mabit said:
My main arguing point here is that there is no fundamental/architectural difference between human learning processes and that of AI's, at least in the case of image generation like we're talking about here.

There may not be a "fundamental difference", but there are clear differences. A person has and continuously generates their own experiences, while an AI is given a curated training set from the people controlling it. Humans have behavioral tics that influence what they do, how they think, how they remember things, and change with time, creating more individuality in what they ultimately create, while an AI is just following a script made by its creators.

Unless and until AI can be granted personhood, the things they create are considered derivative works of its input training set. Currently, the law does not recognize the works of AI to be copyrightable, it's lacking the creative spark that's necessary for copyright protection, which creates an inherent difference in the works produced by AI vs a person. Without that creative spark, it's just deriving its results from its inputs (using complex and imprecise math), it's not adding its own unique creative element to the work.

Maybe this will all change with time, but I think we're going to run into bigger questions regarding AI being considered people, and the moral and philosophical issues that entails, first.

watsit said:
A person has and continuously generates their own experiences, while an AI is given a curated training set from the people controlling it. Humans have behavioral tics that influence what they do, how they think, how they remember things, and change with time, creating more individuality in what they ultimately create, while an AI is just following a script made by its creators.

But that does not change the fact that all of those are essentially just emergent behavior that come out as a result of your neurons firing and strenghtening their connections with time

All that your eyes see is captures of the outside world, just like the ones that can be fed to a network. Those networks have memories that can change with time, and the way each of them learns is affected by tics that influence their output. Is the only tangible difference here that one of them has their input given by humans then?

So if I were to code a neural network exactly like others we have right now, except instead of being given an explicit training dataset it would browse the web and construct its own unlabelled training dataset for learning, it would be ok? Because that's basically what its being doing already according to OpenAI's papers from what I've checked

watsit said:
Unless and until AI can be granted personhood, the things they create are considered derivative works of its input training set. [...] Without that creative spark, it's just deriving its results from its inputs (using complex and imprecise math), it's not adding its own unique creative element to the work.

Yes, my point is that it's exactly how humans create too. You can only draw a horse because you've seen pictures or videos of horses before, seen them personally or seen drawings of them. Those are all inputs in your own training dataset that you'd use to construct a new image for me.

For a "creative spark", that once again just goes back to the "we humans are special, we have souls" argument, which honestly can't be argued for or against in any way because at that point it's just philosophy.

What we can say though, is that every artist, writer, musician that has ever lived has been drawing from their own experiences to create their works. They take what they've learned and play around with it to create something brand new with it, which in turn inspires other people who do the same in a continous process of artistic natural selection. And that's exactly what neural networks are created to do

mabit said:
But that does not change the fact that all of those are essentially just emergent behavior that come out as a result of your neurons firing and strenghtening their connections with time

And that does not change the fact that the way humans operate in the real world is different from the way AI operates in the real world. That's a fallacy, "because we're similar in this one way, we must be similar in these other ways".

mabit said:
Yes, my point is that it's exactly how humans create too. You can only draw a horse because you've seen pictures or videos of horses before, seen them personally or seen drawings of them. Those are all inputs in your own training dataset that you'd use to construct a new image for me.

Except I explained the difference in how we ultimately create what we do. The different types of input, how we come about those inputs, and how those inputs influence us.

mabit said:
For a "creative spark", that once again just goes back to the "we humans are special, we have souls" argument, which honestly can't be argued for or against in any way because at that point it's just philosophy.

I never made that argument that "we humans are special, we have souls", which I don't even agree with. I explained the difference in both the amount of and how we come about our experiences, and how AI is more directly controlled by its creators. I even said this could change in the future (as AI becomes more advanced and become more independent), but as it currently is, there's a difference in a human creating something from their experiences, vs an AI creating something from its experiences.

I would just like to point out that I do quite love the research, particularly the models used for medical applications like identifying cancer cells, viruses, and other conditions to help human doctors. But feeding a model copyrighted works to allow it to output pieces of said works at a prompt and give no credit or royalties to the people who have actually made that trainings set possible is not a good use, and absolutely flies into the face of what this site is about.

If these neural networks actually become good enough to extrapolate from their data sets, or someone makes a trainings set that isn't 99% copyrighted material, we might change our tone. As it stands for the current models inspiration is all we allow it for as far as it's use as a tool goes.

And the moment we can have actual conversations with the damn things to the point where a consciousness can be reasonably assumed we will be revisiting the point of entirely AI generated artwork. Until then it's little more than a fancy tool for automated tracing and collage making from millions of newspapers for every image output.

watsit said:
And that does not change the fact that the way humans operate in the real world is different from the way AI operates in the real world. That's a fallacy, "because we're similar in this one way, we must be similar in these other ways".

There are no fundamental architectural differences between how a neuron fires in a brain and how it fires in a neural network. You can argue about which activation functions most closely emulate real life, or the implementation of a specific algorithm that does the back propragation of the error function to update the weights of the net. But the fundamental way in which the network is constructed is to emulate real life brain tissue remains the same, any differences are particular to single implementations of the same idea

watsit said:
I never made that argument that "we humans are special, we have souls", which I don't even agree with. I explained the difference in both the amount of and how we come about our experiences, and how AI is more directly controlled by its creators. I even said this could change in the future (as AI becomes more advanced and become more independent), but as it currently is, there's a difference in a human creating something from their experiences, vs an AI creating something from its experiences.

You were using the term "creative spark" as one of the basis of the argument, I assumed it to imply that human brains have something extra that cannot be encoded in a neural network, my apologies if that assumption was incorrect

I still can't quite understand what would be the unique human experiences that a neural network wouldn't have though, would you be able to elaborate? What would be the fundamental difference between a human brain and a neural network that would affect the amount or type of experiences that it can have?

mabit said:
You were using the term "creative spark" as one of the basis of the argument,

Because that's how copyright law is written. To be eligible for copyright protection, a work has to show a minimal amount of "creativity", and without that "creativity" the work is treated as either derivative of another (and there's an extra hurdle to separate a derivative work from a transformative work even when creativity is present), or purely factual.

mabit said:
I assumed it to imply that human brains have something extra that cannot be encoded in a neural network, my apologies if that assumption was incorrect

We do have something extra, our neural faculties are much more advanced/nuanced than AI, with many more inputs and pathways that alter ourselves. I wouldn't go as far as to say such a thing "cannot be encoded in a neural network", but I would say we don't currently have the capability to encode it. Maybe at some point in the distant future we'll end up being able to, but that's not the state of it right now.

mabit said:
I still can't quite understand what would be the unique human experiences that a neural network wouldn't have though, would you be able to elaborate?

Emotion (changing how we perceive past and present experiences based on how we feel), a subconscious (subtle details to remembered experiences that we aren't actively aware of, but which change our behavior anyway), flawed memory (misremembering past experiences, which changes how we may express it in the future), inherent biases (giving more weight to certain details over others for no apparent reason). On top of a decades of life experiences continually shaping us in every way imaginable (and more ways unimaginable). That's far more than a computer being asked to pattern match various words to a selected set of images, and then generate images when given an assortment of words.

watsit said:
We do have something extra, our neural faculties are much more advanced/nuanced than AI, with many more inputs and pathways that alter ourselves. I wouldn't go as far as to say such a thing "cannot be encoded in a neural network", but I would say we don't currently have the capability to encode it. Maybe at some point in the distant future we'll end up being able to, but that's not the state of it right now.

Yes, I completely agree. That's why I mentioned that the only difference is in our number of neurons and the number of connections each neuron has

watsit said:
Emotion (changing how we perceive past and present experiences based on how we feel), a subconscious (subtle details to remembered experiences that we aren't actively aware of, but which change our behavior anyway), flawed memory (misremembering past experiences, which changes how we may express it in the future), inherent biases (giving more weight to certain details over others for no apparent reason).

In a human brain, all of those you're describing are chemical changes that affect the electrochemical pulses that synapses carry. The way you're describing it, emotions translate to temporal fluctuations of network weights, while subconscious/flawed memories/inherent biases are all translated to a network's actual biases that are introduced based on the current network's weights and new inputs it receives

Those are all things that exist in neural networks being made today. Biases actually seen as generally undesirable and a big part of network modelling is actually about identifying them and using wichever heuristics you can to remove them from your network (aside from the "emotion" part, which just isn't put in because we really have no need for that sort of mechanic in most tasks)

It's not in any way anything exclusive to human brains. It's something that's not required in closed domain NN applications so we do our best to not let them affect tasks negatively

watsit said:
On top of a decades of life experiences continually shaping us in every way imaginable (and more ways unimaginable). That's far more than a computer being asked to pattern match various words to a selected set of images, and then generate images when given an assortment of words.

What you're saying here it's literally a matter of corpus size then, which also isn't anything limited to the human brain in any way. If anything, you can easily argue that neural networks get more input in a specific domain than any human, and in just a fraction of the time

mabit said:
In a human brain, all of those you're describing are chemical changes that affect the electrochemical pulses that synapses carry. The way you're describing it, emotions translate to temporal fluctuations of network weights, while subconscious/flawed memories/inherent biases are all translated to a network's actual biases that are introduced based on the current network's weights and new inputs it receives

Those are all things that exist in neural networks being made today.

No, I don't think they are. AI doesn't have what we consider "emotion" (and is one of the last things we should ever consider purposely giving it, since emotion is the basis for irrational behavior, and giving AI irrational behavior on purpose as it becomes more advanced and complex constitutes a grave danger). AI doesn't have a subconscious (it learns specifically what we tell it to, it doesn't pick up on and get influenced by perceived details even it's not aware of). Flawed memory would be considered an error and something we'd work to remove. And as you say, biases are generally considered undesirable here and we try to remove. All these things that influence human behavior and particularly the kind of content we create, we purposely try to keep out of AI. Ergo, AI creating something is different from humans creating something.

mabit said:
What you're saying here it's literally a matter of corpus size then, which also isn't anything limited to the human brain in any way. If anything, you can easily argue that neural networks get more input in a specific domain than any human, and in just a fraction of the time

Which is what I've been saying. Yes, fundamentally they're built on the same concepts, but the size and breadth is vastly different, and many of the things we attribute to the basis for human creativity and uniqueness, we try to keep out of AI. So ultimately anything we may perceive as "creativity" from an AI is an illusion from how we made it respond to inputs, just like how people can perceive sentience from an AI, but it's just an illusion created from how it was made to respond to various inputs. And I mean, an illusion is itself created as a result of errors in our brain's ability to process certain combinations of inputs, yet those very illusions give rise to many things pervasive in our culture (the idea of dragons and anthropomorphic animals, were similarly created from our inability to properly comprehend various things over the generations, which spread around and becomes the bedrock for the creative outlet of many people; that's not something I see AI doing anytime soon).

Remember when Flash and 3D rendering created low-effort artwork on image sites? Gonna have to have rules updated on some sites! :P

It seems like my post leaned further into the nature of the AI vs human learning than it probably should have, given all these responses to that specifically... how these programs actually learn, how complex they are, or how they work, isn't really relevant to me or why I don't think generated images should be accepted by e621 or other art galleries/art archives. I don't see this technology being practical for "replacing" artists, so I don't have disdain for it out of fear, Clawstripe raises some good points that I agree with, which are why I have a generally negative view of these AI art generators, or at least the hype around them, their creators and their advocates. Sure, some of the images are funny or look cool, the fact that a program can interpret text into a loosely or even mostly coherent image is interesting, but I don't really think it's art.

My main point is that AI generated images don't involve any effort from the end user, the only person that put in any effort is the programmer(s), there's no artist involved.

These images are generated in minutes, with no input outside of the initial prompt. You could use Stable Diffusion or DALL-E 2 to create 200 "paintings" that look to be acceptably "professional" looking (at first blush) within an hour, obviously putting aside any imposed limits on individual users. These are essentially disposable images, if you don't like what you got from the generation? You can tweak your prompt, or leave it as is, and just hit the button to re-roll the dice until you get something that's coherent, relatively appealing and "close enough" for you to accept. To put these generated images alongside artwork that took time, effort, knowledge, had creative decisions made and a process executed - regardless of that art's quality - is to say that these disposable generated images are equal in artistic merit and value.

For those reasons, I don't think that kind of content is appropriate for this site, and I see AI generated images as the equivalent to a recoloured base or a screenshot of a doll maker game, if you want to share cool looking AI generated images, there's always Imgur, Reddit, and Twitter, they shouldn't be taking up server space on an art archive.

The only exception that could be made are generated images that are dramatically altered from their original state, possibly with a link to the original.

hungrymaple said:
You can tweak your prompt, or leave it as is, and just hit the button to re-roll the dice until you get something that's coherent, relatively appealing and "close enough" for you to accept.

This kind of is where I tend to disagree somewhat. The prompts used to get extremely good results require an understanding of what each keyword means as well as how they would interact with one another. Current good AI generated works tend to require chains of keywords and possibly even customized weighting. Stable diffusion can run on individual computers and can produce quite good results, but the prompts can be hard to understand to a casual user.

Yes, running a prompt that is simple can produce OK results, but specifying things like the medium, the style of works it should prefer, and the quality of those works while giving supplementary details that should be added gives massively improved results. That art piece which won a state competition took a while to tinker with the prompt to get that result. https://www.midjourney.com/showcase/ (the AI used to win that competition mentioned yesterday) has a selection and their respective prompts. Also take a look at the stable diffusion subreddit and the posts with prompt included as their flair. Those prompts can be quite abstract on their own.

I find it closer to the diamond argument. Natural vs artificial. The former is limited by environment and the great amounts of time to generate, the latter is going to be determined by the initial conditions of their creation and is quite fast by comparison. If the latter starts approaching the quality of the former, people will step in with a "vetting" process and/or laws to restrict the latter to protect their business.

Seems like a lot of people here are only familiar with AI image generation in terms of the old web-based sites like DALLE mini. The models we have now are growing exponentially more complex and sophisticated.

hungrymaple said:
My main point is that AI generated images don't involve any effort from the end user, the only person that put in any effort is the programmer(s), there's no artist involved.

That is not strictly true. Sure, you can just use the default settings, plug some words into a text box, and click a button, and maybe you'll get something good. But getting the most out of txt2img generation requires a fair bit of knowledge on prompt engineering and experience with how to manipulate the AI into getting the exact results you want.

However, the best use of image generation is for AI/artist collaboration. I started with a low quality generation from a prompt, and refined it until you can hardly tell it was AI-generated (EDIT: NSFW warning):
https://i.imgur.com/o4ZrMXI.png
Here's another example where I started not from a text prompt, but a drawing:
https://i.imgur.com/hfSDGnA.png

Both of these examples required a great deal of input from me. It necessitated a lot of layering, masking, and manually redrawing certain parts like hands and the face.

AI image generation is so much more than a text box and a button. You can start with little more than a colored sketch and end up with an incredibly detailed masterpiece. We've only scratched the surface of its potential as a collaborative art tool.

So... Stable Diffusion happened. I guess this is the killer app for nVidea's excess GPUs! :D

AI dont have idea or imagination
but they can draw our idea . anybody's dream can be drawn just by words even with fail amtemp its fun to mess aroud with it
it can be used to draw criminal potrait from the victim words
the new replace the old is natural like if u somehow create clean cheep energy it would make oil industry go bankrupt many people lost their job
i guess we should learn coding now

The other thing is i just don't see how you want to tell what is AI drawn and what is not.

Yes, we see a great amount of Highly detailed Pictures, but it also can go another way and just give a middling amount of details. Or what is when they use img2img to refine the details?

Not to mention that you could even finetune your own model to create a specific or even unique "style" in the generation of your own private model

Another somewhat well-known site just revised their upload policy to exclude AI submissions:

2.8 Content Lacking Artistic Merit
Content lacking artistic merit is not permitted on FA, and includes such items as:
-- Submissions created through the use of artificial intelligence (AI) or similar image generators.
[...]
Content created by artificial intelligence is not allowed on Fur Affinity.

AI and machine learning applications (DALL-E, Craiyon) sample other artists' work to create content. That content generated can reference hundreds, even thousands of pieces of work from other artists to create derivative images.

Our goal is to support artists and their content. We don’t believe it’s in our community’s best interests to allow AI generated content on the site.

I'm not sure how that factors into the discussion here, but I figured it was worth mentioning.

kora_viridian said:
Another somewhat well-known site just revised their upload policy to exclude AI submissions:

I'm not sure how that factors into the discussion here, but I figured it was worth mentioning.

As much as some results amazes me, it was starting to get way too spammy there. Specially with fixations on specific body parts. So I'm glad they did something about it.

I have tried it and it seems like plain english is not the best way to tell it what to do. I think it should be more like a programming language. If it was a programming language, it would take more skill to get decent results, but it would give the user more control.

In my opinion, if the user has more control, we could call them an artist and it would be like if they were drawing, exept they would write code instead of drawing themselves.

electricitywolf said:
In my opinion, if the user has more control, we could call them an artist and it would be like if they were drawing, exept they would write code instead of drawing themselves.

I can write a program to make various pictures when given user-specified numbers (with the chosen numbers affecting the resulting picture). Such pictures would not be accepted even if the results appeared vaguely furry-related, and the person providing the numbers wouldn't be considered the image's artist. Being able to format your Google searches to make it return more relevant results doesn't make you the one that scoured the internet to find the pages, it just means Google trained you to make better queries for better results. Similarly, being able to format your request better for these art generators doesn't make you the one who drew the art, it just means you made better requests that gave better results.

watsit said:

electricitywolf said:
In my opinion, if the user has more control, we could call them an artist and it would be like if they were drawing, exept they would write code instead of drawing themselves.

I can write a program to make various pictures when given user-specified numbers (with the chosen numbers affecting the resulting picture). Such pictures would not be accepted even if the results appeared vaguely furry-related, and the person providing the numbers wouldn't be considered the image's artist. Being able to format your Google searches to make it return more relevant results doesn't make you the one that scoured the internet to find the pages, it just means Google trained you to make better queries for better results. Similarly, being able to format your request better for these art generators doesn't make you the one who drew the art, it just means you made better requests that gave better results.

Maybe i am wrong about coders being artists. But i think you may agree that plain english is not the best way to tell those networks what to do.

notmenotyou said:
And the moment we can have actual conversations with the damn things to the point where a consciousness can be reasonably assumed we will be revisiting the point of entirely AI generated artwork. Until then it's little more than a fancy tool for automated tracing and collage making from millions of newspapers for every image output.

We should add "AI generated" to the deletion reason buttons.

lance_armstrong said:
We should add "AI generated" to the deletion reason buttons.

I saw you just deleted my last submission for "AI generated". Is uploading any art which used AI in the creation process not allowed here? Just curious because I'm not seeing it in the TOS.

roflcopter2 said:
I saw you just deleted my last submission for "AI generated". Is uploading any art which used AI in the creation process not allowed here? Just curious because I'm not seeing it in the TOS.

https://e621.net/help/uploading_guidelines

Low quality submissions: ...artificial upscales, AI / neural network edits, AI generated artwork, etc.

AI generated artwork was only added to that list on August 18.

lance_armstrong said:
https://e621.net/help/uploading_guidelines

AI generated artwork was only added to that list on August 18.

Gotcha, thank you!

Tbh though, "AI generated artwork" doesn't really belong under "Low quality submissions". If the goal is to ban it for moral/preventative/legal measures I understand but the output you can get with these is definitely not "low quality". It's often indistinguishable from good quality work from real artists. Not to mention the work involved in crafting a prompt, cleaning up, etc. I imagine in a few months most digital artists will be using at the very least, AI-assisted inpainting, in their workflow.

lance_armstrong said:
https://e621.net/help/uploading_guidelines

AI generated artwork was only added to that list on August 18.

Everybody gangsta until someone starts tracing AI generated images and not telling anyone about it.

Just wanted to say, there's another AI model called Stable Diffusion which is about as good as Dall•E 2, and is free and open source, so you could generate NSFW content if you're running it on your own machine.

mabit said:
[...]

Just a note here, while text-to-image generation models like Stable Diffusion and DALLE are created through deep learning, they aren't neural networks, and neither are neural networks direct analogues of biological neural structures.
I think it's fair to say these models are emulating a form of creativity, but that process is not necessarily similar to how a human brain creates art.
Additionally, these models encode the statistical relationships between words and visual features; unlike humans, they cannot understand the underlying meanings of visual representations.

acidph said:
Just a note here, while text-to-image generation models like Stable Diffusion and DALLE are created through deep learning, they aren't neural networks, and neither are neural networks direct analogues of biological neural structures.
I think it's fair to say these models are emulating a form of creativity, but that process is not necessarily similar to how a human brain creates art.
Additionally, these models encode the statistical relationships between words and visual features; unlike humans, they cannot understand the underlying meanings of visual representations.

Dall-e literally states that it is a neural network, which throws the stable diffusion claim into question as well, so that is going to need a citation. Especially since that is the form of training they went through.
Additionally, you must be a contrarian as ANNs literally were literally inspired in structure based off of how neural connections work, they aren't exact, but then again any form of abstraction will do that.
You would also have to define "understand" in a different way, as their ability to demonstrate something with a defined input could be considered understanding it.

About the only thing you got right in all that is that how they make the works is not the same as how humans do it. A closer comparison is how you visualize things in your head, as that is going to create imagery that is built upon your memories.

acidph said:
Just a note here, while text-to-image generation models like Stable Diffusion and DALLE are created through deep learning, they aren't neural networks[...]

Oh you're right, I thought they were GAN-based but they're actually diffusion models, my mistake
With that said though, diffusion models still employ "common" neural networks that are trained to each step of their reverse diffusion process Ho et al. (2020). I will agree that the learning process I described is not exactly how these latest instances of image generation work, but they are still applying the same concepts, albeit in more complex/refined ways for the domain

acidph said:
and neither are neural networks direct analogues of biological neural structures.

I am gonna need you to provide some clarification on this point though. Neural networks are explicitly bio-inspired algorithms made to emulate basic brain structures, the main differences between them are merely implementation ones (based on our current understanding of biology that is)

acidph said:
I think it's fair to say these models are emulating a form of creativity, but that process is not necessarily similar to how a human brain creates art.
Additionally, these models encode the statistical relationships between words and visual features; unlike humans, they cannot understand the underlying meanings of visual representations.

We can't understand the underlying relations between our neuron synapses either. We know the basic structure and the mechanisms underlying how the chemoelectrical signals are transmitted, but we have little to no clue about how the information is being represented.

I'm worried what might happen if this newest tech becomes mainstream, so I want to put in my two cents.
If we accept machine made images as fully-fledged art, won't that flood the internet with data? Can image boorus survive a year worth of data, uploaded daily? Who will tag it? Machines, right? Will it be flawless, or, considering the pace at which a machine can make images, will it overwhelm any human moderators?
And when the majority of art becomes machine-made, won't machines end up being trained on machine-made art? Will there be people studying art for years, only to make and sell it for pennies to feed the algorithms? Or will art be regurgitated until we enter an era of machine-made abstract art?
Are we in such a dire shortage of graphical content, that we must resort to using a technology that will ultimately make commercial art worth as much as dirt, and cause millions of people to lose employment in a beautiful trade?
I don't know how this tech works; I believe it's a state-of-the-art program that can synthesize data, gathered from people's work (without asking), in order to try and fulfil instructions. It doesn't know when it copied a signature; it doesn't know proportions. Such tech could be used for MRI scans to assist doctors in detecting neoplasms. Why try to force it into places it doesn't belong to?
I am shocked that there are people who equate lines of code to a person, a generic silicon and copper wafer to a unique being with a lifetime of experiences, learning, knowledge, and emotions.
Nietzsche would get PTSD if he learned that we want to kill Man next :D

mabit said:
I am gonna need you to provide some clarification on this point though. Neural networks are explicitly bio-inspired algorithms made to emulate basic brain structures, the main differences between them are merely implementation ones (based on our current understanding of biology that is)

What I mean is that while computational neural networks are bio-inspired, it's not accurate to think of think of them as simulations of biological neurons. Back-propagation does not exist at all in biological neuron structures, for example.

acidph said:
What I mean is that while computational neural networks are bio-inspired, it's not accurate to think of think of them as simulations of biological neurons. Back-propagation does not exist at all in biological neuron structures, for example.

As I said above, that's merely a difference in implementation. Back-propagation is just our way of abstracting the mechanism of synaptic plasticity in an artificial net. It is still made to emulate what's present in actual neurons tho

Edit:
I have no background in neuroscience so I can't say anything with property about the subject, but Fitz, H., & Chang, F. (2019). Language ERPs reflect learning through prediction error propagation. Cognitive Psychology, 111, 15-52. apparently sugests that the exact mechanism of error back-propagation is indeed present in our brains

ausfer said:
I did. We'll see what I get back.

I'm not sure where else I should be sharing this stuff...

There is a discord server for NSFW stable diffusion content, You can try to post your pictures on their furry channel

roflcopter2 said:
Gotcha, thank you!

Tbh though, "AI generated artwork" doesn't really belong under "Low quality submissions". If the goal is to ban it for moral/preventative/legal measures I understand but the output you can get with these is definitely not "low quality". It's often indistinguishable from good quality work from real artists. Not to mention the work involved in crafting a prompt, cleaning up, etc. I imagine in a few months most digital artists will be using at the very least, AI-assisted inpainting, in their workflow.

There is a discord server for NSFW stable diffusion generation. You can try to post your pictures on their furry channel

https://twitter.com/xxylas/status/1573031117447479296

If anyone has more info on this eSix checkpoint for Stable Diffusion, I'm highly interested :O

Supposedly it's on a Discord server, but I couldn't get the link to it.

Honestly, I see these things as a way to get prompts... like, Ok, it generated this kind of tangible but uncanny image, can we use our brains to make it into something legible? That kind of thing. Like uh, what this video does. https://youtu.be/Y0_U8V7buxI

annon-guy said:
https://twitter.com/xxylas/status/1573031117447479296

If anyone has more info on this eSix checkpoint for Stable Diffusion, I'm highly interested :O

Supposedly it's on a Discord server, but I couldn't get the link to it.

I gotchu fam
discord.gg/cNyuGvbT8U
This is the hub for all things furry SD-related. Currently there are 3 different furry models, plus a pony model.

Meta is now into generating videos from text.
https://makeavideo.studio/

I don't care if someone wants to protect the sanctity of art and the honor of artists. Hey, that's on brand for this site, which cares a lot more about artists than most do. But let's just be honest here- if it were really about quality standards, there would be no need to talk about whether it's plagiarism and so on. This is sort of ideologically-driven, which becomes more obvious when personal issues with individuals are brought up. The way I see it, if an image can trick you into thinking it's drawn by a skilled human and you want it gone anyway, it's not because it fails to live up to some objective standard.

Most AI-generated artwork sucks but it's sheer denial to say it's impossible for it to ever live up to human creativity and skill. People who say that sound like luddites. You can imagine a lot of the same arguments being made by anyone who has had their craft taken away by automated processes. A machine can't drive steel like John Henry, it can't outsmart the grandmaster of chess, it can't beat the global any% speedrun of Super Mario 64, and it certainly can't create high-quality images based on prompts. It's a tale as old as the printing press. Nothing makes us inherently special by objective metrics, regardless of how tempting it is to cling to ideals of the noble and irreplicable human spirit. We can never ultimately beat the machines we, ourselves, designed at the games we, ourselves, designed.

This isn't to say artists will wholly be obsolesced. People today still buy original paintings despite the ability to save exact copies digitally! But this is a foreshadowing of how art in general might go. It will be (even more of) a commodity, rather than the only means to secure someone's fantasy of their fursona being inflated like a balloon. Actually, AI artwork has a long way to go before it's able to copy fursonas in particular, but nobody expected things to come as far as they have, so quickly. Everyone who's so sure it's impossible may find themselves in for a rude awakening. Never say never. I would not be so quick to dismiss the potential of technology.

Those of you who scoff at the idea that AIs can replace human artists really need to consider how quickly this has all happened. This stuff has barely been out for any length of time and look at what a tizzy it's already got people in. The fact that it produces low-quality results now should not reassure you when we stand at the very dawn of AI-generated artwork. You are looking at their equivalent to our 6000-year-old cave etchings, and from that perspective, they already have a lot on us. What lies over the horizon? Machines always advance faster than people. It's only going to get better (or worse, if you prefer), and that's going to happen very quickly. You can point out weird hands and soulless eyes all day long, but there is a monkeys-at-a-typewriter phenomenon and it's fairly often that I see a decent AI-generated image pop up. They already have the ability to occasionally fool all but the best-trained eyes, which is to be expected with the sheer magnitude of AI-generated art being pumped out (output rate being one of many inherent advantages they have over humans). That's where we already stand, and the technology will inevitably improve immensely, while human hands and creative potentials aren't going to be seeing much in the way of breakthroughs.

No doubt, some of this art is going to be snuck through no matter how hard anyone tries to fight it, and I don't see why it would warrant punishment if it was seen as worthy of being in the site in the first place.

As for the plagiarism argument, I don't see how it's plagiarism any more than it's plagiarism to copy an art style. Plenty of great artists started out looking like cheap knockoffs of those who came before. They mentally drew from those artists when it came to the decisions they made. Their art was, of course, still their own. And in the same way, regardless of what pools of knowledge these AIs draw from, the art they create did not exist before they brought it forth, and it's not a matter of cropping or moving a few sliders. How is that anything but original?

I get why people still respect artists. Lots of things are more interesting when humans do them. Any sort of competition, from athletics to speedrunning games. But I am not under the impression that it's impossible for someone to make machines do these things better than people. There's a reason they don't let machines compete in chess, and the fact that people have to ban AIs from art competitions says a lot about how unequivocally inferior they supposedly are.

I say all of this not even as a diehard proponent of AI-generated art. I don't want to live in a world where humans are just inferior at everything. I understand how painful that kind of thing is. Maybe people should have thought more thoroughly about whether it's right to do everything in their power to replace people in creative pursuits. But I'm not delusional, I see where this is going, and the genie's out of the bottle whether I like it or not, and whether some websites allow it or not.

NovelAI just released their paid image generation service, with an anime module and a beta for a furry module. They're pretty clearly trained on Danbooru and e621 respectively (they don't really try to hide that or anything). I think the anime model is already wireheading half of Japan, and reportedly it's making AI art discussions hard to follow at work because everyone's posting anime boobs.

The furry model is... it's better than any txt2img I've seen for furry stuff so far. It still does have the usual txt2img weaknesses, so, it can get very confused with unusual poses or anatomies, not to mention multiple characters in a scene, and getting good art out of it is a matter of knowing the random keywords it happens to associate with the "good art" direction in its latent space. But chances are that you can go there, type in the e621 tags for your fetish in the input field, and get out something you can fap to, and no other model that I've seen has done that so far.

lekkiyo said:
Actually, AI artwork has a long way to go before it's able to copy fursonas in particular

Oh, man, if only things were going that slowly. I don't have a good feel yet for how well it works in practice, but people have been reporting success with getting these models to be pretty faithful to their sona with DreamBooth and even just textual inversion (which boils things down to an easily shareable embedding file a few kilobytes large). We might be bargaining over the exact level of convenience and faithfulness, but it's definitely not in the realm of science fiction.

The drawings uploaded to the Discord server have tempted me to conclude aft a few more years have passed AI will deprive the income of those who make a living drawing digitally; will said drawers survive, and if so, how?
Also, in terms of effort, how does one compare using this tool to develop images to humans drawing them from scratch?

foxel said:
The drawings uploaded to the Discord server have tempted me to conclude aft a few more years have passed AI will deprive the income of those who make a living drawing digitally; will said drawers survive, and if so, how?
Also, in terms of effort, how does one compare using this tool to develop images to humans drawing them from scratch?

It will create additional competition by lowering barriers to entry. Including very massive barriers like SKILL. If you face more competition, you are likely to be paid less.

If the coomer, roleplayer, or art enthusiast can write a list of tags or a paragraph of text and get 95% of what they want, they might not want to pay for commissions even if a hand or something else looks fucked up. If they want to copy a specific artist's style, they can train a model on that artist's images. Many hours could be wasted fiddling with these AI models, but it won't be much compared to the 10,000 hours it would take to git gud. The learning curve for using AI software will decrease as these models are repackaged to become more user-friendly, and that will allow more dumberer and younger people to jump on the bandwagon.

Low-skilled artists, not completely unskilled, can use drawings as input for image-to-image translation to enhance them. You can spend an hour drawing something that looks bad, but is detailed enough to guide the AI model to produce exactly what you want. You can adjust model parameters to reduce the amount of dramatic changes made to the drawing, and/or edit it section by section. More people will try to get to this level just so they can get better at using AI.

Highly-skilled artists can do the same types of things with their sketches, and redraw over the results. If you can find ways to cut 4 hours of work down to 2 hours, your output is doubled. The increased productivity could be lower than that, so ask a real artist their opinion if they try that approach.

There are more things that could be done. Ironically, human creativity is what will push these tools to their limits. AI models will become dramatically better after a few years of experimentation, and the hardware they run on will be faster and cheaper. The GPU might not be the only important component, as there are AI accelerators coming to most new consumer PCs soon.

There will be winners and losers because of this. Not every artist will have their income hurt. For some it will be unavoidable.

foxel said:
The drawings uploaded to the Discord server have tempted me to conclude aft a few more years have passed AI will deprive the income of those who make a living drawing digitally; will said drawers survive, and if so, how?
Also, in terms of effort, how does one compare using this tool to develop images to humans drawing them from scratch?

Artists will survive because the internet is going to be flooded with so much AI-made artwork to the point where nobody will care about it. Handmade digital artwork/traditional artwork will remain valuable while AI-made artwork and AI generated artwork will be relatively worthless. A good comparison would be real jewelry vs. fake jewelry. They look the same but the fact that one is fake destroys 90% of its value compared to the real thing.
As for artists who plan on using AI to do most or all the work for them for the purpose of profit, they might not make very much money because their competition will only be other AI artists, not traditional/digital artists.

Another reason would be that lets say a few years into the future, I commission an artist to draw my OC and share it to everyone. People will comment on it, people enjoy the character and the artwork knowing it came from a certain artist, there is a sense of community.
Now lets say I instead use AI to generate a bunch of images in the artists style of my OC, and share it to my friends or social media and whatnot. Who is actually going to care about any of that art? No one. Nobody will give a shit because it wasn't made by the artist, they're just a bunch of fake images that look identical to what they artist would have made.
I think a large portion of the value that comes from commissioning artists is the fact that you also get to share the art with the rest of the community. If all your OC art is AI generated, then you won't really have anything to share to anyone.

Since the decision has been made to not have AI-generated furry artwork here on e621, how about creating a site specifically for AI-generated furry artwork?

Arguments against:

If the "we don't do AI because we like human artists" goes all the way up past the e621 mods/admins, to e621's owners, then this idea is a non-starter.

Also "the stuff we don't like has to go live over there" has lead to some not-great places, historically.

Maybe such a site already exists and I just don't know about it.

Arguments for:

The e621 developers have clearly gotten Danbooru to a high state of tune. They pretty much know what it takes, in terms of server hardware and bandwidth, to host furry artwork at scale. If the business case (ad revenue, possibly driving product purchases) pencils out, they could have "e621 but for AI art" up and running in a week or two.

Alternatively, e621's owners already own another furry art site, which pretty much nobody uses - rebrand it for AI art and see if *that* gets any traffic? It even has "Network" in the name, which could probably be worked into the "AI" branding if desired.

lance_armstrong said:
It will create additional competition by lowering barriers to entry. Including very massive barriers like SKILL. If you face more competition, you are likely to be paid less.

That worries me but it, and the rest of your reply, nuances my view; thank you.

lance_armstrong said:
Low-skilled artists, not completely unskilled, can use drawings as input for image-to-image translation to enhance them. You can spend an hour drawing something that looks bad, but is detailed enough to guide the AI model to produce exactly what you want. You can adjust model parameters to reduce the amount of dramatic changes made to the drawing, and/or edit it section by section.

Highly-skilled artists can do the same types of things with their sketches, and redraw over the results. If you can find ways to cut 4 hours of work down to 2 hours, your output is doubled. The increased productivity could be lower than that, so ask a real artist their opinion if they try that approach.

And now I wonder if AIs will cause some drawers to shift from making artworks for commissioners to making and selling entire worlds. Fearing turns to awe.

thousandfold said:
Artists will survive because the internet is going to be flooded with so much AI-made artwork to the point where nobody will care about it. Handmade digital artwork/traditional artwork will remain valuable while AI-made artwork and AI generated artwork will be relatively worthless. A good comparison would be real jewelry vs. fake jewelry. They look the same but the fact that one is fake destroys 90% of its value compared to the real thing.

Another reason would be that lets say a few years into the future, I commission an artist to draw my OC and share it to everyone. People will comment on it, people enjoy the character and the artwork knowing it came from a certain artist, there is a sense of community.
Now lets say I instead use AI to generate a bunch of images in the artists style of my OC, and share it to my friends or social media and whatnot. Who is actually going to care about any of that art? No one.
I think a large portion of the value that comes from commissioning artists is the fact that you also get to share the art with the rest of the community. If all your OC art is AI generated, then you won't really have anything to share to anyone.

You allay my concerns; many others and I enjoy the skill and intent behind every brush stroke, each detail. While AIs take intent in the form of tags, they still build their works from material they collected—amalgamated wills, as it were. End results of course appeal, but I've oft found that sense of community you mentioned drawing me in further.

notmenotyou said:
...But feeding a model copyrighted works to allow it to output pieces of said works at a prompt and give no credit or royalties to the people who have actually made that trainings set possible is not a good use, and absolutely flies into the face of what this site is about.

If these neural networks actually become good enough to extrapolate from their data sets, or someone makes a trainings set that isn't 99% copyrighted material, we might change our tone. As it stands for the current models inspiration is all we allow it for as far as it's use as a tool goes.

And the moment we can have actual conversations with the damn things to the point where a consciousness can be reasonably assumed we will be revisiting the point of entirely AI generated artwork. Until then it's little more than a fancy tool for automated tracing and collage making from millions of newspapers for every image output.

That is a fundamental misunderstanding of what the technology does. It doesn't "output pieces" of any pre-existing work. Images (copyrighted and otherwise) are used to train the model that the AI uses to generate new images, but the model itself does not contain any representations of copyrighted images, in whole or in part: It's merely an abstracted set of instructions which dictate what outputs are expected when given a specific input or set of inputs. While it is true that the images produced from that model are influenced by the data used to train it, they are completely unique works and not copy-paste collages of stolen assets or tracings of pre-existing works.

Having said that, I will now address the implicit, broader argument that stylistic imitation is inherently a violation of copyright even if no specific visual elements are copied from any specific pre-existing work.

Consider: If referencing copyrighted images in order to imitate the style of another artist is IP theft, then should not every source-accurate depiction of a copyrighted franchise (MLP, Pokémon, Zootopia, Sonic, etc.) be purged from the site? Note that we're not talking tracing here (that would be directly using pieces of the original image(s) to construct a new one, which, as I noted above, is not what the AI generation is doing): If someone looks at a bunch of MLP screenshots in order to learn the abstract rules needed to accurately draw the ponies as they appear in the show, then is that not functionally the same as how an AI model is trained? Why should the fact that it's a person doing the referencing instead of a machine make a difference? They are still using copyrighted material to construct a model that they can then use to infinitely reproduce works in the style of another artist, after all.

Suppose this person then used their experiences to write an instructional book on pony drawing techniques, then published it online? Should they now be obligated to pay royalties on sales of that book to each and every artist whose work they viewed? Or was the act of distilling that information into an abstract set of guidelines sufficient to transform it into an original work, distinct from the works referenced to construct it? I would make a case for the latter: the AI model is itself is a transformative work created by referencing copyrighted images, rather than an infringing artifact which directly contains copyrighted material and, by extension, so too are the images generated with the model.

One need only consider the method by which an artist learns their craft to see why this is the case: Artists do not have their talent bestowed upon them by god in a flash of divine inspiration, not do they will their ideas into existence ex nihilo. An artist wishing to draw a specific character does not need to first logically derive that character's appearance from first principles. Artistic extrapolation is nothing but the ability to take inspiration from the world around you and recombine it in ways that are greater than the sum of its parts; It's absurd to try to separate the artist from their world (including other art they may have seen in their lifetime) and treat them as possessing some unique, inherent "spark" that alone makes them capable of producing True and Honest Original Works™.

I'd go so far as to argue that it isn't even possible for a human to create art without having first been "trained" on other artworks: Imagine asking an artist who had literally never seen furry art before to draw an anthro fox. Would their lack of prior experience with pre-existing (potentially copyrighted) furry artwork make their depiction "more original" than any other furry art ever drawn? Would this exceptional originality automatically make the work a High-Quality Submission if it were uploaded to e621? Or would their lack of reference material for their subject cause them to create something which, though pure of all copyrighted furry art influences, most furries would reject?

Now imagine that the artist had never seen any visual art before and was given the same request. If "lack of copyrighted references" = originality = quality, then logically this hypothetical artist would be the most original and best furry artist the world had even known! But, of course, we wouldn't expect that to be the outcome, would we? Why, then, do we implicitly conflate these three attributes by categorizing all AI-generated art as Low-Quality Submissions, with no regard for the objective quality of each individual work?

By observing examples of the subject that one wishes to depict (including other copyrighted artworks), then generalizing from those examples to reckon how to produce a desired result from a given set of inputs, art can be created: It doesn't matter if it was a human or an AI who did the observing and generalization. Nihil sub sole Novum.

spacies said:
Images (copyrighted and otherwise) are used to train the model that the AI uses to generate new images, but the model itself does not contain any representations of copyrighted images, in whole or in part: It's merely an abstracted set of instructions which dictate what outputs are expected when given a specific input or set of inputs. While it is true that the images produced from that model are influenced by the data used to train it, they are completely unique works and not copy-paste collages of stolen assets or tracings of pre-existing works.

That doesn't necessarily make the statement false. Copyright doesn't just protect the specific pixels of the image, but the creative elements that made up the image. If you take an image and compress it to a JPG, for example, the data completely changes, and even the pixels change, but it's still considered the same image since the creative elements that went into it are still there. Or for a closer analogy, sheet music is an "abstracted set of instructions" for a recorded piece of music. Copyright doesn't protect just the final mastered recording, but also the underlying structure and creative elements put into it, such as the sheet music used to create the recording. So even if someone takes some music, breaks it down to an abstract set of instructions (a sheet music representation), then creates a new recording from those instructions, even with alterations, that could still be infringing. So, if it breaks down an image to get its underlying structure, then creates an image using that structure as a basis, there can be a claim depending on what the final result incorporates of the input. This applies to humans to: if you learn art by studying a bunch of images, then at some point make an image that ends up looking substantially similar to one of the images you studied (even if you didn't realize it or intend to), that can still be an infringement claim. Sure, you can fight back saying you didn't copy any of the protected elements of the image, and maybe you'll succeed, or maybe not. Depends on what the images in question actually are, you can't make a blanket statement that it won't be infringing.

There's a reason many companies turn a blind eye to fan work, so if they happen to produce something similar they have plausible deniability. But if an AI knowingly uses some image as a source and ends up creating something substantially similar, even without directly copying pixels, that can still be infringement.

watsit said:
Or for a closer analogy, sheet music is an "abstracted set of instructions" for a recorded piece of music. Copyright doesn't protect just the final mastered recording, but also the underlying structure and creative elements put into it, such as the sheet music used to create the recording. So even if someone takes some music, breaks it down to an abstract set of instructions (a sheet music representation), then creates a new recording from those instructions, even with alterations, that could still be infringing.

This is not a good comparison point, as the laws regarding music are an entirely different can of worms .

votp said:
This is not a good comparison point, as the laws regarding music are an entirely different can of worms .

Maybe, but the point is, breaking down a copyrighted work to "an abstracted set of instructions" that are used to help create an output doesn't prevent copyright from still applying. It's going to depend on each individual output given all its possible inputs.

watsit said:
That doesn't necessarily make the statement false. Copyright doesn't just protect the specific pixels of the image, but the creative elements that made up the image. If you take an image and compress it to a JPG, for example, the data completely changes, and even the pixels change, but it's still considered the same image since the creative elements that went into it are still there. Or for a closer analogy, sheet music is an "abstracted set of instructions" for a recorded piece of music. Copyright doesn't protect just the final mastered recording, but also the underlying structure and creative elements put into it, such as the sheet music used to create the recording. So even if someone takes some music, breaks it down to an abstract set of instructions (a sheet music representation), then creates a new recording from those instructions, even with alterations, that could still be infringing. So, if it breaks down an image to get its underlying structure, then creates an image using that structure as a basis, there can be a claim depending on what the final result incorporates of the input.

I was refuting the specific claim that AI art generators "output pieces" of copyrighted works and that they are "a fancy tool for automated tracing and collage making from millions of newspapers for every image output", which is a factually incorrect statement.

Beyond that, your counter-example doesn't hold up to scrutiny: As you admit in your reply, compressed jpgs and musical notations are both representational methods of translating the information of an image or song. I specifically stated that "representations of copyrighted images" are not stored in the model, and I meant it. No representation of the information contained in an image is added to the model during training: the model is an original data construct which is shaped by exposure to the training set. It's less "a book of sheet music for a bunch of specific, individual songs" and more "A book of music theory describing chords, scales, tempo, etc. and how they can be used in different ways to achieve different aesthetic ends". That's what I meant by abstract instructions. Maybe I should have used a different term?

This applies to humans to: if you learn art by studying a bunch of images, then at some point make an image that ends up looking substantially similar to one of the images you studied (even if you didn't realize it or intend to), that can still be an infringement claim. Sure, you can fight back saying you didn't copy any of the protected elements of the image, and maybe you'll succeed, or maybe not. Depends on what the images in question actually are, you can't make a blanket statement that it won't be infringing.

There's a reason many companies turn a blind eye to fan work, so if they happen to produce something similar they have plausible deniability.

The fact that these sorts of legal issues also apply to human artists actually supports the argument I made in the latter part of my post: If human artists can also be found to be infringing by unintentionally copying a substantial part of a work they've seen in the past, then is that not evidence that human artists also learn art by observing (copyrighted) materials and building a model, just as the AI does?

But if an AI knowingly uses some image as a source and ends up creating something substantially similar, even without directly copying pixels, that can still be infringement

These AIs cannot knowingly do anything: Ultimately, a human is deciding on the desired output and then trying to define a set of prompts and settings which bring about that desired output. So this situation is really no different from a human artist making a work by hand.

watsit said:
Maybe, but the point is, breaking down a copyrighted work to "an abstracted set of instructions" that are used to help create an output doesn't prevent copyright from still applying. It's going to depend on each individual output given all its possible inputs.

well, as anyone who was conscious around the time that Fortnite: Battle Royale was taking off might know, for choreography, at least, the constituent parts of a piece (single dance moves in that case) are not copyrightable, only larger parts of the performance are. so if you extend that logic to traditional art there's some point where you've broken the original piece down enough to where what remains is no longer copyrightable.

spacies said:
The fact that these sorts of legal issues also apply to human artists actually supports the argument I made in the latter part of my post: If human artists can also be found to be infringing by unintentionally copying a substantial part of a work they've seen in the past, then is that not evidence that human artists also learn art by observing (copyrighted) materials and building a model, just as the AI does?

To a point. As I've said earlier, there is similarity in how these AI work compared to humans. But there are notable differences as well, particularly when it comes to agency. An AI is just doing what it's told with what it's given, no questions asked. Humans are far more varied with many more inputs and functional behaviors that influence their results.

spacies said:
These AIs cannot knowingly do anything

I mean an AI is known to use some source. e.g. that one that used e6's most popular images as a training set, if that AI ends up producing an image substantially similar to an image here, it is known to have had access to that image and there would be infringement as a result.

darryus said:
well, as anyone who was conscious around the time that Fortnite: Battle Royale was taking off might know, for choreography, at least, the constituent parts of a piece (single dance moves in that case) are not copyrightable, only larger parts of the performance are. so if you extend that logic to traditional art there's some point where you've broken the original piece down enough to where what remains is no longer copyrightable.

If it's broken down enough, and the pieces remain separate, yes. Where that line is will be different depending on what exactly you're talking about, which will be on a case-by-case basis. My point is simply that you can't make a blanket claim that breaking a work down to "an abstracted set of instructions" which are then used to help create another image isn't copyright infringement because "they are completely unique works and not copy-paste collages". They can be copyright infringing depending on what the actual output is for a given case.

darryus said:
well, as anyone who was conscious around the time that Fortnite: Battle Royale was taking off might know, for choreography, at least, the constituent parts of a piece (single dance moves in that case) are not copyrightable, only larger parts of the performance are. so if you extend that logic to traditional art there's some point where you've broken the original piece down enough to where what remains is no longer copyrightable.

A good point! Even if the AI model contained representational fragments of the images used to train it (which, I cannot stress enough, it does not) that would still not necessarily be infringement, provided that substantial parts of the original works aren't copied over wholesale into the generated images.

watsit said:
If it's broken down enough, and the pieces remain separate, yes. Where that line is will be different depending on what exactly you're talking about, which will be on a case-by-case basis. My point is simply that you can't make a blanket claim that breaking a work down to "an abstracted set of instructions" which are then used to help create another image isn't copyright infringement because "they are completely unique works and not copy-paste collages". They can be copyright infringing depending on what the actual output is for a given case.

See my earlier reply to your post. You're misunderstanding what I mean when I say "abstract instructions" (a regrettable lack of clarity on my part).

watsit said:
To a point. As I've said earlier, there is similarity in how these AI work compared to humans. But there are notable differences as well, particularly when it comes to agency. An AI is just doing what it's told with what it's given, no questions asked. Humans are far more varied with many more inputs and functional behaviors that influence their results.

I mean an AI is known to use some source. e.g. that one that used e6's most popular images as a training set, if that AI ends up producing an image substantially similar to an image here, it is known to have had access to that image and there would be infringement as a result.

If it's broken down enough, and the pieces remain separate, yes. Where that line is will be different depending on what exactly you're talking about, which will be on a case-by-case basis. My point is simply that you can't make a blanket claim that breaking a work down to "an abstracted set of instructions" which are then used to help create another image isn't copyright infringement because "they are completely unique works and not copy-paste collages". They can be copyright infringing depending on what the actual output is for a given case.

While it is true that the AIs lack agency, the humans using those AIs do not: I'm trying to advocate for AI generative art as a legitimate artistic medium, not argue that StableDiffusion is sentient! :P

Joking aside, humans may have a far more complex "training algorithm" than the AI, but that doesn't disprove my main point, which is that originality is not an intangible property of human thought: what we think of as "original" is actually just a complex recombination of other references and inspirations (some of which were copyrighted), which were in turn inspired by other things, etc. Such that every person could be said to have their own "model" trained on the world around them.

To your second point: If a furry artist is also known to spend a lot of time on e621, then would not the same suspicions be cast upon them if they were to accidentally create an infringing image that resembles something posted to e621? Why is the AI any more of a liability in that case?

Finally, I think you're consistently misunderstanding what I mean by the phrase "abstract instructions" (admittedly a problem with the way I worded it). The AI model is specifically not composed of any representation of any content taken from any of the images in the training set. It is an original work that is shaped by exposure to that representational data and can be used to predict the kinds of outputs that might be expected given a specific input: Claiming it's representational is like claiming TVTropes is nothing but an archive of TV and movie scripts.

The fact of the matter is, AI is here to stay, the genie is out of the bottle and instead of looking for an 'easy-out', we need to create a ruleset for how to apply it and do so responsibly. Banning it is a knee-jerk reaction that will only lead to infighting between people that support it and people that don't. In the end, implementation is easier and causes fewer issues. You won't have people using AI while hiding it so art puritans won't go after their throats, and you won't have issues like AI art spam if you make it so only a limited number of said tags can be uploaded by each user in a certain span of time to avoid flooding.

From what I saw, any moral arguments tend to stem from a lack of understanding of how NNs work. I really hope the issue gets resolved as I do heavily use AI and some of the results are arguably e6 quality worthy. I did post a bunch of pics yesterday and they got deleted(admittedly, I didn't search deep enough into the ToS to know it was an issue at the time) but I did find others have also uploaded AI-generated art and their art is still up for some odd reason. I guess mine looked 'too good'? Anyway, here is a gallery with a bunch of pics and you'll be the judge of how long it takes before AI takes over:

https://imgur.com/a/OHJJbXR

Also, you are welcome to tell me which people this art is 'stolen' from. I am sure you can find a similar 'style' but a lot of artists use a similar style to each other. Meesh and twinkle-sez have similar styles but you wouldn't say they copy each other. In the end, a war against AI is a war already lost. The bubble of people that deny how good it can be will get smaller and if e6 doesn't offer a platform for it, people will migrate to somewhere that does. Let's not mention people commissioning AI art(yes, that is happening already) will have no platform to put on art(at least what they consider art) in their character galleries.

As for the lacking creativity argument, would you suggest that with the advent of electronic music, people like Skrillex lack creativity because they can press 3 buttons on a laptop and produce beats that took hours or days to make using traditional synths and then some would be outright impossible without that tech? No? Then why would people using AI be different from that? It's a new tool, it needs to be treated as such instead of making a huge deal out of it and banning it at every turn. I understand banning AI when the art it created was low quality, but that is starting to not be the case. And given e6's role in the general furry community, I feel if there is any place AI art SHOULDN'T be banned, is here. Besides, anyone who doesn't want to engage with it can simply blacklist the tags for it.

ckitt said:
if e6 doesn't offer a platform for it, people will migrate to somewhere that does.

okay.

you say that like it's going to be some massive exodus. like, I don't think most people see AI-generated images as much more than a novelty. we're humans, we like stories, we like art created by other people, substance, AI isn't able to provide that, and, until AI is actually sentient, it never will.

also, holy fuck, comparing peeps who write prompts into an AI that was trained passively to musical artists that put days of actual fucking manhours into creating actual art is insane.

darryus said:
okay.

you say that like it's going to be some massive exodus, like, I don't think most people see AI-generated images as much more than a novelty. we're humans, we like stories, we like art created by other people, substance, AI isn't able to provide that, and, until AI is actually sentient, it never will.

I am not worried about a 'massive' exodus, I just think it's bad to want the community to split due to the removal of a valid expression of art regardless of your personal view of it. It clearly shows some disrespect to the opinions of others and you could say it's a form of witchhunt rather than a reasonable response to a new emerging tech.

As for AI not being able to provide stories created by people, dungeon AI and GPT-3 models would disagree with that, and they would tell you themselves if you asked them. But we are talking about something else here. Stable diffusion and models like it for art generation, yes, that AI can't provide stories by people. It is a tool used by people to bring out the stories they have to say or characters they want to show. I could make a comic out of AI images. It will look like crap currently, inconsistent in clothing, and maybe some scenes will look especially weird, but sure enough, those issues will be ironed out in time, but just for the sake of argument say we are a year ahead in time and it's possible. Wouldn't you admit that this is a story 'I' made, and brought to a picture format via the use of AI? How would it be any different if I drew the panels?(aside from taking credit for the art on them) Likewise, if an artist made pictures out of a story written by an AI like GPT-3, would you argue that the artist is inspired by the AI or that the AI made the artist a tool to bring its stories out?

My point is that what matters in this context is agency. If there is an active agent in that whole process, the credit will be given to the agent. AI, as you clearly said yourself, has no agency as of yet, so it's still the person giving the prompts the same way an artist prompts themselves when they draw. 'I want to make X character, in a pool, under the sun, hand over her eyes casting a shadow on her face, laying on a pink inflatable'. Those are the prompts and the artist will then generate an image based on those parameters. The only difference between that and using an AI, is that instead of creating the model for all that stuff in your own brain and developing the technical skill to apply it on a medium, you outsource that work to the AI. You are still in control of the creative part of the process and you need to wrangle the AI throughout the creation of that piece much like you'd need to learn techniques to draw it. One is just easier than the other and that's all there is to it.

darryus said:
also, holy fuck, comparing peeps who write prompts into an AI that was trained passively to musical artists that put days of actual fucking manhours into creating actual art is insane.

I used to make edm, it really doesn't take as long as you think if you have your ideas already set out. The hard part is the creative process of figuring out what you want to make. It takes away from the skill needed to learn each individual organ to the point where you can play complex melodies on each one of them. You could literally copy a midi file and make slight changes just enough to fool people into thinking it's original. Some might spot it and figure it out, but most will not care as long as it's a better banger than the other song and in the end, those people come out winning because now they have two similar songs they can enjoy. The comparison wasn't between AI and music artists, it was people using AI and music artists using program suites to create songs they otherwise don't have the technical skills to create. Skrillex could hire professional synth makers to make a new synth that has the sounds he specifically needs and then hire a performer to play it so he can record it. Or, he can use a program to do both of those things for the one-time price of that program by placing bars of varying lengths higher or lower on the screen and pressing a 'play' button. Sorry if I wasn't clear the first time. The focus isn't if AIs are people, the focus is if it's a tool that can be used by people. And it is. So it should be treated as such instead of shunned because it's different and new.

watsit said:
There's a reason many companies turn a blind eye to fan work, so if they happen to produce something similar they have plausible deniability. But if an AI knowingly uses some image as a source and ends up creating something substantially similar, even without directly copying pixels, that can still be infringement.

It wouldn't be. The way these NNs are trained, the information from the image is completely thermalized across the network, and it's literally impossible to reconstruct any of the images that went in. Beyond that, the total information from each image is in the range of bytes at most, and in many case fractions of a bit, it's just those fractions are aggregated over thousands to millions of images, and variations of said images. No part of any of the training images exist in a trained network. That information is literally not there anymore. You can show that mathematically beyond a reasonable doubt, which is a standard far higher then would be needed for a civil court. Any legal challenges would fail

As for getting a similar image from a well trained NN, it would be it's own image and not based off of training data because again, in a thoughtfully trained NN, that original data just isn't there anymore. It be like two artist working from a similar pose drawing a similar image.

I get why admins might be concerned about legal liabilities, but realistically, you'd have a massively greater chance of being sued/takedown'd by Nintendo for the Pokemon porn here before you ever got sued/takedown'd about an AI generated image.

darryus said:
you say that like it's going to be some massive exodus. like, I don't think most people see AI-generated images as much more than a novelty. we're humans, we like stories, we like art created by other people, substance, AI isn't able to provide that, and, until AI is actually sentient, it never will.

AIs are able to write short stories now, really well. You'll likely see a merger of these two ideas in the not to distant future to make images with stories attached, or possibly even full, multi-panel comics.

I get where Ckitt is coming from, I think they're a little bit too passionate here, and might be stretching things a bit. But it's only a bit. There will always be space for human generated content. But within the next 5 years you will see full AI images that surpass the vast majority of what you see on E621, and within 10 (maybe less) you'll see all forms of generated art (pictures, comics, even videos) produced better than most humans can make, and most importantly faster. People will navigate to where there's more and where they can get cheaper art created.

Once QPCs are in general use, QNNs will likely go far beyond even this. I don't even want to think about what happens then.

reallyjustwantpr0n said:
I get where Ckitt is coming from, I think they're a little bit too passionate here, and might be stretching things a bit. But it's only a bit. There will always be space for human generated content. But within the next 5 years you will see full AI images that surpass the vast majority of what you see on E621, and within 10 (maybe less) you'll see all forms of generated art (pictures, comics, even videos) produced better than most humans can make, and most importantly faster. People will navigate to where there's more and where they can get cheaper art created.

Once QPCs are in general use, QNNs will likely go far beyond even this. I don't even want to think about what happens then.

I am a bit annoyed still that the art of my characters was taken down due to "low quality" when 1. It had far better quality than most pictures here, enough so to fool people into promoting it to the daily popular, and 2. That there are AI-generated pictures on here of far worse quality that are allowed to stay. But rules are rules even if not applied unilaterally. I do understand the need for a framework on the site to avoid AI flooding and will gladly wait for that to happen. But sadly I don't see any active dialogues towards it, just outright alienation instead of trying to make the pieces fit together.

ckitt said:
I am a bit annoyed still that the art of my characters was taken down due to "low quality" when 1. It had far better quality than most pictures here, enough so to fool people into promoting it to the daily popular, and 2. That there are AI-generated pictures on here of far worse quality that are allowed to stay. But rules are rules even if not applied unilaterally. I do understand the need for a framework on the site to avoid AI flooding and will gladly wait for that to happen. But sadly I don't see any active dialogues towards it, just outright alienation instead of trying to make the pieces fit together.

if there's any ai-generated art on here it's either not approved or was approved by mistake.
also, getting on the today popular page isn't really impressive at all, you need like ~10 score. there's literal shitposts that got as much score as your posts before being deleted. and using the term "fooled" is probably not helping your case and neither is saying you're just using AI to avoid having to pay artists.

I think there's a deeper discussion to be had here, and I think there is an insane amount of frustrating denial about where things are headed.
So, I have a few pieces here on e621. I had to take a long break from art after just barely getting started due to life happening way too hard, but in my short time, I did a commission for over $100, I got one of my art pieces as one of the top 50 scoring feral jenna from balto pics on the site, and I've recently learned by searching the taglists that one (i know that's not enough to impact the model much at all) of my art pieces was used in the training of one of the more popular furry stable diffusion variants. I bring all that up to just say, even though I wasn't around for long, I don't think it's out of line to call myself an artist, or to state that I was intending to and capable of jumping back in and make some money doing it.

My main thing is that I think the current policy (here and elsewhere) is shortsighted, though I don't know what the best solution is. We've got a couple of months, 6 at the high end before these models start producing consistently coherent results. I think there is absolutely zero chance that every new professional concept artist working in the industry is not using AI as part of their toolkit within a year or so. That's a field that uses 3d, photobashing, painting, and photo texturing etc to crank out art in record time, and some of the most amazing and highly regarded art out there is by pro-concept artists. Why would a company hire someone who refuses to use AI and as a result is 1/50th the speed? Basically my point is that this phase of 'oh it's not real art', will be short-lived. Though I completely understand all the worries behind it from the booru being flooded by samey ai generated art, to worrying about peoples' art being used to train models. For what it's worth, I don't consider purely txt2img stuff on its own to make the person prompting an 'artist'. I'm not saying I'm above it. I enjoy making stuff solely with text, but while I do consider the output to be 'real' art. I think of it as the AI being the artist, and the prompter to be the art director or artist commissioner. Anything involving editing or img2img is somewhere in between.

but I want to show a quick example of something here:
this was a quick test and shows a small part of one of my pieces (original on the left). It was a particularly rushed piece and never really made it past a sketch phase, but I was done with it and I still kinda liked the way it came out. I ran it through an img2img stable diffusion model, and it completed my sketch into a full painting for me. It took it's own creative liberties and by tweaking the settings I can decide how much liberty it takes, and I can also go back into photoshop and tweak what it does and back and forth until I'm completely satisfied. As an aside I would absolutely want to tweak these if I was doing this forreal. Neither is the style I would be going for. I was just seeing what it could do.
NSFW: https://cdn.discordapp.com/attachments/1031880246624010280/1031880287426199562/wolfX3.png

I'm confident I could make it to the middle one on my own today, though due to my inexperience it would take me many many hours. I'm also confident I could develop the skillset to get the style on the right one, though it would take me at least a month or two to get there practicing that style, depending on how much time I had to practice.

Point is, I would have to be completely nuts to not take advantage of this time saving ability or artistic super power. Also editing AI outputs to tweak the direction or fix mistakes will also help me become a better artist. Now I'm not claiming that everything people make with AI, nor even everything I'd want to make with it will start out with a colored and lighted sketch like this, but as time goes on, and I don't mean years from now, it will be completely impossible to tell the difference between AI generated art and human made art. And more and more abilities to add one's own artistic flair to the output will be increased. Even now if I finished a few more pieces I could train with textual inversion or dreambooth and have it learn my particular style and apply that.

Where will the line be drawn? Will professional level artists have to abandon their toolset when posting to e621 or fa (or anywhere else it seems)? Will everyone just have to lie when they use AI? I don't know what the solution is, but everywhere is adopting a zero tolerance policy and people are exaggerating the "art theft" aspect of it. Though I absolutely agree with the putting artists out of the job thing.
For the record, I will not be uploading AI generated or AI assisted art here or anywhere under this pseudonym, I will use an alternate name and only post where it's not banned if I start to do that.

ckitt said:
I am not worried about a 'massive' exodus, I just think it's bad to want the community to split due to the removal of a valid expression of art regardless of your personal view of it.

It was never accepted in the first place, so nothing's being removed that used to be allowed here.

The Uploading guidelines says:

Bad things to upload:

• Low quality submissions: ... AI / neural network edits, AI generated artwork, etc.

Not every site needs to cater to everything; just as we don't accept stories, or audio work, or sculptures, or fursuits, or photos, without also having artwork created by a person, we don't accept AI generated artwork without it also having artwork created by a person

reallyjustwantpr0n said:
No part of any of the training images exist in a trained network. That information is literally not there anymore.

That doesn't matter. Step A: Have access to the work, Step B: Produce something substantially similar to the work. That's copyright infringement. It doesn't matter how you go from A to B.

pandorasrabbithole said:
Where will the line be drawn? Will professional level artists have to abandon their toolset when posting to e621 or fa (or anywhere else it seems)?

No one is saying you can't use AI as a tool in creating artwork. But you still need to put in enough of your own work to make the artwork yours, rather than just what the system spit out with little (if any) touch-ups.

AI's an interesting animal, I will say that. I toyed around with DALL-E to create some amusing images but it's not the same as a customized art piece drawn by a real human being. I am a little scared about the future of AI art generators but there has to be an upper limit somewhere. How close can they get to replicating an artist's work? Because if they managed that, it could potentially devalue the work of many artists in the fandom.

watsit said:
No one is saying you can't use AI as a tool in creating artwork. But you still need to put in enough of your own work to make the artwork yours, rather than just what the system spit out with little (if any) touch-ups.

do you mean that's the official policy of uploading here? or do you mean that that's your own philosophy for AI art.
Also the example I posted, would that fall under making it my own?

pheagleadler said:
AI's an interesting animal, I will say that. I toyed around with DALL-E to create some amusing images but it's not the same as a customized art piece drawn by a real human being. I am a little scared about the future of AI art generators but there has to be an upper limit somewhere. How close can they get to replicating an artist's work? Because if they managed that, it could potentially devalue the work of many artists in the fandom.

I follow AI pretty closely. There is no upper limit. And we are currently in the MS paint phases of what would eventually become photoshop. Only in the world of AI years of progress are shrunken down to months or weeks.

Also, if you wanted to you could make an imitation of anyone's style today that would fool most people by finetuning a model.

pandorasrabbithole said:
do you mean that's the official policy of uploading here? or do you mean that that's your own philosophy for AI art.

It's the policy:

NMNY said:
Using AI for supplementing regular art, like helping with backgrounds or as pose references, are fine, but we are definitely drawing the line at people either uploading unmodified AI artwork, or compositing AI works and applying mild touch ups.

post #3590930 is an example of a piece that started with an AI generated image, and used it for a pose reference that they redrew in their own style, changing a number of things, fixing up obvious issues, and swapping the character for a different one.

watsit said:
It's the policy:
post #3590930 is an example of a piece that started with an AI generated image, and used it for a pose reference that they redrew in their own style, changing a number of things, fixing up obvious issues, and swapping the character for a different one.

awesome. thanks for the info

watsit said:
It's the policy:
post #3590930 is an example of a piece that started with an AI generated image, and used it for a pose reference that they redrew in their own style, changing a number of things, fixing up obvious issues, and swapping the character for a different one.

oh sorry, one more thing. do you think the example I showed would pass?
seems different since the AI added detail that would be kept in the final, though I'd have to make more edits to the full pic.

darryus said:
and using the term "fooled" is probably not helping your case and neither is saying you're just using AI to avoid having to pay artists.

The coping here is insane. I said 'fool' to dismiss the notion that the art is 'low quality'. And isn't the argument you guys have 'if you don't want to pay for art make it yourself'? Now suddenly if we make art ourselves using whatever tool is available you STILL get to whine? Nah, I think you got other issues and it isn't the AI or people like me. Are you saying that if I didn't have that option I would pay for artists? Because if that's your take, you already assumed too much and you are dreadfully wrong. No one is losing money from me as I wasn't planning on buying.

Try and argue the points instead of going for stupid gotchas. What issue do you have against AI art that is rational?

edit: What I do get as a point is people 'commissioning AI art' because it will be cheaper and easier to get but then you'd be admitting that AI is the same quality as 'real' art. Did germany stop producing power tools when cheap Chinese knockoffs came around? No. It still serviced people that wanted 'quality' and still made a banger economy out of its industry. Normal art isn't going anywhere, in fact most normal artists will be using AI along with their skills to produce better pictures as soon as these models become mainstream enough to cater to more niche communities. I wonder if then you'll stop consuming art altogether and just hang around in museums remembering the 'good old days'.

I don't understand why some people insist, that e621 should host AI "art". Just use r34 for it if it is that important. They'll probably give a shit about it, as they usually do with other stuff.

watsit said:
It was never accepted in the first place, so nothing's being removed that used to be allowed here.

Not every site needs to cater to everything; just as we don't accept stories, or audio work, or sculptures, or fursuits, or photos, without also having artwork created by a person, we don't accept AI-generated artwork without it also having artwork created by a person

Last time I checked AI images fell under the category of digital artwork so your categorization of it as a separate entity because it is AI generated, is your opinion and nothing concrete or based upon anything but your personal feelings on the subject. Also, the rule about AI generations was made when AI art looked like someone took a dumb and smeared it on a canvas, hence why it was placed under 'low quality'. And, aren't forums a place for the community to discuss issues they may have, or are we supposed to just sit silently and watch when things that interest us get an unfair treatment?

watsit said:
That doesn't matter. Step A: Have access to the work, Step B: Produce something substantially similar to the work. That's copyright infringement. It doesn't matter how you go from A to B.

So, all artwork ever created is copyright infringement, because that's practically the level of 'substantial similarity' you are talking about here without realizing it. You clearly need to do some reading on the subject before attempting to use this point. There are no grounds for copyright infringement on AI and there never will be, stop using that really, really ill-informed point so the conversation can progress further than senseless issues that stem from a lack of education on the subject.

Pandora made a great point about AIs use becoming more prominent, maybe not in 6 months or a year, maybe in 5 years but whatever the timeframe may be, what's your argument against building a framework on the site to host that type of art? And I mean actual argument, not your personal feelings on it.

watsit said:
No one is saying you can't use AI as a tool in creating artwork. But you still need to put in enough of your own work to make the artwork yours, rather than just what the system spit out with little (if any) touch-ups.

Define 'enough' in this context. How many hours or minutes of your time should be put into a piece of art to be considered yours? Does that go for every skill level? Should a veteran artist take the same time as a newbie artist? How many parts of the picture need to be changed? Is the background okay to be left alone if someone made the whole character? Is it okay to leave the character if the person made an intricate background? Is it okay if it just colors lineart? If it just shades it? Is it okay if the picture is already fully drawn but just enhanced? How much enhancement is okay? Can the enhancement cause significant changes to the character or setting? How much micromanagement do you think you can put up with before you start laughing at the term 'artistic freedom'?

Maybe the easier way is to instead of managing the people trying to create new content, you let them loose and manage the way that content is vetted and held up to standard. Isn't that what people are after in the end? Having quality pictures for their viewing pleasure? It's getting people what they want and making everyone happy in the end. You seem to have a mentality that someone must get the short end of the stick when it doesn't even have to be this way. What gives? Why wouldn't the admins consider changing the rules in the face of new information?

dubsthefox said:
I don't understand why some people insist, that e621 should host AI "art". Just use r34 for it if it is that important. They'll probably give a shit about it, as they usually do with other stuff.

I guess it's the familiarity and if the pictures are furry-oriented of OCs, I am not even sure if r34 would accept that kind of flooding. It seems like you just want to pass the bucket. What I don't get is why it can't be properly tagged here and people that don't like it can just blacklist it, or have its own category so it doesn't mix with normal art. I mean, countless of solutions, only 'real' reason I see is lack of will to even discuss it in a serious manner without it devolving to 'if you don't like it leave' instead of a good rational explanation. You don't see how that is off-putting?

ckitt said:
I am not even sure if r34 would accept that kind of flooding

That's funny. They are literally ripping all artworks from e621 with a bot.

ckitt said:
why it can't be properly tagged here and people that don't like it can just blacklist it... only 'real' reason I see is lack of will to even discuss it in a serious manner without it devolving to 'if you don't like it leave' instead of a good rational explanation...

The answer to that is:

• Moral
• The believe in true art. ^{_{(I know it's a little dramatic expressed)}}
• Possible legal issues.
• And the just simply don't want AI art here.

The thing is, I haven't seen a single moderator, janitor, active user in the forum, that said:"yes I want to see a giant flood of AI art on e621" Or maybe I am just blind for the other comments. Those who are active here and want to do more than just having their daily evening fap, have pretty much the same opinion somewhere between "ehh" and "no" ^{_{(from my observation, prove me wrong. I have to admit I skipped through the thread)}}