metin,
@metin@graphics.social avatar

When generative AI is trained with AI-generated data, it becomes degenerat(iv)e AI.

tomcom,
@tomcom@mastodon.nl avatar

@metin I thought that would be something you'd be all for???

metin,
@metin@graphics.social avatar

@tomcom 🙂 I welcome AI as a tool for certain operations, like denoising an image, but I dislike the generative varieties, because they're trained with stolen work, and cause creatives to lose customers and motivation. Also, generative AI gives the same amount of non-satisfaction as a Google image search, compared to creating by yourself.

tomcom,
@tomcom@mastodon.nl avatar

@metin so if you're against them because they use other people's work, then why are you against them training from their own output?

Surely this is a type of training you'd support?

daniel,
@daniel@social.dhelonious.de avatar

@tomcom @metin @tomcom @metin This was a joke about the fact that generative AI actually degenerates when trained on AI-generated training images. This is because AI-generated images tend to have defects and other artifacts.

For example, AI sometimes creates hands with too many fingers. If you train an AI on such images, it learns that it is normal for hands to have any number of fingers and the problem increases.

tomcom,
@tomcom@mastodon.nl avatar

@daniel @metin so long as a human has gone through and curated images, then you could definitely use AI images for training data.

metin,
@metin@graphics.social avatar

@tomcom @daniel Human curation, consent of each author, and compensation for the use of the material. But I'm afraid that's almost impossible, because a thorough AI training needs many thousands of images. A solution could be a platform that pays creators each time their images are used for AI training.

Recently, Apple made a multi-million deal with Shutterstock for AI training, but I haven't read yet whether the actual image creators get anything from that.

tomcom,
@tomcom@mastodon.nl avatar

@metin @daniel this could actually be a big business in the future, because companies will pay money for good training data. Some of that money should go to the creators.

The only problem is that tracking down the original creators for every random image they've trawled from the internet might be a significantly bigger task than finding the images in the first place.

I'll add that training on AI generated images sidesteps these issues entirely.

metin,
@metin@graphics.social avatar

@tomcom @daniel Yes, but AI images have always been trained on real people's images to start with, no matter how much it becomes diluted by retraining with AI output. It remains a derivative of copyrighted works.

  • All
  • Subscribed
  • Moderated
  • Favorites
  • ai
  • Leos
  • DreamBathrooms
  • thenastyranch
  • magazineikmin
  • everett
  • InstantRegret
  • Youngstown
  • mdbf
  • slotface
  • rosin
  • modclub
  • kavyap
  • GTA5RPClips
  • cubers
  • JUstTest
  • osvaldo12
  • tacticalgear
  • Durango
  • khanakhh
  • ngwrru68w68
  • provamag3
  • anitta
  • tester
  • cisconetworking
  • ethstaker
  • megavids
  • normalnudes
  • lostlight
  • All magazines