Why does AI have so many issues when it comes generating believable weapons? It can get so many things right, but guns are still a massive issue.
Artificial Intelligence and Guns
Why does AI have so many issues when it comes generating believable weapons? It can get so many things right, but guns are still a massive issue.
What caliber does this even use?
It has problems with all sorts of things, you just don't notice because you're more interested in guns than the rest. It fucks up any sort of vehicle, machinery, etc when you look at the details the way you look at guns. Hell it can't even get 's hand right. Like WTF is going on with her pinky?
Noob it's with a integrated loudner
that's a .45, because they don't make a .46
3-pounder round shot.
GOATSE
'anime face' has a very common structure to work off of. 'gun' does not. Every gun has a wildly different form factor and stylization
I think it's the same problem it has generating hands.
AI right now works by trying to copy a bunch of images it's seen which have been categorized by type. We have tons of art-tags for hair styles, or facial expressions, and the like, so it has a lot of information about what it's meant to plug in where. Hands meanwhile are very complex organs, ten fingers with three knuckles each that can be in a wide variety of poses, and none of them are labeled with exceptions like 'fist' or 'peace-sign'.
Guns are similar. Most people when they draw a gun will just label it 'pistol' or 'rifle'. They won't individually label the parts of the gun in a drawing- the chamber, magazine, sights, etc. Just the broad type of gun it is.
So when a machine is told 'gun', it grabs every picture of a gun, be it a pistol, shotgun, minigun, or even a flintlock, cause those are all 'gun' and it hasn't been told the constituent parts of the gun like has been told the constituent parts of the face. So it smashes them together cause the machine hasn't been given the context of what makes up a gun.
Not an expert but the entire approach is a dead end.
You basically need to teach an algorithm to work like an artist.
All AI arts is just a collection of morphing things from a large dataset together instead of actually knowing the actual logic of drawing things.
Yeah but conceptualization and inspiration is something TRUE AI would have, and that's beyond our capabilities.
A machine would need independent thought to create information where there isn't any for such a thing. Right now machines are incapable of new ideas, so for where computers are right now we'd need to teach it the individual parts of the gun and how they go together, rather than teaching it the gestalt of the gun and hope it figures out the rest on it's own.
Actually, it's all about 2 things:
- empirical data
- context
The latter fleshes out the former. If you start feeding the AI with a bunch of clearly labeled and differentiated gun models and gun parts, it will start building a conext for what, for example, an M16A4 is, and what its receiver looks like. With more angles and lighting conditions provided, it can build up color and form data, which will eventually make it possible to create an AI version of an M16A4 from any angle in any light conditions. The problem is, people haven't been diligently feeding any AI with such PrepHole-centric data, at least not publicly.
There is the infamous example that Google and co's image recognition recognizes black people as monkeys. And feeding more data to those algorithm didn't fix that problem.
Wonder what you'd get if you fed a computer only images of black people and then asked it to make a white person.
I know on PixAI they got several engines for making solely asian women.
>I know on PixAI they got several engines for making solely asian women.
With huge dicks or 7 vaginas, yeah.
>Wonder what you'd get if you fed a computer only images of black people and then asked it to make a white person.
Same as currently just the opposite,
If you take e.g. https://civitai.com/models/62778/majicmix-sombre
and tell it black or white you get Asian models with black and white clothing or hair.
If you train it exclusively on sub Saharan Africans, and tell it white, it would, at most, be capable of producing an African albino.
>Wonder what you'd get if you fed a computer only images of black people and then asked it to make a white person.
wonder no more, anon:
>Midjourney
Yea, that checks out.
That's pretty funny.
You just need to get people to solve captchas where they click on pictures of certain gun parts.
That way Google can keep track of who knows about guns too.
>That way Google can keep track of who knows about guns too.
Uhhh, what. Meds perhaps?
Current AI isn't even AI, they're neutral network approximations. You're looking at literal
>neuron activation
In code. That's why it can't do what actual artists can, all it does is vomit out the equivalent of a discrete "drawing function" for the "neurons" activated:
https://futurism.com/the-byte/watch-ai-die-neural-network
That's the reason why even the most basic things like perspective, proportions and anatomy are literally always wrong.
You basically need an universal trained AI with basically everything, which isn't possible with our actual data storage and computing capacities.
This is the dumbest, most pseud tech-bro omg science bullshit I have ever read
We call the constructs "neurons" out of convenience, they are not analogous in any way shape or form to an actual neuron.
All "AI" does it apply a ridiculous shitload of statistics to arrange data in a way the user finds desirable (sometimes)
Exactly. Artists can make distinctions that AI can't. If you ran this image through Stable Diffusion, the AR-70 and 92 on the floor wouldn't be recognizable.
>All AI arts is just a collection of morphing things from a large dataset together instead of actually knowing the actual logic of drawing things.
This is actually not true and is a common misconception. Neural networks dont "mash together" images, and actually the network learning process is literally teaching the network the "logic" (in a hyper complex multi-thousand dimensional matrix-space mathmatical sense) of how to "draw" these things.
this is not how humans learn to draw. humans need understanding of 3D space and shapes. once you understand those, you can draw pretty much everything with only seeing a few examples. humans don't need to see toddler images before they learn to draw like toddlers. just shows how dishonest this entire marketing ploy of "AI learns like humans" is
Not true. AI can depict things with stunning accuracy, it just doesn't have enough input data for specific guns.
If you did a lora for a specific kind of gun - let's say a thompson machine gun - it could create them with reasonable accuracy. It just depends on how long you want to train it for.
Listen little blonde sweetheart I just want to love and protect you
>Hand over the headpats now anon, and nobody gets hurt
I do so willingly
What if she wants your money too?
How feasible is a double-barelled magnum with an underslung attachment?
literally just this except with an over/under derringer instead of a tokarev. very feasible
>feasible
God I wish.
I want that fucking thing more than anything.
Don't let your dreams be dreams. Get to work.
I need one of these for a 1911 to fire dragon's breath or napalm fuck your shit rounds
She can have the whole world if she wants it
'AI' just imitates patterns. It doesn't understand the logic behind these patterns, so unless you babysit it it won't be able to generate complex structures properly, such as hands or guns.
Think of it this way:
Ask 1000 artistically gifted people off the street to draw a shotgun, then ask an artist to combine those 1000 images into one. The 1000 people of the street might or might not know what a shotgun is but the artist tasked to combine the 1000 definitely doesn't and just guesses by common elements. There you have your strange AI gun.
Now ask those same 1000 people to draw a blonde e-boi and go through the same process with the clueless artist. Blonde e-boi is a way more agreed-upon concept, whereas shotgun is not only way too vague but most people off the streets (or in this case Pixiv artists) might not even be too aware how they might look and work.
In addition, artists often just make up guns, often for rule of cool.
And that's honestly the same sort of problem that nogunz artists have when they draw guns that make no sense IRL. They remember snapshots of guns they've seen in movies or TV and they try to draw their own based upon those images they remember, but since they don't fundamentally understand the guns and they're just recalling little details here and there instead they end up mishmashing features which don't make sense, like drawing a pump handle on a break-action shotgun, or a toplever on a magazine-fed shotgun, ejection ports not aligned with barrels, sights that don't make sense, etc.
It's mostly based on real life images, similar to a google search for shotgun.
Like others have pointed out AI only imitates a common structure and pattern, but what you need to keep in mind is that it's not a real 3D pattern, it's only the replication of common 2D elements in images.
The amount of "thinking" happening is along the lines of,
Gun has straight lines converging to a common point and some round things on the end, and some dark stuff under the beige blocky oval things (grip sticking out under the hand).
There is no understanding of the function of the parts involved.
This is the same for any interaction with any kind of object, it's trained on 2D images, not on 3D datasets and it doesn't have a 3D dateset to compare the 2D output to and no logic understanding what it is assembling.
On its own AI is currently already good at creating portraits but no interaction with objects or the environment. If you want interactions to make sense you need to add additional data, dept information, a simple mannequin, hand templates etc. This is possible but can not be expected from an auto-generated image without added manual input.
Depends on the type of art, those interested in productivity see it as a tool, those making oil-paintings etc. don't care. If you draw digital pinups for a living you might be fucked as AI can replace most of the the job and create custom stuff more in line with the potential customers fetishes (if there is a data-set for it if it's too complicated). Those only drawing for fun as a hobby welcome the addition of a new tool, but might lose some drive to get better themselves.
>It's mostly based on real life images
Wouldn't that be the case only for something imitating photos? You can't teach an AI on a dataset and then tell it to generate images in an art style that wasn't included in that dataset.
The base dataset (1.5 , SDXL etc.) is created using real images, giving it an anime-style is additional training of this dataset.
The new dataset is only getting trained on Anime https://civitai.com/tag/sdxl by default it's photographs, same with the previous ones.
>photographs
small correction, meant photographs, old paintings and whatever you can use without legal problems.
Because the algorithm was deliberately fucked with to avoid producing realistic guns. No joke.
So you mean to tell me that every single AI art bot, no matter what they are trained on, for some reason would fuck up guns specifically when they most of the time can't make a picture to look coherent in the first place?
Get off it
>every thread talking about AI
>"but muh art this muh artist that"
moron stop bringing that up, literally no one fucking cares, go back and doompost on your containment board >>>/ic/
Are artists that mindbroken by AI?
Literally 21st century equivalent of luddites.
Yes. My waifu who does art for commissions is absolutely booty blasted by AI art.
>they just copy other artists and steal their work
>meanwhile turns people's LinePlay avatars into art ignoring the fact she is using other people's are and ideas as the basis for her inspiration
Considering they shit up and derail every AI related threads I'd say yeah
It's not art, it's image generation, happy? Now go back
Art or image generation, doesn't matter. What in this thread about images that originate from AI is to be discussed if we aren't supposed to discuss images that originate from AI?
>Thread about AI art
>"STOP TALKING ABOUT MUH ART AND ARTISTS"
What did he mean by this?
To properly understand the morality of AI art generation in the modern era, we need to carefully study Judge Dredd comic books from 37 years ago.
Well that's depressing. We should have listened.
tags: Todd Howard, the creator of the Skyrim series, holds you hostage with an assault rifle for not purchasing Starfield, the sequel to the Oblivion series.
No wait Todd wait! I was waiting for the game of the year edition I swear!
Todd is not convinced.
I've yet to see actual REAL artists who know their craft complaining about AI. It's mostly just shitters.
It's 'better' at more abstract guns, like sci-fi plasma weaponry. There's a lot of factors that influence the AI into being worse at firearms than other aspects.
First, firearms are often hard to make look right, period. Look at all the cringe kino of videogame developers, movie/TV prop producers, and other amateurs trying to make realistic/plausible gun replicas, they often fall very, very short.
Second, there often isn't a ton of high-quality image data that is properly tagged and curated to train the AI to correctly identify different guns from each other, or even something like holding a radio with a long antenna. Guns also have a lot of similar surfaces that each need to be angled differently in just the right ratios to look 'right', many AI models can struggle depicting these kinds of precise, logical engineering surfaces. You could ameliorate this by making LoRa's with good, focused training data, but it's a deeper issue.
If you look at the amount of attention to detail put into tagging all the physical anatomy/clothing/position options in images, then compare that to the corresponding tags for various weaponry related things, it's a stark difference in breadth and depth. A lot of the time, the artsy types just aren't well-versed, or even interested, in depicting realistic firearms.
Several reasons. The top three being Bad samples, see pic related for a M4. If you don't know what a cat looks like and someone keeps giving you pictures of dogs you won't get a cat.
The second one is term wise 'gun' isn't very specific for an AI to try to understand.
Third is the devs often deliberately sabotage it's learning when it comes to guns because CA tech land says gun bad
Yes marin is very cute, but what does she have to do with AI
She is a cosplayer, she can basically be for Gojo anything dhe could dream of?
>Kurumi Tokisaki? Get me that clock lense
>Chris Yukine? Let's 3D print her revolvers
>Revy? Help me with that water tattoo.
>Igawa Asagi? Really? Uh, oh, hehe.
That is what they dream about the AI, to be capable to all and none a master of all, but that would mean all of us become obsolete, and that we cannot allow.
I dont think her relationship with that guy will last very long
Nah, considering she is a self-insert of the creator and Gojou her fantasy husbando, and this being anime means you may expect them rolling kids and whatnot...
But seriously, on topic, a lot of people won't mind the silly structure of the AI drawings, it will pass most of the time if you aren't a PrepHoleommando.
Why does she has a kitten stuck in her ass?
It's kinda an easy way to understand things, like when you basically tell old people ChatGPT is like a Super Google which tries to go the extra mile, they suddenly understand how to use that stuff.
You basically need to teach AI how to generate 3D models of basically everything and teach the AI the basics of art generation like perspective, proportions. And then combine those things.
Soi basically how artists actually operate.
AI sucks, it can't even get boobs right, and there's probably no greater number of training images available on the entire internet.
gun
the charging pipe
for all your shooting & smoking needs
anti shill countermeasures
No one has trained one with enough well tagged gun pictures.
Even with some LoRAs helping with hands and guns, they're still a major ass-ache to get right. Hopefully some more finely-categorized models come out soon.
The problem is guns are a lot more niche than anime porn. I bet you it was a calculated moves for all these big engines to let you make porn, cause they knew that would be the biggest demand for this sort of thing. Maximize engagement so that people can work out the kinks for you.
Guns are much too niche since you'd only need them as a prop in character art, or trying to come up with some fantasy gun design.
PrepHole should have a stable diffusion general like PrepHole and /h/ where we all generate military chicks and funny weapons. Cmv.
Minute variations in biological anatomy or clothes and things are acceptable but guns are very finely built objects that adhere to extremely strict proportions. If you want variation, that means the AI won't duplicate what it was trained on 100% of the time, but guns MUST be perfectly duplicated, every time.
It can kinda do 'em though. This is from Midjourney 5.2
Made this one with sdxl. Not bad honestly.
Yeah, SDXL is much cleverer than SD1.5. I think Midjourney's models are somewhere in the middle. Not as smart as SDXL but smarter than 1.5 - thus, with more room for storing data about guns.
Did you know that analysis has revealed internal depth representations of scenes within Stable Diffusion models? It's already doing that.
SDXL is definitely better at understanding your prompt.
I don't know what you thought my prompt was but that ain't it chief
>you will never go to war with your skelebro
Why must life be so cruel?
Minus her lazy eyes.
Please put that gun away your highness, I swear I don't know anything about missing pairs of undergarments!
Guns are slowly getting there with newer models.
Also it seems to overfit M4 style rifles
>LeMat automatic
>ask the AI to make a slimegirl
>get this
I don't even know what to say.
Hey man, a holes a hole.
depends to a large degree on if the model was trained with images containing the concept.
Actually surprised the xl one I just played with knew what I meant.
You'd usually add a "lora", a dateset trained on the type if image you're looking for and add it to the model pushing it in the right direction.
But dedicated slime girl data seems to currently only available for 1.5 ... so the guns only make sense if you assume they are currently getting eaten.
>slimegun
I want one!
images of guns are not exactly common
most images of guns aren't exactly clear
many images of guns are very different from guns
the average stock images from a image dump site rarely contain guns
the typical A.I. programmer isn't interested in guns
people asking for guns off an A.I. is rare
this all congregates to a programmer that isn't going to feed their A.I. many images, nor are they going to try to fix algorithms that are producing flawed guns, meaning guns are gonna look funky
personally, i kinda like it, its like the guns from old B movies that pretend their PVC & cardboard mess is the future of hand held weaponry
It's any complicated object in general, stupid. Just try it with instruments or tool.