Re:I Am Tired of AI
by grecy on Friday September 27, 2024 @05:41AM
(#41668161)
And it will get a lot better quickly. Ten years from now it will not be slop.
[ Parent
| on HN ]
Not sure about that. Stable Diffusion came out a bit over 2 years ago. I'm not sure that Stable Diffusion 3's, or Flux's, output is artistically _better_ than the original; it's better at following directions, and better at avoiding the most grotesque errors, but if anything it perhaps looks even _more_ generic and same-y than the original Stable Diffusion output. There's a very distinctive AI _look_ which seems to have somehow synced up between Dalle, Midjourney, SD3 and others.
You can generate AI images that do not have the "AI look":
https://ideogram.ai/assets/image/lossless/response/icQM0yZQQ...
And it's been two years since SD v1, a model that was not able to generate faces well, and it only output blurry 512x512 1:1 image without further finetuning, I tested v1.5 a few minutes ago and it's worse than I remember.
Or it will all be slop as there us no non-slop data to train on anymore
No, I don't think that's true. What will instead happen is there will be expert humans or teams of them, intentionally training AI brains rather than expecting wonders to occur just by turning the training loose on random hoovered-up data.
Brainmaker will be a valued human skill, and people will be trying to work out how to train AI to do that, in turn.