#AIArt #StableDiffusion Sorry for noob AIArt question but I'm fascinated. Why when I search for "dog" does it not just return basically a search for dogs in its sample set? Are there random "filters" or "combinations" or whatever it's doing that it executes as a default? And if so, is there a way to turn those off? Certain this is cringe for some, sorry but learning.
@CardboardRobot
I don't really understand the question. When you talk to something like stable diffusion, you aren't doing a search at all!
And it doesn't have a sample set that it keeps around and uses. What it has is a big neural network that was trained with a lot of pictures of dogs and things, but the training set isn't kept around.
What happens is that from the training set the neural network has learned what a dog looks like in general, so when you say " dog " it will produce a picture that it thinks is a picture of a dog, with some other random stuff that it thinks sort of belong in the dog pictures.
Hope that helps!
@CardboardRobot yeah, it's really freaking wild frankly! You feed a billion images with descriptions to this neural network in training mode, then you take the images away so that only the network is left, put it into generation mode, feed words in, and images come out! Sometimes really amazingly surprising images. And sometimes more or less just what you asked for. It's all magic! :) I think everyone involved was surprised by how well it's worked. Interesting times!
@ceoln Thanks, that actually does help a lot! I thought it was freshly combining things from some data set it was scouring. Not my field, just fascinated like I say. Though still confused why it thinks a dog looks different from what it's learned so far. But my learning set is pretty meager at this point on the subject.