Comments by "clray123" (@clray123) on "Rick Beato"
channel.
-
6
-
6
-
3
-
2
-
2
-
1
-
@1wibble230 Err.. I train AI (language) models. I know how it's done and what they can reproduce, and I can assure you there is no "sensory input" and that it is not at all how humans would learn or create anything.
The target function in training for which loss is minimized is to reproduce the original dataset faithfully. What you are confusing for "creativity" are residual errors and side effects, the paths not covered by the dataset, but in some spatial vicinity to it, or "hallucinations" as they are called by some AI folks. (AI researchers have historically a tendency to introduce misleading anthropomorphizing teminology at every step, lying to themselves all the way.)
As for your fancy shmancy image prompts, I can assure you that if the model was not trained on something resembling your prompt in the first place (as individual image labels), it would not be able to react in any sensible way to your prompt. You get what just you put in, mixed up. Try for once applying some really creative, detailed prompt, rather than parroting the "style prompts" prepared as examples to you, and see what you get out of these models then.
1
-
1
-
1
-
1
-
1
-
1