5 ms·

>"The output represents the model’s understanding of what is useful, aesthetic, pleasing, etc. and that, together with data filtering and cleaning that general

>"The output represents the model’s understanding of what is useful, aesthetic, pleasing, etc. and that, together with data filtering and cleaning that general image generating AI companies do,2 is what the companies consider most valuable, not the training data.3" This didn't make any sense to me. Without the curated training data (images) how are they making the models? No matter what, putting images into your machine then selling the output generated with them and not compensating the original creators is going to be seen as problematic. Machines aren't people.

lodovic 4y ago
I compare it to reading a big stack of comic books and then trying to create your own, painstakingly reproducing the layout, color palettes and character styles. Or listening to (eg) Pink Floyd and then making a song that really matches their music. YouTube is full of those, it's really similar to SD images. But still not the same as sampling as that requires actually duplicating parts of the original. The reason we have copyright for a limited time is to promote the arts, not to give people a monopoly on a specific style so they can milk it. (although that is what seems to be happening)
seydor 4y ago
They are never going to compensate the artists. It's cheaper to hire 1000 designers to make 100000 images of artistic styles they are going for
- krisoft 4y ago
  uhm. What you are proposing is compensating the artists. Those 1000 designers are the artist in question then. > It's cheaper to hire 1000 designers to make 100000 images of artistic styles they are going for I bet that you are massively underestimating the cost of that.
- gugagore 4y ago
  I think you underestimate the scale of data these models are trained on by many orders of magnitude.
  - Ajedi32 4y ago
    I suspect you might be able to get pretty good results by training the system on video/CGI/other images that can be easily mass produced, then fine-tuning on a much smaller number of drawings and other stylized images.
wvenable 4y ago
> Machines aren't people. There's no reason why that is the significant detail. Why does it matter? If you can look at millions of images over your lifetime and faithfully reproduce famous works of art by hand, aren't you just as wrong?
- EamonnMR 4y ago
  Humans can be trusted not to do that thing, and get in trouble if they get caught.
  - wvenable 4y ago
    Is it not humans giving the AI prompts to do these things? Why does the machine in the middle matter?
- shagie 4y ago
  Machines can't create copyrighted works. Setting aside the question of "is the model a derivative work", running the program cannot create a work that is copyrighted. Only humans (and not monkeys) can hold a copyright. And thus, the questions are: "is generating a model based on the data set a derivate work" and the unasked question "is asking the model to generate a work in the style of {artist} a derivative work by the person asking the model?"
  - wvenable 4y ago
    > Running the program cannot create a work that is copyrighted If think you're going to need more clarity on what you mean by that. Programs are used to create copyrighted works all the time. And machines can and do create copies of other people's copyrighted works. > And thus, the questions are: "is generating a model based on the data set a derivate work" and the unasked question "is asking the model to generate a work in the style of {artist} a derivative work by the person asking the model?" My point is you can take the machine or model out of the question entirely. If you learn stuff and then produce something new with what you learned, is that a derivative work? That's already a complex question but it has nothing to do with how you learned it. It depends entirely on the output and has little to do with the input.
cma 4y ago
> No matter what, putting images into your machine then selling the output generated with them and not compensating the original creators is going to be seen as problematic. Machines aren't people. What about a company where you submit images and it tells you which faces are in them?

lodovic 4y ago

I compare it to reading a big stack of comic books and then trying to create your own, painstakingly reproducing the layout, color palettes and character styles. Or listening to (eg) Pink Floyd and then making a song that really matches their music. YouTube is full of those, it's really similar to SD images. But still not the same as sampling as that requires actually duplicating parts of the original. The reason we have copyright for a limited time is to promote the arts, not to give people a monopoly on a specific style so they can milk it. (although that is what seems to be happening)

seydor 4y ago

They are never going to compensate the artists. It's cheaper to hire 1000 designers to make 100000 images of artistic styles they are going for

krisoft 4y ago
uhm. What you are proposing is compensating the artists. Those 1000 designers are the artist in question then. > It's cheaper to hire 1000 designers to make 100000 images of artistic styles they are going for I bet that you are massively underestimating the cost of that.
gugagore 4y ago
I think you underestimate the scale of data these models are trained on by many orders of magnitude.
- Ajedi32 4y ago
  I suspect you might be able to get pretty good results by training the system on video/CGI/other images that can be easily mass produced, then fine-tuning on a much smaller number of drawings and other stylized images.

wvenable 4y ago

> Machines aren't people. There's no reason why that is the significant detail. Why does it matter? If you can look at millions of images over your lifetime and faithfully reproduce famous works of art by hand, aren't you just as wrong?

EamonnMR 4y ago
Humans can be trusted not to do that thing, and get in trouble if they get caught.
- wvenable 4y ago
  Is it not humans giving the AI prompts to do these things? Why does the machine in the middle matter?
shagie 4y ago
Machines can't create copyrighted works. Setting aside the question of "is the model a derivative work", running the program cannot create a work that is copyrighted. Only humans (and not monkeys) can hold a copyright. And thus, the questions are: "is generating a model based on the data set a derivate work" and the unasked question "is asking the model to generate a work in the style of {artist} a derivative work by the person asking the model?"
- wvenable 4y ago
  > Running the program cannot create a work that is copyrighted If think you're going to need more clarity on what you mean by that. Programs are used to create copyrighted works all the time. And machines can and do create copies of other people's copyrighted works. > And thus, the questions are: "is generating a model based on the data set a derivate work" and the unasked question "is asking the model to generate a work in the style of {artist} a derivative work by the person asking the model?" My point is you can take the machine or model out of the question entirely. If you learn stuff and then produce something new with what you learned, is that a derivative work? That's already a complex question but it has nothing to do with how you learned it. It depends entirely on the output and has little to do with the input.

cma 4y ago

> No matter what, putting images into your machine then selling the output generated with them and not compensating the original creators is going to be seen as problematic. Machines aren't people. What about a company where you submit images and it tells you which faces are in them?