AI generated art

[Mon0Zer0]

 

[Mon0Zer0]

 

[Laura R Hepworth]

Just a word of caution about AI art generators, there is a ton of debate right now over the unauthorized use of images to train the AI models and whether or not it is a violation of copyright. I've played around with several AI art generators (particularly Artbreeder), but have backed off of it for now until I see how this controversy with it turns out. Personally, I really hope it works out that they aren't considered copyright violations as AI art generators have so much potential as inspiration/reference material. The issue hasn't come up in court yet (that I know of), but that's probably only a matter of time.

 

[Mon0Zer0]

Just a word of caution about AI art generators, there is a ton of debate right now over the unauthorized use of images to train the AI models and whether or not it is a violation of copyright. I've played around with several AI art generators (particularly Artbreeder), but have backed off of it for now until I see how this controversy with it turns out. Personally, I really hope it works out that they aren't considered copyright violations as AI art generators have so much potential as inspiration/reference material. The issue hasn't come up in court yet (that I know of), but that's probably only a matter of time.

Definitely. This is just a bit of fun for the time being. I suspect they'll be too useful to concept artists to hold back, though. It's awesome to photobash / 3d model a scene and then run it through a diffusion network to add all the textures. It's not a million miles away from what concept artists already do!

 

[Laura R Hepworth]

Definitely. This is just a bit of fun for the time being. I suspect they'll be too useful to concept artists to hold back, though. It's awesome to photobash / 3d model a scene and then run it through a diffusion network to add all the textures. It's not a million miles away from what concept artists already do!

Oh, it definitely is fun! It could be super useful too so I hope this can all get sorted out. The difference though is that all the images concept artists use for their photobashing or matte paintings have been properly licensed and the AI generators haven't purchased the licenses to the images they trained with. I also think they'll be able to figure it all out, but, as a freelance artist, I'm personally a little uneasy with it right now as I need to know for sure it won't come back to bite me later since what I make would be used in commercial projects.

 

[Robert Zwilling]

I don't think you can use other people's work without paying for it. But the digital generation's motto is that if it ain't nailed down, it's free to take, until you get to court anyway. It works great using your own images.

 

[Mon0Zer0]

I don't think you can use other people's work without paying for it.

You're not using someone else's art.

The algorithm is a set of data that has analysed an artist's body of art and created a statistical model that answers the question: "What defines an artist's style?"

Starting from a seed image based on perlin noise, the programme creates a series of shapes defined by the user's prompt and then uses the statistical model to build on those shapes in combination with what it understands about objects and perspective, and the way the artist uses brush strokes to render an object. The programme is not combining elements of an artist's work, it's performing a routine that 'says', "based on the content and the shapes how would [artist] tackle this object".

It is trained on a body of work, but the models it uses to render an object in no way contain that work, only a mathematical representation of the artist's style as a series of weights and biases.

The grey area revolves around whether an artist owns a style or not.

I suspect that the answer is likely to be "no" on the grounds that technique and use of materials can't be copyrighted and pretty much every artist, knowingly or not, owes a huge debt to some other artist ("swiping" is the term professional comic artists use). Until someone is taken to court it's impossible to say for sure. I'm not sure the kind of artists that are used in the training data are well off enough to bring such a case, and unless it affects Banksy or Damien Hurst it's hard to see it ever getting there (in which case Blek Le Rat may want to have a word with Banksy!).

 

[Daysman]

Saw this recently...

 

[Harpo]

I just heard about DALL•E mini thanks to a couple of friends

DALL·E Mini – Generate Images From Any Text Prompt

In this article, we explore Craiyon (also known as DALL·E mini), which is an artificial intelligence model that generates images from any prompt you give.

wandb.ai

Here’s an example, based on a well-known Gary Larson cartoon

 

[Harpo]

Of course, as with any new art form or technology, the “early days” are over once people start using it to be silly.

 

[Swank]

I just heard about DALL•E mini thanks to a couple of friends

DALL·E Mini – Generate Images From Any Text Prompt

In this article, we explore Craiyon (also known as DALL·E mini), which is an artificial intelligence model that generates images from any prompt you give.

wandb.ai

Here’s an example, based on a well-known Gary Larson cartoon

View attachment 90278

You can't un-see that!

 

[Laura R Hepworth]

You're not using someone else's art.

The algorithm is a set of data that has analysed an artist's body of art and created a statistical model that answers the question: "What defines an artist's style?"

Starting from a seed image based on perlin noise, the programme creates a series of shapes defined by the user's prompt and then uses the statistical model to build on those shapes in combination with what it understands about objects and perspective, and the way the artist uses brush strokes to render an object. The programme is not combining elements of an artist's work, it's performing a routine that 'says', "based on the content and the shapes how would [artist] tackle this object".

It is trained on a body of work, but the models it uses to render an object in no way contain that work, only a mathematical representation of the artist's style as a series of weights and biases.

The grey area revolves around whether an artist owns a style or not.

I suspect that the answer is likely to be "no" on the grounds that technique and use of materials can't be copyrighted and pretty much every artist, knowingly or not, owes a huge debt to some other artist ("swiping" is the term professional comic artists use). Until someone is taken to court it's impossible to say for sure. I'm not sure the kind of artists that are used in the training data are well off enough to bring such a case, and unless it affects Banksy or Damien Hurst it's hard to see it ever getting there (in which case Blek Le Rat may want to have a word with Banksy!).

What annoys me is that this whole situation could easily have been avoided if the companies working on the AI art generators had simply posted a public call explaining what they were doing, what sort of images they needed, how the image would be used, etc. and asked for anyone alright with their photographs/artwork being used in such a way to submit them to them at a dedicated submissions email address. Sure, it might have taken longer, but it would have prevented the whole issue of using unauthorized images and had no need to be sorted out in court.

 

[Mon0Zer0]

These are from Midjourney. Rick & Morty by Ralph Steadman:

Dog Emperor:

Pig in a wig:

Cyberpunk:

 

[Mon0Zer0]

What annoys me is that this whole situation could easily have been avoided if the companies working on the AI art generators had simply posted a public call explaining what they were doing, what sort of images they needed, how the image would be used, etc. and asked for anyone alright with their photographs/artwork being used in such a way to submit them to them at a dedicated submissions email address. Sure, it might have taken longer, but it would have prevented the whole issue of using unauthorized images and had no need to be sorted out in court.

I think these images are a by-product of image recognition systems, rather than something they intended. Dall-e2 has been trained on 635 million images from the web, I don't think it would be even remotely possible to do that, at that scale.

 

[Laura R Hepworth]

I think these images are a by-product of image recognition systems, rather than something they intended. Dall-e2 has been trained on 635 million images from the web, I don't think it would be even remotely possible to do that, at that scale.

The problem is that the AI modeling still starts with images that they don't have the rights to use. I've been reading a lot bout the issue lately and both sides of the argument make very good points. I genuinely hope it works out in favour of the AI programs though as I would love to be able to use them in my commercial work and have full confidence that it won't somehow land me (or my client) in legal trouble. It just bugs me that there was a way around it and nearly all of them opted for the easy and legally questionable webscraping method instead. I have seen an AI portrait generator that sourced all their training images in-house and with full permissions from all involved. So, it is possible, just much harder. On one hand, I don't see much difference between the more advanced AI art generators and an artist using indirect reference, but, on the other hand though, I also recognize the illegality of taking someone's photos/artwork and using them without permission and the concern of whether or not the AI generator will actually create fully new image or simply a reiteration of the original. I find the tech rather fascinating so I really hope they get things sorted out quickly so I can feel at ease to go back to playing with it to my heart's content.

 

[Swank]

The problem is that the AI modeling still starts with images that they don't have the rights to use. I've been reading a lot bout the issue lately and both sides of the argument make very good points. I genuinely hope it works out in favour of the AI programs though as I would love to be able to use them in my commercial work and have full confidence that it won't somehow land me (or my client) in legal trouble. It just bugs me that there was a way around it and nearly all of them opted for the easy and legally questionable webscraping method instead. I have seen an AI portrait generator that sourced all their training images in-house and with full permissions from all involved. So, it is possible, just much harder. On one hand, I don't see much difference between the more advanced AI art generators and an artist using indirect reference, but, on the other hand though, I also recognize the illegality of taking someone's photos/artwork and using them without permission and the concern of whether or not the AI generator will actually create fully new image or simply a reiteration of the original. I find the tech rather fascinating so I really hope they get things sorted out quickly so I can feel at ease to go back to playing with it to my heart's content.

Couldn't one just use a random photograph paired with the artistic style from somewhere else?

 

[Laura R Hepworth]

Couldn't one just use a random photograph paired with the artistic style from somewhere else?

No, because there is no such thing as a 'random photograph.' All photographs are copyrighted to the photographer automatically. Using any of them without permission is copyright violation. While there are public domain image sites, unfortunately, a number of these sites are having problems with image piracy where people are uploading images that they don't have the rights to and so don't have any legal standing to make them available for public use.

 

[Swank]

No, because there is no such thing as a 'random photograph.' All photographs are copyrighted to the photographer automatically. Using any of them without permission is copyright violation. While there are public domain image sites, unfortunately, a number of these sites are having problems with image piracy where people are uploading images that they don't have the rights to and so don't have any legal standing to make them available for public use.

You are capable of taking a photo and granting yourself the right to use it.

 

[Laura R Hepworth]

You are capable of taking a photo and granting yourself the right to use it.

Unless you are the photographer of said photo, no, you can't. It's still stealing and people have been sued for it. Even bloggers have been sued or issued DMCA takedowns for the use of photos they didn't take themselves (and even when they give proper image credit and linkbacks).

 

[Mon0Zer0]

AI generator will actually create fully new image or simply a reiteration of the original.

On that point the answer is yeah it is a brand new image.

I've been testing it and running my own 2d and 3d work through and bypassing the perlin noise seed stage and it produces new works around those images. If you arrange some 3d primitives in blender into a shape (a castle / a hill / stream etc) and then render out a rough scene, the programme will add lighting / textures as if it were a painting all along. It doesn't always get it right, so I've been overpainting / comping different passes together.

Those are brush strokes in the style of someone else around shapes / colour / depth that I define up front, and definitely not a patchwork quilt of other people's images as if it were photobashed. It's pretty remarkable.