3 ms·

AI training on AI-generated images works like a very complicated Reinforcement Learning from Human Feedback (RLHF), only the best images have a chance to be rel

by GaggiX 4y ago

AI training on AI-generated images works like a very complicated Reinforcement Learning from Human Feedback (RLHF), only the best images have a chance to be relevant enough to be noticed by a scraper, after being selected by the user who used the AI, and they have probably been shared several times by random people. Also, a lot of AI-generated content doesn't look like trash to me, I've seen a lot of good images, the fact that AIs can generate trash doesn't seem to be a problem to me, I can also generate junk and share it on the net if I want.

nottorp 4y ago
> Also, a lot of AI-generated content doesn't look like trash to me, I've seen a lot of good images Translation: There is a lot of human generated original art that you're not aware of, and when you see it in the "AI"'s output you think it's original.
- GaggiX 4y ago
  So you are implying that an image generated by a generative model must resemble closely to an image in the dataset, right? Because I don't think that's true, you can try to prove me wrong if you want, I'll give you a useful resource, the LAION dataset that Stable Diffusion was trained with is completely open, generate or find an image generated by Stable Diffusion and find the most semantically similar images in the dataset using: https://rom1504.github.io/clip-retrieval https://rom1504.github.io/clip-retrieval
- dwringer 4y ago
  People have the same argument all the time about musicians ripping off each others' compositions, no AI needed for that.
  - nottorp 4y ago
    Same for movie scripts, but sometimes they do fail and something original squeezes through. I'm saying this can't be the case with a pattern matching/combining algorithm.
    - dwringer 4y ago
      I've been playing around with stable diffusion a bit, and if you use an iterative process starting with either an extremely vivid and "code-golfed" prompt (like verses of poetry, concise prose, or elegant lisp), or your own illustration, and then feed the output back in again (with some fuzzing) along with the prompt, or a modified prompt, or a totally new prompt, after 2-3 generations the result becomes something that, although it undeniably integrates existing works in myriad ways, looks wholly unique and shaped by the user of the AI system more than by any of the individual contributing parts. I don't think this is terribly different from my creative process when I play music; I feel I have to work just as hard to avoid copping other peoples' stuff unconsciously.

by GaggiX 4y ago

nottorp 4y ago

> Also, a lot of AI-generated content doesn't look like trash to me, I've seen a lot of good images Translation: There is a lot of human generated original art that you're not aware of, and when you see it in the "AI"'s output you think it's original.

GaggiX 4y ago
So you are implying that an image generated by a generative model must resemble closely to an image in the dataset, right? Because I don't think that's true, you can try to prove me wrong if you want, I'll give you a useful resource, the LAION dataset that Stable Diffusion was trained with is completely open, generate or find an image generated by Stable Diffusion and find the most semantically similar images in the dataset using: https://rom1504.github.io/clip-retrieval https://rom1504.github.io/clip-retrieval
dwringer 4y ago
People have the same argument all the time about musicians ripping off each others' compositions, no AI needed for that.
- nottorp 4y ago
  Same for movie scripts, but sometimes they do fail and something original squeezes through. I'm saying this can't be the case with a pattern matching/combining algorithm.
  - dwringer 4y ago
    I've been playing around with stable diffusion a bit, and if you use an iterative process starting with either an extremely vivid and "code-golfed" prompt (like verses of poetry, concise prose, or elegant lisp), or your own illustration, and then feed the output back in again (with some fuzzing) along with the prompt, or a modified prompt, or a totally new prompt, after 2-3 generations the result becomes something that, although it undeniably integrates existing works in myriad ways, looks wholly unique and shaped by the user of the AI system more than by any of the individual contributing parts. I don't think this is terribly different from my creative process when I play music; I feel I have to work just as hard to avoid copping other peoples' stuff unconsciously.