4 ms·

All of the AI art I see has a sort of melted look.

iakov 5y ago
I agree. It's a mish-mash of melting shapes, nothing resembles anything. I'm really confused why would someone call this "perfectly fake". Have those people seen the real covers? The covers were meant to catch your eye, to evoke some kind of emotion. Compared to them, most of the examples evokes only repulsion in me.
- JoeAltmaier 5y ago
  I have the same reaction. They're like an alien with only a description of human activities but no real knowledge of behavior or society, tried to create a painting. Everything is strangely squished, there's no clear dividing line between one thing and another, nothing is a functional symbol but just an impression of a symbol. "Humans like yellow; lets put a yellow shape in there!"
- amelius 5y ago
  They should put a human in a fMRI scanner, and place this in the feedback loop of the GAN. When the human feels repulsion, the fMRI scanner picks this up and sends negative feedback to the network.
  - octopoc 5y ago
    The human would most likely quickly be desensitized and only output feelings of self hatred and suicide. But to be sure we should start a multibillion dollar startup that generates jobs for millions of starving third worlders.
- astrange 5y ago
  This technology has been developing very fast; pictures look a lot more realistic than they used to, and you can guide the process with real images along the way. (Zoetrope is a good notebook for this.) The main issue is that it's hard to generate good looking faces without a specialized model (because we're so picky about humans), and the companies generating models like DALL-E don't want to teach them how to generate faces in case of bias.
- lariati 5y ago
  GAN art is really a one trick pony. I loved it at first and got quite into it but I have seen enough of these images now that the novelty has worn off and they are just kind of shit. Physical painting inspired by GAN art might be cool but I don't have the talent for that. There is also something lost for sure when I know I could make a 100 of these album covers in maybe 20 minutes of tinkering around?
hypertele-Xii 5y ago
With some craft it's possible to produce better work. Here's two of mine for example: https://hypertele.fi/0416e43936606257 https://hypertele.fi/0416e43936606257 https://hypertele.fi/a78fd2489d1b8df8 https://hypertele.fi/a78fd2489d1b8df8
severak_cz 5y ago
Because this is actually what is it after all - you take some pictures, melt them and cast new ones.
rectang 5y ago
There are many ways to avoid that. One way is to use an AI to perform "style extraction", and then apply it to existing images. Here's a presentation for Papers We Love San Diego on a paper describing such a technique: https://www.youtube.com/watch?v=KFORVMT8AWs https://www.youtube.com/watch?v=KFORVMT8AWs I would argue that the proper way to do this while respecting other creators is to create your own original source images and then train the model based on those, before then applying the style to some other original image of your own.

I agree. It's a mish-mash of melting shapes, nothing resembles anything. I'm really confused why would someone call this "perfectly fake". Have those people seen the real covers? The covers were meant to catch your eye, to evoke some kind of emotion. Compared to them, most of the examples evokes only repulsion in me.

JoeAltmaier 5y ago
I have the same reaction. They're like an alien with only a description of human activities but no real knowledge of behavior or society, tried to create a painting. Everything is strangely squished, there's no clear dividing line between one thing and another, nothing is a functional symbol but just an impression of a symbol. "Humans like yellow; lets put a yellow shape in there!"
amelius 5y ago
They should put a human in a fMRI scanner, and place this in the feedback loop of the GAN. When the human feels repulsion, the fMRI scanner picks this up and sends negative feedback to the network.
- octopoc 5y ago
  The human would most likely quickly be desensitized and only output feelings of self hatred and suicide. But to be sure we should start a multibillion dollar startup that generates jobs for millions of starving third worlders.
astrange 5y ago
This technology has been developing very fast; pictures look a lot more realistic than they used to, and you can guide the process with real images along the way. (Zoetrope is a good notebook for this.) The main issue is that it's hard to generate good looking faces without a specialized model (because we're so picky about humans), and the companies generating models like DALL-E don't want to teach them how to generate faces in case of bias.
lariati 5y ago
GAN art is really a one trick pony. I loved it at first and got quite into it but I have seen enough of these images now that the novelty has worn off and they are just kind of shit. Physical painting inspired by GAN art might be cool but I don't have the talent for that. There is also something lost for sure when I know I could make a 100 of these album covers in maybe 20 minutes of tinkering around?

hypertele-Xii 5y ago

With some craft it's possible to produce better work. Here's two of mine for example: https://hypertele.fi/0416e43936606257 https://hypertele.fi/0416e43936606257 https://hypertele.fi/a78fd2489d1b8df8 https://hypertele.fi/a78fd2489d1b8df8

severak_cz 5y ago

Because this is actually what is it after all - you take some pictures, melt them and cast new ones.

rectang 5y ago

There are many ways to avoid that. One way is to use an AI to perform "style extraction", and then apply it to existing images. Here's a presentation for Papers We Love San Diego on a paper describing such a technique: https://www.youtube.com/watch?v=KFORVMT8AWs https://www.youtube.com/watch?v=KFORVMT8AWs I would argue that the proper way to do this while respecting other creators is to create your own original source images and then train the model based on those, before then applying the style to some other original image of your own.