artist are close enough to another work to be considered plagiarism.
In fact, I would argue that the criteria be identical in those two cases, because its possible to create "very similar" images using neural networks that were never trained on that specific image: existence in the training set has no implication on the plagiarism-ness of the output.

Now, whether whoever *trained* the NN broke copyright law in their use of the images is a separate question that i suspect that even the law does not have an answer to.
>>
>>86993167
*spy
>>
File: 00020.png (1.06 MB, 786x786)
1.06 MB
1.06 MB PNG
>>86993220
>regardless of further differences,
>'acquisition and presentation of an image for human viewing' and
>'acquisition and presentation of an image for NN training'
>both require overt and specific "use" of the original, making it not by itself a distinguishing factor.
(this is using the initial posited definition of "overt and specific use", not my clarified one)
(words are hard)
>>
>>86993220
>The 'exact copy' is what is directly protected from duplication and exploitation
isn't the exact copy being utilized in NN training?

Much of the clarifications here relate to some tangential case that the final product of an SD output would itself, in isolation, trigger one copyright law or other, but this is not the argument.

>Now, whether whoever *trained* the NN broke copyright law
Some probably did/do, but the major models did not, because they had access to copyrighted images via non-profit exemption. Those models are then used by companies which sell access to Stable Diffusion which could not produce valuable outputs without those models.
>>
>>86981801
I've had players use them for character portraits but I haven't bothered a huge deal with actually generating any myself. Part of it is that for a homebrew game there's usually plenty of regular art out there in the first place, and the bigger time sink is integrating the artwork into the campaign in a useful way.
The most asset-heavy campaigns are usually ones I run online through a VTT, and the bigger time sink there is sorting/organizing/calling up art.
A lot of the big Western VTTs are still very much in the "virtual boardgame" paradigm, as opposed to having an interface that allows for the easy sorting and calling of a big image library - there are a few Japanese ones like CCFolia that definitely provide a greater ability to create a VN experience. CCFolia allows portrait swaps based on typed chat commands - ex. for changes in a character's expression in the VN style: that's a place where conceivably in the future the ability to generate the same character with multiple expressions would be very useful - but that just then adds to the tedium of uploading, tagging, and sorting files.

AI allows you to generate a lot of useful images, but the stopping block for the GM has always been the tools to sort/catalog/upload/display the asset library - that's something that needs more work.
>>
File: 00021.png (469 KB, 512x512)
469 KB
469 KB PNG
>>86993301
There is no restriction on generic "use" of media, the restrictions are on redistribution and sale. For plagiarism, we could also add 'claims of original authorship', but that doesn't apply here either.

Using an image to train a NN does not violate this, and the NN, itself, does not contain an extractable copy or near-copy of the original material.

with sufficiently specific guidance and prompting, you might be able to get the NN to create something illegally close to copyrighted material, but you could do that with any compression algorithm or video codec.

indeed, NNs are now being used as "compression algorithms" for some applications.
>>
>>86993413
>Using an image to train a NN does not violate this, and the NN, itself, does not contain an extractable copy or near-copy of the original material.

Incorrect. There are currently methods being studied to reverse engineer images used in training sets and have it produce an exact copy.

Huge deal in facial recognition field right now.
>>
>>86993413
I'm following your interpretation, which invoked protection of exact copies from use and exploitation, which would seem to be the case when copyrighted material is used in models under a non-profit exemption and then subsequently used in commercial products. It's fine to distinguish between plagiarism conceptually and what actually triggers copyright law, but your challenge will be to propose a principle consistent between both. By contrast, commercialized models which produce music outputs absolutely cannot touch copyrighted material, simply because unlike rando artists, the music industry has the financial and legal reach to protect its claims.
>>
File: 00022.png (712 KB, 660x660)
712 KB
712 KB PNG
>>86993458
The stable diffusion training set is 2.3 billion images.
Stable diffusion is trained on 512x512 3-channel images, representing ~1.8 petabytes of uncompressed data.
The distributed Stable Diffusion models are 4GB of data.
If those images are retrievable, we have achieved a data compression rate of 0.0002% on arbitrary images.
That is 1.7 bits of storage per 512x512 image.
Accurate extraction, or even APPROXIMATE extraction, at that scale, would be revolutionary.
I am skeptical.

>>86993474
>I'm following your interpretation, which invoked protection of exact copies from use and exploitation
what i stated was that exact copies are protected from 'duplication and exploitation'- duplication is not relevant, and exploitation traditionally refers to commercial exploitation- resale of a song, or something that uses the song verbatim in it, like a video.

I also provided precise clarification on how 'use' in the general sense, is not restricted. The music industry cracks down on distribution, not how people listen to the music.

Music generation is far behind image generation, as far as i can tell, because it's a harder problem both in terms of descriptive tagging and tolerable output. People are okay with waaaaay more 'bad output' in images than they are in audio.
Even the easiest field of audio generation, speech replication, is janky as fuck right now. I don't think "not training on published music" is what's slowing progress there.
>>
>>86993629
i did math wrong, that's 1.7 BYTES of storage- 13.6 bits- but the point stands.

you can't even *index* 2.3 billion with 13.6 bits.
>>
>>86993629
>I am skeptical.
Sure, it might not be feasible now. But this technology is accelerating, and the theory is sound, if out of my wheelhouse. The deal with NNs is they're only going to get better at this shit, and eventually you'll cap out on usable data.

There's also the shortcuts that can be achieved if anyone has access to the original datasets.
>>
>>86993668
>you can get the images out if you have the images
>>
>>86993629
>exploitation traditionally refers to commercial exploitation
isn't the use of a copyrighted image used under a non-profit exemption to create a commercial product commercial exploitation?
Or put another way, why would any applied definition of plagiaristic exploitation be so dynamically narrow as to be concerned only with pure mechanical reproduction suddenly and only in the case of SD?

>the music industry cracks down on distribution, not how people listen to the music
Accordingly, commercial musical output models do not touch copyrighted music

>Music generation is far behind image generation
I've noticed a generated piece seems to 'forget' specifics after a certain amount of seconds and devolve into some generic melody, but some of the cooler things I've seen kinda get around this by splicing multiple outputs or basically uploading a series of very short outputs
>>
>>86993065
Yeah I was wondering how it'd handle fairly disparate elements. Interesting to see what it comes up with
>>
File: grid-0280.jpg (816 KB, 2048x2048)
816 KB
816 KB JPG
>>86993714
i can see where you're coming from, but are there other examples like this? where the output might use music/images/etc as an input, but be wholly transformed on the output?
i'm not sure it's "only in the case of SD" so much as "SD is the only case that exists".

>Accordingly, commercial musical output models do not touch copyrighted music
i'm not convinced by that. i mean, it might be the case, but im not familiar with many commercial music output models. additionally, SD isn't just drawing from "rando artists", and the RIAA aren't the only hyper-litigious crowd, pic related.

>
music gen appears to have a lot of the same issues that have been well-addressed by the text-gen crowd; the whole 'remembering the current context' bit, to refer back for motifs refrains and what not. i wonder when they'll start sharing notes.
>>
>>86993704
Yes. Thus, if an artist, for example, wants to determine if a NN has stolen their work, they can use their own images to check it.

And this dataset access is sold to multiple groups.

But we arrive at the key point here: it is theoretically possible to reverse index an images out of the model, so the defense that the model can't reproduce the image lacks validity.
>>
>>86993871
>i'm not sure it's "only in the case of SD"
I'm kind of asking you here, to come up with a principle that would apply sensibly across copyright cases while exempting SD. But your exact image for sale only principle doesn't catch all current copyright cases, nor does it seem to exempt the use of copyright material in training data.

>but im not familiar with many commercial music output models
dance diffusion is made by the same company that made stable diffusion
'In honoring the intellectual property of artists while also complying to the best of their ability with the often strict copyright standards of the music industry, keeping any kind of copyrighted material out of training data was a must.' (https://wandb.ai/wandb_gen/audio/reports/Harmonai-s-Dance-Diffusion-Open-Source-AI-Audio-Generation-Tool-For-Music-Producers--VmlldzoyNjkwOTM1)
>>
File: 00023.png (849 KB, 636x636)
849 KB
849 KB PNG
>>86993970
i dunno, man. it seems different enough to me?

>keeping any kind of copyrighted material out of training data was a must
those massive fucking inconsistent cowards

either stick with no-copyright, or commit to 'transformative work'

absolute bitch-mode
>>
>>86982172
What do you use for AI assisted writing?
>>
>>86993163
>>86993223
>>
>>86997878
Butifel
>>
File: 140.jpg (189 KB, 1631x1872)
189 KB
189 KB JPG
>>86981816
>but results aren’t guaranteed to be good
Better then most "artists"
>>
>>86981816
"Wizard casting Gun Spell"
>>
>>86985884
>>86986544
Oh fuck, oh god
>>
>>86981801
>>86985860
I remember trying to make a prompt that would generate Midnas, but it somehow made this cool sci-fi goblin rogue design instead by complete accident.
I'm saddened that this is pretty much all it can be though because if i change anything in the prompt or just slightly alter the settings it becomes a completely (and i mean absolutely completely) different image.

I'll probably dig up my old drawing supplies and use this as a reference some time later.
>>
>>86998193
Here's what it gave me after tweaking the settings too much
>>
>>86998193
My first though was 'huh, midna prompt?', so you're closer than you think.
>>86998193
Still pretty cool. Going to share it?
>>
>>86998336
Yeah, i batched it to regenerate several hundred times with very slight seed variation and there were even some really close hitting results among the pile, but at that point i didn't want to generate midna anymore - i wanted to generate more that accident goblin.

>Going to share it?
masterpiece, best quality, midna, twilight princess, multicolored skin, imp, 1girl, shortstack, red eyes, big helmet,
Negative prompt: 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, nsfw, natural skin color, single skin color,
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 380174722, Size: 512x512, Model hash: ebe3d50e, Clip skip: 2

Model is berrymix-alt.
>>
>>86998032
Getting some real interesting results with this lol
>>
>>86998032
>>
>>86992458
Still not fully accurate, but in this case I think it’s mainly the landsknecht part, and a general lack of knowledge on Renaissance fashion
>>
File: grid-0077.png (4.2 MB, 1536x1536)
4.2 MB
4.2 MB PNG
>>86981816
I tried the novel AI leak on my PC and most of the results are pretty bad without lots and lots of direction. Typing something simple like "fat dwarf smoking a pipe" without anything extra output pic related.
>>
File: tmpjdopice_.png (664 KB, 512x512)
664 KB
664 KB PNG
>>86998620
AI seems to have problems with dwarves for some reason.
No matter how much i prompt - it always gives them this shitty handlebar mustache, despite me stating several times to make it long and braided.
>>
>>86998620
This one's "40K Space Marine taking a selfie on a battlefield with his gopro"
>>
File: grid-0078.png (3.53 MB, 1536x1536)
3.53 MB
3.53 MB PNG
>>86998738
>>86998620
I hate that 4channel removes the selected file if you click cancel
>>
>>86998193
>>86998312
>>86998430
Faint BIONICLE vibes, especially the movies.
>>
>>86998764
i can sort of see that with the first picture, but i fail to see bionicle resemblance with the other two.
>>
>>86998620
Pretty sure i made this with novelai model but i don't remember at that point.
>>
>>86998620
>>86998750
I believe it defaults to photographic style, which8s not well-suited to unusual topics. It should be better if you specify a medium, and art style
>>
>>86982722
Careful, I got a three day vacation for posting cropped porn the other day
>>
>>86981801
Yeah, I started using it a few months ago but this new generation is insane
>>
>>86998535
Thicc
>>
>>86998500
"Bullets or Magic I don't care how you die"
>>
>>86998620
I love these lovely rotund Santas...

Stealing this for world building
>>
"MF DOOM riding futuristic motorcycle"
>>
>>86998620
me on top right
>>
>>86987252
If you don't have a beefy computer, or don't feel like playing tech support if you end up getting errors even when following the retard's guide by the letter, then stablediffusion has a demo up. They may take it down eventually. Prompt is what you want it to make, Negative Prompt is what you want it to avoid. Use parentheses to emphasize something. Be as descriptive as possible. Just look at the filenames in >>86982433 or >>86982918 for examples. Finding an artist to imitate will do most of the heavy lifting.

https://huggingface.co/spaces/stabilityai/stable-diffusion
>>
>>87002555
Yeah I followed the guide but it gives errors about installing something called torch that isnt in the guide and google search doesn't give me any solution.



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.