Create a 7-string Guitar with Midjourney AI

[JimF]

I liked the AI art when it was the odd Instagram account making weird looking, uncanny-valley-esque images that seemed a bit off. But now I'm seeing more and more of it, it just reminds me of those people who spend all that time and effort trying to find a convincing distorted midi guitar sound because they can't play guitar.

 

[Rob Joyner]

"Stole" in the sense that it did the ML analog of looking at them. Which is what humans do. If you want to attempt to actually make an argument, you need to get clear on how the processing the model does is any different to what the human artist does to the extent that you want to call one theft and the other inspiration.

writing a prompt that combines billions of copyrighted images without permission to generate generic outputs for easy profit IS NOT studying.
It's theft.
It's the same as creating a data set with all of tony Iommi's riffs without his permission to generate hundreds of riffs writing some poorly written prompt and sell them as original music and make a profit.

And honestly I'm out of this discussion. You want to make moral gymnastics to defend something that's absolutely immoral. It affected me and thousands of designers in the game industry already and hinders our ability to pay bills and ultimately buy our beloved guitars.

The world is a sad place right now with this culture of "it is what it is, toughen up".

 

[Albake21]

"Stole" in the sense that it did the ML analog of looking at them. Which is what humans do. If you want to attempt to actually make an argument, you need to get clear on how the processing the model does is any different to what the human artist does to the extent that you want to call one theft and the other inspiration.

The human mind is just so much more powerful than an AI. An AI is pixel by pixel copying actual work... the human mind is only interpreting that work but being able to meld and create new life out of it. AI is literally just shitting out more of the same 1s and 0s when humans are creating their own numbers.

I fully understand the argument that an AI is just doing what humans do. And that the argument is humans can only create from what they know. But I refuse to believe these AI are able to actually create like how the human mind can. It's just not logically possible.

 

[narad]

writing a prompt that combines billions of copyrighted images without permission to generate generic outputs for easy profit IS NOT studying.
It's theft.
It's the same as creating a data set with all of tony Iommi's riffs without his permission to generate hundreds of riffs writing some poorly written prompt and sell them as original music and make a profit.

And honestly I'm out of this discussion. You want to make moral gymnastics to defend something that's absolutely immoral. It affected me and thousands of designers in the game industry already and hinders our ability to pay bills and ultimately buy our beloved guitars.

The world is a sad place right now with this culture of "it is what it is, toughen up".

If the Tony Iommi model then generates a ton of riffs that sound like they could Iommi but are completely new? That's not stealing, it's creating.

Yea, if you're run-of-the-mill artist you're fucked, but so are a ton of other people. Truckers are going to be fucked when fully autonomous driving is a thing. The amount of people about to get fucked by language AI is unbelievable. But trying to pretend that this is a moral issue is wrong. The AI doesn't need to train on anything new or copyrighted to do what it does. It doesn't need a bunch of random deviantart stuff to work. Artists right now are really over-valuing their importance in the model with these arguments. A ton of for-contract fiver type designer hires are probably done for right now, and that's with the models that exist now. The image generation models of 2 years ago were comparatively garbage. 2 years from now where's it going to be?

The human mind is just so much more powerful than an AI. An AI is pixel by pixel copying actual work... the human mind is only interpreting that work but being able to meld and create new life out of it. AI is literally just shitting out more of the same 1s and 0s when humans are creating their own numbers.

I fully understand the argument that an AI is just doing what humans do. And that the argument is humans can only create from what they know. But I refuse to believe these AI are able to actually create like how the human mind can. It's just not logically possible.

This is a nice thought but is shown time and time again to be untrue. The AI doesn't copy the actual work, btw. The AI doesn't have access to the work when it's generating.

 

[Rob Joyner]

If the Tony Iommi model then generates a ton of riffs that sound like they could Iommi but are completely new? That's not stealing, it's creating.

Yea, if you're run-of-the-mill artist you're fucked, but so are a ton of other people. Truckers are going to be fucked when fully autonomous driving is a thing. The amount of people about to get fucked by language AI is unbelievable. But trying to pretend that this is a moral issue is wrong. The AI doesn't need to train on anything new or copyrighted to do what it does. It doesn't need a bunch of random deviantart stuff to work. Artists right now are really over-valuing their importance in the model with these arguments. A ton of for-contract fiver type designer hires are probably done for right now, and that's with the models that exist now. The image generation models of 2 years ago were comparatively garbage. 2 years from now where's it going to be?



This is a nice thought but is shown time and time again to be untrue. The AI doesn't copy the actual work, btw. The AI doesn't have access to the work when it's generating.

Please explain how DeviantArt and Artstation have opt in and out buttons to allow your data to be scraped by these data sets. You have no idea what you're talking about.
It's not INTELLIGENCE it's just a data set that uses images from the internet.

MJ has a database of 3.2 billions images by artists to generate images. There's nothing intelligent about it. It's a mish mash of data.

 

[narad]

Please explain how DeviantArt and Artstation have opt in and out buttons to allow your data to be scraped by these data sets. You have no idea what you're talking about.
It's not INTELLIGENCE it's just a data set that uses images from the internet.

MJ has a database of 3.2 billions images by artists to generate images. There's nothing intelligent about it. It's a mish mash of data.

I don't need to explain that -- you can get rid of that data altogether. It's not magically going to make the technology stop working or go away. There is plenty of images and artwork that exists in the world that is not copyrighted. There's really no beating it unless you can consistently offer something that a client likes better than the automatically generated images. It sucks, but you can't ban technology just because you don't like it.

Mish mash of data? I'm not sure you know how ML works...

 

[Rob Joyner]

I don't need to explain that -- you can get rid of that data altogether. It's not magically going to make the technology stop working or go away. There is plenty of images and artwork that exists in the world that is not copyrighted. There's really no beating it unless you can consistently offer something that a client likes better than the automatically generated images. It sucks, but you can't ban technology just because you don't like it.

Mish mash of data? I'm not sure you know how ML works...

Maybe this will help. Please educate yourself.

What are text-to-image AI/ML models?
A text-to-image model takes input from a user in the form of a natural language prompt and produces an image matching that prompt. To condition that capability the model needs to be trained on a huge collection of images, media, and text descriptions scraped from the web and collected in the form of a “dataset ” in order to extract and encode an intricate statistical survey of the dataset's items. Images are generated from an input prompt by assembling visual data that attempts to best simulate the statistical correlations between text in the dataset and images in the dataset in order to produce "acceptable" results.

Some of this data is the copyrighted work of artists and the private data of the public. As these models produce derivative works based on probability and statistics, they are prone to reproducing biases, stereotypes, and copyrighted works present within the datasets. Essentially, it could be described as “an advanced photo mixer” generating potential derivations based on statistical probability. **we are committed to an accurate description of the technology and the issues facing them, so we decided to update our language for a more detailed look at the issues. Read the original verbiage below**


What are these unethical practices?
Images and text descriptions across the internet are gathered and taken by a practice called data mining and/or data scraping. This technique allows AI/ML companies to build the massive datasets necessary to train these AI/ML models.

Stability AI funded the creation of the biggest and most utilized database called LAION 5B. The LAION 5B database, originally created on the pretext of “research” contains 5.8 billion text and image data, including copyrighted data and private data, gathered without artists, individuals, and businesses' knowledge or permission.

MidJourney, Stability AI, Prisma AI (Lensa AI) & other AI/ML companies are utilizing these research datasets that contain private and copyrighted data, for profit. They did so without any individual’s knowledge or consent, and certainly without compensation.

 

[Albake21]

This is a nice thought but is shown time and time again to be untrue. The AI doesn't copy the actual work, btw. The AI doesn't have access to the work when it's generating.

You're not wrong. I've seen this conversation hit this exact wall numerous times. I still think it's very debatable, but only time will tell.

 

[narad]

Maybe this will help. Please educate yourself.

What are text-to-image AI/ML models?
A text-to-image model takes input from a user in the form of a natural language prompt and produces an image matching that prompt. To condition that capability the model needs to be trained on a huge collection of images, media, and text descriptions scraped from the web and collected in the form of a “dataset ” in order to extract and encode an intricate statistical survey of the dataset's items. Images are generated from an input prompt by assembling visual data that attempts to best simulate the statistical correlations between text in the dataset and images in the dataset in order to produce "acceptable" results.

Some of this data is the copyrighted work of artists and the private data of the public. As these models produce derivative works based on probability and statistics, they are prone to reproducing biases, stereotypes, and copyrighted works present within the datasets. Essentially, it could be described as “an advanced photo mixer” generating potential derivations based on statistical probability. **we are committed to an accurate description of the technology and the issues facing them, so we decided to update our language for a more detailed look at the issues. Read the original verbiage below**


What are these unethical practices?
Images and text descriptions across the internet are gathered and taken by a practice called data mining and/or data scraping. This technique allows AI/ML companies to build the massive datasets necessary to train these AI/ML models.

Stability AI funded the creation of the biggest and most utilized database called LAION 5B. The LAION 5B database, originally created on the pretext of “research” contains 5.8 billion text and image data, including copyrighted data and private data, gathered without artists, individuals, and businesses' knowledge or permission.

MidJourney, Stability AI, Prisma AI (Lensa AI) & other AI/ML companies are utilizing these research datasets that contain private and copyrighted data, for profit. They did so without any individual’s knowledge or consent, and certainly without compensation.

Yeaaaa....I'm pretty educated on the topic. It is actually already clear that the author of this text is not well versed in ML or intentionally disingenuous with their choice of words here -- they choose to describe the model as "an advanced photo mixer", which erroneously makes it sound like snippets of individual artworks are chosen to produce the final image. That's not how any of this works.

Yes, currently these models are built on big web scrapes of images without contributing anything to the authors. However, that doesn't mean this particular dataset is necessary for producing text-to-image models that would still trounce a huge amount of design jobs over the next 5 years. Whether authors need to be compensated for data usage in this way is also a completely grey area. But even if the usage policies that inevitably crop up here are as pro-artist as you could imagine, companies will just buy the data. And then it's done. You don't need to do it ever again. There's no need to continually reimburse artists for these models, and as these models are not, in fact, data mixers, there's no way to determine the extent to which particular works contributed to any specific generated image. New models will be more data efficient and require fewer samples. The types of samples that are desirable won't be those produced by the general art community. If I wanted an ML music model, I'm not going to scrape everyone's shitty soundcloud. I'm going to target specific vetted music that is more universally praised, and just pay to use it. That doesn't help the other 99%. So your every day artists are still basically shit out of luck, because they weren't actually important to the ML models.

You're not wrong. I've seen this conversation hit this exact wall numerous times. I still think it's very debatable, but only time will tell.

Yea, there's no great counter-example to say it's impossible. Just that many before have set some arbitrary goal as being the definitive ceiling that automation will never reach, and again and again, a machine manages to do it. It also seems weird to me now though to come down hard on the "they can't do it" side, right when two of the most significant advances in decades have just dropped, and with no real sign of letting up. If anything, all the publicity is only drumming up more of the necessary funding to procure compute and data for the next iteration of models.

 

[bostjan]

Please educate yourself.

Oops, might not want to explain entry-level machine learning concepts to @narad, the guy who has been involved with this stuff from its novelty.

 

[Rob Joyner]

Yes, currently these models are built on big web scrapes of images without contributing anything to the authors.

you just ended your own argument altogether here.




Just because you wrote a couple of If and Else's doesn't mean you know what you're talking about.

 

[BlackMastodon]

I hate AI generated images. They fucking weird me out. No thanks.

Fuck AI art. Unhuman.

My first thought at seeing this picture was "thanks, I hate it."

Agree with these guys, this blows and it's ugly af.

 

[StevenC]

Why are we arguing about this when AI can't produce art in any sense?

And secondly, why are we arguing about this when all AI "art" looks like shit?

 

[bostjan]

Welcome to the internet.

Everything looks/sounds/tastes like shit to everyone and we argue about everything. Also, everyone is smarter than the experts and no one really knows what point they are trying to make anyway.

 

[LordCashew]

It looked cool until I saw the bridge and that gives me the creeps.... I don't know how to explain it, it's nightmarish to me lol

It is ineffably evil. Perhaps it triggers whatever instinctual part of our brain helps us avoid infections, because I somehow feel like it's infested with parasites.

 

[wheresthefbomb]

It is ineffably evil. Perhaps it triggers whatever instinctual part of our brain helps us avoid infections, because I somehow feel like it's infested with parasites.

Yeah it tickles the "unspeakable horror" part of my brain. I don't need to know why, I just know that I instinctively, viscerally, do not fuck with it.

 

[BlackMastodon]

Also, everyone is smarter than the experts and no one really knows what point they are trying to make anyway.

Have you been reading my live journal again??

 

[narad]

you just ended your own argument altogether here.


View attachment 119191

Just because you wrote a couple of If and Else's doesn't mean you know what you're talking about.

Your problem is that you're conflating midjourney with AI art. AI art is going to basically destroy most of the freelance art industry, probably the current model architectures are sufficiently good enough to do it already (even with their propensity for drawing extra digits), and it's not going to need to compensate artists in a significant way to do it. AI art != midjourney. Midjourney is just the first tiniest drop of this technology that the public is experiencing at a level where the implications for the industry are clear.

Now you can criticize the dataset that was used to train midjourney as not being constructed or used properly, but that's not a good basket to place all your eggs. (a) because you're basically trying to create a new precedent for "fair use" since many other services use publicly available images in nearly identical way), and (b) because future iterations of these models and services will just use datasets in a way that is free and clear, be it by using only free use images, or by compensating some artists, but probably not many and probably not much. So then your "this is evil" argument falls apart, because you tried to make it about copyrights for the images used in the dataset, instead of what I imagine is the true source for this outrage -- that it's clearly going to do a better job at generating images for most purposes than artists that would have otherwise been contracted to do so.

As to the medical images, there's a good analogy there wrt most of the art that was scraped. Sure, those images shouldn't be in the dataset (really, shouldn't have been uploaded to some publicly crawlable website in the first place) but do you think those images are important to the ability of the model to generate images? Someone points out that there are images that shouldn't be in the dataset. Images are filtered out of dataset. Model is retrained. Model still does what it does. What now?

 

[nickgray]

because I somehow feel like it's infested with parasites.

 

[LordCashew]

View attachment 119220

I feel that these palpably grotesque caricatures are in some way emblematic of advanced technology's potential effect on humanity.

Or maybe it simply reveals a truth we don't want to see...