discussion About IA generated art

[momich]

TLDR: AI is not going to affect the big art, its going to mess up the small creators.

 

Hi, first of all I'm a hobby artist, mech ing student and a big LMG fan. i wanted to comment on linus take on AI art, not from a "U WORNG" but more on a informative stand point and for discussion.

To get a better perspective I'm gonna re frame the issue, instead of art lets talk games since you guys are more familiar with that.

Lets say someone makes a AI game generator for free or very little money, the AAA game industry probably is not going to care about that because it stands on big names, big values, stablished fanbase and blablabla but.

What about small creators? Lets say that you wanna play Super Meat Boy and instead of buying it you go to the AI and enter "SMB platformer" and the ai gives you a incredibly similar copy of that, after playing a bit of it would you still boy SMB? i mean yeah if you want to support the creators but if everyone was like that we would not have the boat website. it would bite into the market of SMB and worse, Would it be okey if the AI was train with SMB as an input? i mean probably completely legal, since every shooter has the same mechanics and nobody thinks of that as Plagiarism but what about the AI that directly uses parts of the material that got sampled to train it and it can absolutely copy what makes the game itself (for art it would be the style, try it just type in an AI generator x thing in y artist style lol). How do u think your steam page would look like if 12yr Timmy could upload his "fallout but in space™" game for 1 dollar? the indie scene is already fill with games making it hard to stand out. And so and so.

Its a very deep and complex issue but its not all doom and gloom. Something like that could put in equal ground little Timmy and someone that knows how to program, making dreams easier to archive.  The true difficulty of the game making process wouldn't be on budget or knowledge, it would be creativity. Sounds good but in an economic system, What's the cost of exponentially growing the supply?.

All that applies to art but even more directly. Like, how do you see art? probably as an superficial commodity, something you just hang and looks good but for some its just way more than that, just like that one game that changed your life and your mom still calls super mario.

 

Sorry for bad English lol.

[Vishera]

12 minutes ago, momich said:

ia gives you a incredibly similar copy of that

Sounds like a copyright violation, the IP holder can sue.

 

[Roswell]

Sidestepping the enormous gap in complexity from 2D images to games…

 

What are you going to feed it? 99.999% of games aren’t open source.
 

You can’t feed it video because there would be no distinction between positive and negative gameplay experiences without an army of reviewers flagging and clipping millions and millions of videos into “good, engaging, and fun” vs “speed running, negative, memes, etc”.

[momich]

1 minute ago, Roswell said:

Sidestepping the enormous gap in complexity from 2D images to games…

 

What are you going to feed it? 99.999% of games aren’t open source.

yeah you are right, you cant compare the complexity of the 2, but that's not the point. The comparation is more of a translation form art to gamers

Oh and 

this is old, yeah rough but remember ia generated images from 10 yr ago? lol. 

[momich]

13 minutes ago, Vishera said:

Sounds like a copyright violation, the IP holder can sue.

 

yeah thats kinda the point 

[Vishera]

3 minutes ago, Caroline said:

I work with AI and can barely get it to modify a picture

Through my A.I endeavors i managed to create a painting that looks like it was painted by hand, and a really beautiful painting too.

It depends on the algorithm you use.

[TetraSky]

Every jobs are at one point or another going to go extinct as we progress toward the future.
It's sad to say, but I for one like AI generated arts a lot more than some of the stuff that passes for "art" these days.

[momich]

First, im so glad to have your perspective, i really think that ia is amazing.

6 minutes ago, Caroline said:

The same as now? most games on Steam are utter bs.

yeah, that but x10.

7 minutes ago, Caroline said:

here's an universe of difference between generating a picture out of GBs of samples and a game

the comparations are not empirical, is more illustrative, im not gonna pretend that everyone's is an art nerd. but as you may know "we are years from that point" is a when not an if, it will come and we will face it just think of the difference between the computer we took to the moon to an ipod.

 

16 minutes ago, Caroline said:

I'm more interested in using AI to power robot maids with optional LED cat ears and enhanced hydraulics/armor than some silly games 

same bro, same.

[momich]

12 minutes ago, TetraSky said:

Every jobs are at one point or another going to go extinct as we progress toward the future.
It's sad to say, but I for one like AI generated arts a lot more than some of the stuff that passes for "art" these days.

yeah i mentioned that im an Mech ing student for that, my job would be literally just replace manual labor with mechanical one but it shows that you dont remove an employee, you just move his position to a higher one that needs more expertise, the issue is more like if its going to be like that or if we are just gonna destroy the actual people there with something morally dubious.

[Roswell]

27 minutes ago, Caroline said:

I work with AI and can barely get it to modify a picture

Plenty of AI image tools exist and are part of professional workflows. AI modified images are literally everywhere. Magazines, billboards, signage, sites, etc etc.

[Eigenvektor]

1 hour ago, Caroline said:

Nowadays even using the toilet is a copyright violation because someone might have patented dumps.

Just wanted to point out that copyright and patents are not the same. You could try to get a patent on "a method of removing concrete waste from a person's body". If that is granted you could then sue for patent infringement if someone uses the same method. Of course they could point out there's prior art and you'd more than likely lose it again.

 

That's not the same as copyright which is mostly automatic and does not need to be filed like a patent does. When you create a piece of art you have copyright on it and can sue other people if they create art that is more or less a carbon copy of yours.

 

As far as games go, (depending on where you live) you could maybe patent the concrete implementation of an algorithm, but not the algorithm as such. So you could potentially get a patent on a super efficient path-finding algorithm, but you can't patent path finding algorithms as such.

 

Copyright on the other hand will apply to any in game assets such as in game art (splash screens, loading screens), textures, sounds, …. If you train your AI on existing games you'd very quickly land in hot water if it creates copies of these assets in your own game.

 

2 hours ago, momich said:

Lets say that you wanna play Super Meat Boy and instead of buying it you go to the AI and enter "SMB platformer" and the ai gives you a incredibly similar copy of that, after playing a bit of it would you still boy SMB?

That's really not how AI works at this point in time. It has no understanding of what "a platformer" is. You wouldn't be able to tell it "go play Duke Nukem" to get it to recreate that game (or a similar game), because it doesn't have human level understanding of what "play" even means. There is no Artificial General Intelligence yet.

 

You need to train neural networks on a huge data set for thousands of hours to be able to reliably recognize images that contain a cat. AI is mostly about pattern recognition and about recreating patterns (which can be used to "create art"). But that's not the same a creating an entire game which needs far more than a single image to be playable and fun.

 

The AI in the video above may be able to recreate what you see in GTA when you press keys. As impressive as that is, it will not be able to create new missions for you to play or anything it hasn't encountered before (e.g. "random" encounters). You could maybe use an AI to more quickly generate a framework in which you can then implement your game, which will still require creative input. However you'd quickly run into the issue that you can't exactly "fix bugs", because you can't exactly debug/modify the NNs output by hand. You would have to retrain it (for thousands of hours) instead.

 

You can see that they're using only a very small part of the game. If you want it to learn/recreate the entire map, you'd need to train for a lot longer. Now add dynamic day and night cycles and you'll need even more time. Most likely that would be much more expensive than creating a game in an existing engine by hand (at least for now)

[LAwLz]

The "anti-AI-art" crowd is just the latest group of Luddites.

People have always been against technological progress when it threatens their jobs, regardless of what their jobs were. 

In 1779 people were against knitting machines that created textile fabrics.

In the 1800's people were against cameras.

In 2022 people are against AI generated images.

[05032-Mendicant-Bias]

5 hours ago, momich said:

Lets say someone makes a AI game generator for free or very little money, the AAA game industry probably is not going to care about that because it stands on big names, big values, stablished fanbase and blablabla but.

We are currently very far from that, I'd rekon decades away. ML models are still quite narrow. You can make ML to do chatbots, ML to translate description to images, to write a description of an image, fold proteins, route PCB tracks and lots more.

I'd argue to get AI to write a good game, you need to get very close to artificial general intelligence. Making something a human perceive as a good game is a lot more abstract than creating an image a human perceive as a good image.

 

2 hours ago, Eigenvektor said:

Copyright on the other hand will apply to any in game assets such as in game art

What I understand of Copyright law is anachronistic even for today's use. I find it stupid that copyright of a work last seventy years after the death of an author, a medical patent last twenty years, and an IT patent lasts twenty years (in IT technology can become obsolete after two years).

Right now I think AI generated content has no copyright at all, Legal Eagle (lawyer and youtube creator) made a video discussing the topic:

 

4 hours ago, TetraSky said:

Every jobs are at one point or another going to go extinct as we progress toward the future.
It's sad to say, but I for one like AI generated arts a lot more than some of the stuff that passes for "art" these days.

Right now and for the foreseeable future AI generators are going to automate stock image generation. I use DallE to make scenario rapresentations for my D&D campaign. I would never have contracted an artist to make pictures I use once in a game between friends anyway.

I think any image used for professional purposes is still going to require professionals. Just like putting cameras in everybody's pockets didn't eliminate the need for professional photographers.

[Roswell]

9 minutes ago, 05032-Mendicant-Bias said:

AI generators are going to automate stock image generation

 

9 minutes ago, 05032-Mendicant-Bias said:

any image used for professional purposes is still going to require professionals.

 

These statements are mutually exclusive. Stock photos are mostly used by professionals. Amateur use is a small fraction of the market.

[Eigenvektor]

3 minutes ago, 05032-Mendicant-Bias said:

What I understand of Copyright law is anachronistic even for today's use. I fijnd it stupid that copyright of a work last seventy years after the death of an author, a medical patent last twonty years, and an IT patent lasts twenty years (in IT technology can become obsolete after two years).

That's an entirely different discussion. For the time being, recreating a game will (most likely) violate copyright law. Whether that is fair or should be the case for games that are 5, 10 or even 20 years old is a different topic. I only brought up patents because Caroline was intermixing copyright and patents.

 

Aside from that, if the technology someone created is obsolete within two years, holding a patent on it that lasts for twenty is somewhat meaningless?

 

3 minutes ago, 05032-Mendicant-Bias said:

Right now I think AI generated content has no copyright at all, Legal Eagle (lawyer and youtube creator) made a video discussing the topic:

He does mention that art created by a machine has no copyright because only a human can hold copyright. That's not the same as "AI generated images cannot infringe on copyright" though.

 

As I said above, a game and/or its assets are protected by copyright law. If you recreate that game, whether manually or assisted by an AI, you are violating their copyright. Which means, more likely than not, you will get in legal trouble if you tried to publish it.

[Arika]

i will say it was amusing to see furries losing their shit over novelAIs furry stablediffusion model.

 

As a hobby artist myself and someone who has been delving into AI generated content for a while, i love where things are going.

 

Why not use AI for inspiration for your next piece? whether that is a text prompt, or a full image.

AI doesn't have to replce you, embrace it and use it to your advantage.

[05032-Mendicant-Bias]

1 hour ago, Roswell said:

These statements are mutually exclusive. Stock photos are mostly used by professionals. Amateur use is a small fraction of the market.

Professionals aren't going to sell stock images. Stock images are a cheap way to get to a concept and used in early stage concept by professionals. AI generation is going to be a step in the creation of the professional content.

 

Amateur and casual users can instead use AI generation instead of searching a good picture with google image search.

 

1 hour ago, Eigenvektor said:

As I said above, a game and/or its assets are protected by copyright law. If you recreate that game, whether manually or assisted by an AI, you are violating their copyright. Which means, more likely than not, you will get in legal trouble if you tried to publish it.

I do not understand the legal side, I wonder what the threshold fo violation is. I once read: "Copy from one and it's copyright theft, Copy from one thousands and it's research"

 

If you ask AI to replicate something with very minor violations, and try to profit from that, that should count as copyright violation.

 

If you ask AI to do something new, you sell it, and someone make a copy and sell that, you are not protect by copyright.

 

That's how I understand it, but I might be wrong.

[Roswell]

15 minutes ago, 05032-Mendicant-Bias said:

Professionals aren't going to sell stock images.

What? Stock images are very often shot by professionals, it’s a huge source of revenue in the photography world.

 

18 minutes ago, 05032-Mendicant-Bias said:

Stock images are a cheap way to get to a concept and used in early stage concept by professionals.

Again, what? This is not what we use stock images for. Stock images usually end up composited with other elements, edited, or simply dumped as is into a project.

 

If I (or anyone else in my field) needs inspiration or a piece for a mock-up/concept, we’re not typically paying for a stock image since it’s never going into production.

[Eigenvektor]

19 minutes ago, 05032-Mendicant-Bias said:

If you ask AI to replicate something with very minor violations, and try to profit from that, that should count as copyright violation.

I assume you mean "variations" in the first part. Otherwise, agreed.

 

19 minutes ago, 05032-Mendicant-Bias said:

If you ask AI to do something new, you sell it, and someone make a copy and sell that, you are not protect by copyright.

Not sure how true that is. While an AI as such can't hold a copyright, its human creator or operator may very well hold a copyright on things they have "instructed" it to create. I don't think a company would go "Oh hey, this is public domain btw, because our AI created it for us rather than a human employee". I guess we'll find out once the first of such cases hits the courts, i.e. someone replicates AI created art and the company that operates the AI sues them.

 

3 minutes ago, Roswell said:

Again, what? This is not what we use stock images for. Stock images usually end up composited with other elements, edited, or simply dumped as is into a project.

From what I've seen, our marketing department uses stock images exactly they way you describe. They're used for ads and also as part of illustrations or documentation everywhere. They do often add variations to it though, to make it appear less "stocky".

[jones177]

 

Some see new tech as a threat and others embrace it. 

 

As for its use in games I see it as an aid in random generation. Artists would have to do the base work and then the AI would generate the variables. So no NPC would look alike and no building would look alike. Nothing would look alike unless they were supposed to. 

 

A "fallout but in space™" would be possible with unique aliens ,architecture and scenery on every planet.

It will be a literal game changer.

1 hour ago, Eigenvektor said:

From what I've seen, our marketing department uses stock images exactly they way you describe. They're used for ads and also as part of illustrations or documentation everywhere. They do often add variations to it though, to make it appear less "stocky".

I did work for a company that used stock images to depict me, the design team.  So the frumpy nerd became Chad in a suit. 

[tikker]

8 hours ago, momich said:

the indie scene is already fill with games making it hard to stand out. And so and so.

There's another side to this though: it's hard to stand out because so many are doing the same thing. Once a concept gains popularity a thousand copies will suddenly spawn. Think wave shooters in VR, battle royale stuff when that was the bomb, the interest of being like dark souls. I think the  potential problems that we see with AI-generated stuff isn't unique to AI. It has alwasy been there and is amplified by it, truly showing you that it takes skill and effort to make something that stands out.

8 hours ago, momich said:

TLDR: AI is not going to affect the big art, its going to mess up the small creators.

So it'll be like any technological advance.

7 hours ago, momich said:

yeah i mentioned that im an Mech ing student for that, my job would be literally just replace manual labor with mechanical one but it shows that you dont remove an employee, you just move his position to a higher one that needs more expertise, the issue is more like if its going to be like that or if we are just gonna destroy the actual people there with something morally dubious.

What makes it morally dubious in your eyes? Human art functions quite similar to it in my opinion. Most if not all of the art one creates is inspired by "training" on other people's work or experiences you had. Let's say you enjoy the cel-shaded art style and also love pixel art. Now you combine the two to make cel-shaded pixel art works. Have you created a new, interesting, unique art style or are you just "ripping off" other peoples work from either department? Is doing this ourselves different from asking an AI to do it?

 

As was mentioned earlier, I think the need for professionals will remain. Practical situations will require such experience to judge what is needed and make the right calls. What AI will change I think, is that the simple and cumbersome things will disappear or lessen. Instead of needing to commision a pink shark with leopard spots, one will just ask an AI to generate it and go from there. I mentioned rotoscoping in another thread about this as another example. Corridor did a video about this. Some people's job is rotoscoping frames all day. With AI(-assisted) rotoscoping they will be "out of a job", but instead can perhaps focus on the real job more, which is editing the movie and use their expertise more effectively by only having to guide the AI instead.

3 hours ago, 05032-Mendicant-Bias said:

If you ask AI to replicate something with very minor violations, and try to profit from that, that should count as copyright violation.

It probably can, since you don't own the rights to that artwork. Fan art is in the same position AFAIK. You don't own the rights to that material, so without permission of the original owner you have no right to reproduce and profit from it from a copyright perspective.

3 hours ago, 05032-Mendicant-Bias said:

If you ask AI to do something new, you sell it, and someone make a copy and sell that, you are not protect by copyright.

If it's new, why wouldn't you? You initiated the creation of it, so while the AI cannot copyright it, I think you should be able to get copyright for it if it's really something new.

[Kisai]

We've had this conversation a few times on the forum.

 

On 9/4/2022 at 7:49 PM, momich said:

TLDR: AI is not going to affect the big art, its going to mess up the small creators.

 

Hi, first of all I'm a hobby artist, mech ing student and a big LMG fan. i wanted to comment on linus take on AI art, not from a "U WORNG" but more on a informative stand point and for discussion.

To get a better perspective I'm gonna re frame the issue, instead of art lets talk games since you guys are more familiar with that.

Lets say someone makes a AI game generator for free or very little money, the AAA game industry probably is not going to care about that because it stands on big names, big values, stablished fanbase and blablabla but.

The AAA industry would love it if they could fire everyone and produce procedurally generated games from procedurally generated assets at the drop of a hat. The problem with that line of thinking is that "you ain't fooling me", why would I spend money on the garbage that is no better than monkeys with typewriters. No seriously. http://news.bbc.co.uk/2/hi/3013959.stm

 

AI, in it's current generation is basically learning how to "copy the style" by throwing random noise at the wall and comparing it to the ground truth. If you over expose it to one specific image, phrase, or audio sample, it will learn how to do that sample perfectly. It will however not be able to create a variation on it. Case in point, a TTS I trained, learned to say "what the f***" because multiple subjects said it exactly as "what the f***", and thus the TTS learned to say it exactly how an english speaker would say it, and between all the subjects in the corpus, when it says it, it sounds like a blend of three of them. Now if I change the word to "truck" or "fork", it doesn't sound like it at all.

 

On 9/4/2022 at 7:49 PM, momich said:

What about small creators? Lets say that you wanna play Super Meat Boy and instead of buying it you go to the AI and enter "SMB platformer" and the ai gives you a incredibly similar copy of that, after playing a bit of it would you still boy SMB? i mean yeah if you want to support the creators but if everyone was like that we would not have the boat website. it would bite into the market of SMB and worse, Would it be okey if the AI was train with SMB as an input? i mean probably completely legal, since every shooter has the same mechanics and nobody thinks of that as Plagiarism but what about the AI that directly uses parts of the material that got sampled to train it and it can absolutely copy what makes the game itself (for art it would be the style, try it just type in an AI generator x thing in y artist style lol). How do u think your steam page would look like if 12yr Timmy could upload his "fallout but in space™" game for 1 dollar? the indie scene is already fill with games making it hard to stand out. And so and so.

 

AI could only help a small creator. Instead of needing a million dollar budget to produce all the assets, or even the game code itself, you could pretty much ask the AI to produce a game, and give it ever so specific details, and then you clean up the rest. It does not replace the need for artists or programmers, because the AI has no creativity. It can only do what it was trained to do, and nothing more. It can not think outside the box. AI has in fact shown us our worst intentions, when you train it for NLP (Natural Language Processing), where it will pick up all the racist, sexist, and other bigoted language, and various other micro-aggressions. Like even having an AI learn from ONLY public domain resources is bound to expose that in the worst way, because many of the PD sources like Alice in Wonderland and The Wizard of Oz were produced in a time when it was "OK" to mistreat minority groups. To say nothing of what the underlying premise of "Alice's Adventures in Wonderland" actually is (nonsense.)

 

For NLP to reflect our current, more "PC" culture that is going to be offended by any maliciously off-color language, we can't simply let AI learn from our past, it has to learn from our present as well.

 

On 9/4/2022 at 7:49 PM, momich said:

Its a very deep and complex issue but its not all doom and gloom. Something like that could put in equal ground little Timmy and someone that knows how to program, making dreams easier to archive.  The true difficulty of the game making process wouldn't be on budget or knowledge, it would be creativity. Sounds good but in an economic system, What's the cost of exponentially growing the supply?.

All that applies to art but even more directly. Like, how do you see art? probably as an superficial commodity, something you just hang and looks good but for some its just way more than that, just like that one game that changed your life and your mom still calls super mario.

 

Sorry for bad English lol.

 

AI has a better chance of producing a novel, film, or video game that looks like Alice in Wonderland than it has of producing anything coherent to a human. AI learning , as it stands right now, requires further and further specific details to get anything that starts to resemble what a human might call "artistic", but the AI does not possess artistic intent. It does not interpret. It just reads the NLP input and looks inside it's database of where these words are cross linked, then looks at it's "samples" of those words to create an image, or a text, or audio. This is why TTS voices only sound human. What do you think would happen if you trained an AI to learn how to speak animals languages? https://www.popsci.com/technology/artificial-intelligence-animal-language/ 

 

People are all too quick to villainize AI, that it will take jobs, when it's really the more mundane, time-consuming efforts that might get replaced. Basically an AI can be a "concept artist" for people who aren't artists to communicate to actual artists what they want. 

 

Many artists get (rightly) upset with human clients that are too nitpicky, or can't communicate effectively what they want, and try to micromanage the details. An AI will have no such complaints, and can then in fact show them what those micromanagement's might be seen as.

 

But it's still throwing spaghetti at a wall. No two identical prompts will produce exactly the same image, text or sound. Even two versions of the same model will not. We're at the point right now where we are giving the model too much freedom to learn from internet sources, when we should be curating models to produce "correct" input. 

 

For example. Look at the LAION project, which is what all the image generators are learning from, and look up "moon"

 

So you have, 5 images of the actual moon, taken on cameraphones, one painting, two comics, one image that might be the moon, a scribble, and cartoon banana-shape crescent moon. So when you type just "moon" into DALLE or something, it will likely guess at the context.

 

So between all of these, none of them are a full moon. With the one in the bottom-right likely being a photo-of-a-photo

 

This is what would happen if you tried to use AI to design artwork, let alone a game. It will take ONLY what it learned, and have no idea how to compose a scene that makes sense to a human, only what it learned contextually from it's learning sources.

 

https://www.theverge.com/2022/9/5/23337580/openai-dall-e-text-to-image-generator-outpainting-native-function

 

Case in point.

So the AI somehow figured that the "dogs playing poker" must have been in a much larger room with what looks like more dogs at a table playing a game as well, and to the left another poker table with yet another dog. This is actually not the first time we've seen this technique in AI.

 

https://openai.com/blog/musenet/

 

This is the exact same idea, except with music.

 

AI is still at a point in time where it doesn't create anything, it merely learns to mimic a style, or things it's been shown/things listened to/things read, it has no creativity of it's own. It's the Chinese room argument still. It does not understand what it is doing. It's just really fast plagiarism at worst, and really fast "scratch asset" generator at best. At least when it comes to trying to make it create things.

 

As for copyrights. Copyrights clearly can be had on the training corpus. It can not be had on the outcome. That's already been ruled.

https://www.copyright.gov/comp3/chap300/ch300-copyrightable-authorship.pdf 306, 313.2

 

309:

Quote

As discussed in Sections 304 through 308, a work may be copyrightable (i) if it is eligible for copyright protection in the United States, (ii) if the work has been fixed in a tangible medium of expression, (iii) if the work was created by a human author, (iv) if the work constitutes copyrightable subject matter, and (v) if the work contains at least a minimum amount of creative authorship that is original to the author

 

AI fails point 3 automatically, and fails point 5 if you merely try to copyright the output of the AI. There is no creative authorship made by AI models. They do not possess any ability to create something they were not trained on, which means anything they produce is at minimum, a derivative work. While something produced by AI may not be copyrightable, that doesn't mean you can't transform it into something that is.

 

Which leads back to the idea of producing a game by AI. If I, "A human", decided to let an AI do only the art and music, but I write the scenario and use an off-the-shelf game engine such as Unity or Unreal. Am I entitled to authorship? Am I entitled to all the money made when the AI model obviously has learned from other copyrighted and non-copyrighted sources? This is a question that we're kind of left in the lurch about.

 

I would, at best, use AI to create the art assets that are not tangibly important. eg, if I'm making a "flappy bird", then the background itself is of no importance, only the character sprite, and the obstacles. So I would likely hire someone to produce "the bird" and "the obstacles", but all other art aspects of the game, the background, the music, could just be AI generated, and if someone were to insist that "the AI stole my artwork, used in your game", I could remove the claimed asset and replace it with another. That's the safest use.

 

But think about a film or a television show. Even today, shows that were made 30-50 years ago, are routinely edited to remove licensed music that was used in the 70's, 80's and 90's because the DVD or streaming site can not (or will not) negotiate a licence. So there's an option on the table. Run the original music through something like musenet, delete the original music, punch it up a bit, so it still fits the scene. Or just offer two sound tracks. One with the audio in dispute removed, and one with the audio using the AI produced filler music. Is that copyright infringement? No. But it may be a derivative work of a copyrighted material that should not itself alone be considered copyrightable.

 

Honestly, when you look at the big picture, we are actually no closer to star trek holodeck's than we were 20 years ago. All that has changed is the computational power, so we can take bigger shortcuts with AI. What would take weeks to render at 480p in 1999, can now be rendered in real time, raytraced, radiosity, caustics, etc in real time on a high end GPU in 4K. The same has happened with AI, so TTS systems and computer vision systems of 20 years ago were abysmally awful. Companies that sprang up to do TTS systems were so over-protective of their technology that, it literately got adopted by absolutely nobody, out of the fear that they would lose money if someone were to pay for their product once and then produce limitless output. That didn't happen. It wasn't until audio ebooks became a thing, that that fear would come true, and people still prefer a real human to read a book to an AI that that can not make any distinguishing voices or emotional impact. We've improved the "output" quality and speed without doing very much to make AI's actually understand what they are doing.

 

 

 

[Neroon]

The first thing I thought when I saw AI generated art, was like this is gonna be fantastic for board and cardgames. So many cards and game pieces could simply use this art, to make whatever you need. And yes that does affect artists and I will say this, the effect is arguably larger then we might think.

 

If loads of artists are no longer needed for this, people will quit that line of work, and others might never get good at it, because there is so little money in it.

 

I would compare it with photography, everyone can take pictures, but making a good picture is something different. So how do people learn? Obviously through a hobby, but they can increase the quality really quickly if they get paying assignments, buy better glass etc. To be clear, I'm not saying the quality of the products is the most important part, not at all, but all of it contributes to people sticking to it, doing more with it, learning etc.

With photography most people already get out of it, because there is so little money in it. Now with art, there will be a decent chunk of money that disappears, because why pay someone for art, when you can generate it.

[Kisai]

7 hours ago, Neroon said:

 

With photography most people already get out of it, because there is so little money in it. Now with art, there will be a decent chunk of money that disappears, because why pay someone for art, when you can generate it.

Wedding photographers are still a thing. School and Graduation photos are still a thing. "Old Timey" Photos are still a thing. Photos taken on rides at amusement parks are still a thing.

 

What has happened with photography is that "the camera" replaced "the painter" in what used to be a very time consuming process that only rich elite people could do. Was there 1:1 artists for every one who wanted a painting of themselves? Nope. And that's why a lot of arguments about "AI will take our jobs" is nonsense. They're taking "jobs" that people were not going to pay for in the first place, the same argument big corporations use for piracy when they put a dollar value on losses.  People were not going to buy it anyway, so it's not a "loss", it's a failure to set a price the market will bear. There will be people who object to paying for anything, and you'll never get money out of them, but calling their presence a loss is the same as people using AI to generate low quality materials instead of paying for "an artist" to do the same.

 

I'm sure AI will replace some low-effort jobs, but people will not watch a TV show, Cartoon, or Video game (all routinely "crunched" while the IP holder is unwilling to pay for enough staff) that is too low of quality, and an AI will only be able to produce "low quality" things without human guidance.

 

Like I'm absolutely sure AI could create the same visuals of South Park or Rick&Morty, as the character designs are not complicated like Japanese cartoons or even Disney cartoons. However, AI would be far more likely to be able to create an entire episode of south park without human effort than it would Rick&Morty. The AI could just read the news headlines and produce something affirming or denouncing the subject without actually understanding it, because that's all South Park really is at it's core is a Satire, it doesn't need to be correct or honest. Rick and Morty however has dialogue that contains a lot of invented words, and the "aliens" and stuff are "original" though obviously inspired by things. So perhaps an AI given the opportunity could create something visually similar, but because AI's are not creative, at best it just regurgitates something already seen in these shows, like a fanfic.

 

That is pretty much what people had been using GPT for. Creating cursed fanfic-level content. 

https://www.gwern.net/GPT-3#repetitiondivergence-sampling

Quote

Repetition/gibberish (mystery). Autoregressive language models trained by likelihood (prediction) loss all share an extremely annoying problem: when you generate free-form completions, they have a tendency to eventually fall into repetitive loops of gibberish.

I can confirm this happens on at least one TTS model I use, but didn't train myself. It will sound natural, but occasionally will emit something like a howl or stutter because the word is too short, and typical training corpus prefer long samples, not short ones.

 

https://www.gwern.net/GPT-3#copyright

Quote

Creativity: 
GPT-3 has, like any well-educated human, memorized vast reams of material and is happy to emit them when that seems like an appropriate continuation & how the ‘real’ online text might continue; GPT-3 is capable of being highly original, it just doesn’t care about being original and the onus is on the user to craft a prompt which elicits new text, if that is what is desired, and to spot-check novelty.

And that's kind of the problem with what an AI may do. It might be able to do things, it just doesn't know it should, and requires more specific prompts to elicit something original out of it. An AI will absolutely take any shortcut it can learn to use. Humans avoid those shortcuts because they know it ultimately creates low quality output, but an AI? When an AI is given low quality inputs, it produces low quality output. It doesn't know to discard low quality samples, it doesn't know that it's low quality. In TTS training corpora, which tends to be shared with ASR's, they're often given 16khz audio for ASR and 24khz for TTS, yet if a human listens to these recordings, it's immediately obvious that the recording is low quality. Sure a TTS can work with it, but it will never produce anything better than the 24khz audio, because that's what it learned, to sound like a 24khz recording.

 

[Neroon]

11 hours ago, Kisai said:

Wedding photographers are still a thing. School and Graduation photos are still a thing. "Old Timey" Photos are still a thing. Photos taken on rides at amusement parks are still a thing.

 

What has happened with photography is that "the camera" replaced "the painter" in what used to be a very time consuming process that only rich elite people could do. Was there 1:1 artists for every one who wanted a painting of themselves? Nope. And that's why a lot of arguments about "AI will take our jobs" is nonsense. They're taking "jobs" that people were not going to pay for in the first place, the same argument big corporations use for piracy when they put a dollar value on losses.  People were not going to buy it anyway, so it's not a "loss", it's a failure to set a price the market will bear. There will be people who object to paying for anything, and you'll never get money out of them, but calling their presence a loss is the same as people using AI to generate low quality materials instead of paying for "an artist" to do the same.

 

I'm sure AI will replace some low-effort jobs, but people will not watch a TV show, Cartoon, or Video game (all routinely "crunched" while the IP holder is unwilling to pay for enough staff) that is too low of quality, and an AI will only be able to produce "low quality" things without human guidance.

 

Like I'm absolutely sure AI could create the same visuals of South Park or Rick&Morty, as the character designs are not complicated like Japanese cartoons or even Disney cartoons. However, AI would be far more likely to be able to create an entire episode of south park without human effort than it would Rick&Morty. The AI could just read the news headlines and produce something affirming or denouncing the subject without actually understanding it, because that's all South Park really is at it's core is a Satire, it doesn't need to be correct or honest. Rick and Morty however has dialogue that contains a lot of invented words, and the "aliens" and stuff are "original" though obviously inspired by things. So perhaps an AI given the opportunity could create something visually similar, but because AI's are not creative, at best it just regurgitates something already seen in these shows, like a fanfic.

 

That is pretty much what people had been using GPT for. Creating cursed fanfic-level content. 

https://www.gwern.net/GPT-3#repetitiondivergence-sampling

I can confirm this happens on at least one TTS model I use, but didn't train myself. It will sound natural, but occasionally will emit something like a howl or stutter because the word is too short, and typical training corpus prefer long samples, not short ones.

 

https://www.gwern.net/GPT-3#copyright

And that's kind of the problem with what an AI may do. It might be able to do things, it just doesn't know it should, and requires more specific prompts to elicit something original out of it. An AI will absolutely take any shortcut it can learn to use. Humans avoid those shortcuts because they know it ultimately creates low quality output, but an AI? When an AI is given low quality inputs, it produces low quality output. It doesn't know to discard low quality samples, it doesn't know that it's low quality. In TTS training corpora, which tends to be shared with ASR's, they're often given 16khz audio for ASR and 24khz for TTS, yet if a human listens to these recordings, it's immediately obvious that the recording is low quality. Sure a TTS can work with it, but it will never produce anything better than the 24khz audio, because that's what it learned, to sound like a 24khz recording.

 

Have you ever looked in that field? Aside from that type of photography being mindnumblingy dull, the problem is that every idiot with a DSLR thinks they can make pictures. So if you are even remotely serious about this stuff, it's hard to find jobs that are somewhat interesting and/or pay decent.

 

Sure there are exceptions, like there are always, but the point is that the field itself is extremely tough.

 

Low effort jobs? Have you seen the AI stuff? Midjourney Showcase

That's not low effort, if you can do that as a person, you are absolutely talented.

 

As for animation, yeah I don't know where that is at now, but that would create consistency issues and so many other things, so that's obviously not what we should be comparing it to.

 

It feels like you are either comparing it with a different field (3D and animation), or you are lagging behind with what technology can do these days.