• barsoap@lemm.ee
    link
    fedilink
    arrow-up
    5
    ·
    edit-2
    1 year ago

    The Stable Diffusion 1.5 base model seems to recognise the art style from prompt, though it’s a bit spotty and it doesn’t seem to understand it well. None of the fine-tuned models I have understand it, some even spit out realistic images instead of some kind of line art.

    The theme “woman devouring her son” isn’t well-understood, either, in many examples it simply seems to interpret it as “anguish”, it’s not a given that you even get two subjects.

    It generally wants to… avoid the theme? Never seen ancestral euler differ so much from euler. “eating, anguish, female, male” is the gist of the prompt it can’t make more sense of it CLIP isn’t GPT.

    As to the outputs: Unusable in general, though have one to prove I’m not talking out of my arse, you can load it up in ComfyUI (unless imgur strips that info, also, the setup is trivial).

    If it was an AI model it doesn’t seem to have been SD. Maybe SD 2 but I don’t have the base model lying around and none of the downstream models that I have are anywhere close to fine-tuned for shoddy corporate art. No, I won’t download 2G worth of floats just for this post this has already been unproductive enough as-is.

    Taking Goya’s “Saturn devouring his son” and running it through img2img would likely result in something usable enough to sift through and find something decent, am too lazy to try right now. SD really benefits from being given non-textual directions.