• 0 Posts
  • 682 Comments
Joined 11 months ago
cake
Cake day: August 21st, 2024

help-circle
  • not necessarily. image generation models work on a more fine-grained scale than that. they can seamlessly combine related concepts, like “photograph”+“person”+“small”+“pose” and generate plausible material due to the fact that all of those concepts have features in common.

    you can also use small add-on models trained on very little data (tens to hundreds of images, as compared to millions to billions for a full model) to “steer” the output of a model towards a particular style.

    you can make even a fully legal model output illegal data.

    all that being said, the base dataset that most of the stable diffusion family of models started out with in 2021 is medical in nature so there could very well be bad shit in there. it’s like 12 billion images so it’s hard to check, and even back with stable diffusion 1.0 there was less than a single bit of data in the final model per image in the data.