|
Now I am curious if we get in the technical side of things how this could play out.
For what I gather, the IA for standard diffusion at leasts, when it gets a prompt to create such and such, then goes to make an image of pure random pixels or noise, then it uses the knowlege it got from looking at millions of pictures, not the pictures itself, and correct me if im wrong here, so the files in the IA are not said images but the knowlege gained from looking at them. So then the IA proceeds to sort the random pixels into what it thinks could be the most probable order to do what we ask from it.
So if it is as I think it works, then I dont know if we can say that any image used during training is used directly on the end result of the stable diffusion process, yes the AI knows about those images and might use them as reference, but the images are not shipped along the AI and only the knowlege gained from the learning is shipped on the files that then are used to generate whatever it is asked from the AI.
TLDR: The original images are not distributed or used directly but what the AI learned from them is what then creates the pictures we want.
|