• @Zeth0s
    link
    English
    5
    edit-2
    1 year ago

    Neither. Going from a prompt to an image is a stochastic non-linear transformation based on billions of parameters.

    Your brain also performs stochastic non-linear transformations of inputs. Just in a different way.

    • Natanael
      link
      fedilink
      English
      21 year ago

      Do I have to explain with math how my high level abstract reply applies?

      Most generative ML rely on probabilities. The averages are over multidimensional complex data structures representing patterns extracted from the inputs. Like average faces when you prompt it for a face (try training it on different sets of faces and look at how the output differs, you really do see it retain averages of the patterns in the input such as average skin color and haircuts). I wasn’t talking about linear arithmetic averages.

      • @Zeth0s
        link
        English
        3
        edit-2
        1 year ago

        My comment is that they simply are not averages, that’s it.

        As a simpler example, it is like saying that a polynomial plus some noise is an average… It’s simply not.

        The stochastic and non linear parts are the reason these models create original images, unless overfitted.

        If it was a weighted average you’d have identical, smoothed, most likely non sensical images for identical prompts.

        And this is not the case.

        That’s all my comment.

        • Natanael
          link
          fedilink
          English
          0
          edit-2
          1 year ago

          You still misunderstand my use of “average”. I am once again not talking simple averages over simple arithmetic numbers.

          Look at the outputs of models trained on majority white faces vs diverse faces. If you still don’t understand what I mean by averages then I guess this conversation is hopeless

          Yes there’s noise in the process.

          But that noise is applied in very specific ways, it still fundamentally tries to output what the training algorithm indicated you would expect from it given the prompt, staying in a general neighborhood preserving the syntax / structure / patterns in the input training data related to the keywords in your prompt. Ask for a face without more details and you get a face looking like the average face in the input, usually white in most models, western conventional haircuts, etc, because that’s representative of its inputs, an average over the extracted structure in the inputs. The noise just tweaks some selection of representative features and their exact properties. It is still close enough to average that I feel it is fair to call it average, because it so rarely output extremes (other than when the model just breaks down and produce nonsense).