Iranian computer network engineering MsC. He/Him. FOSS developer, strong believer in GPL. ProleWiki editor (Maintainer of the Persian instance).

The character in pfp is an OC. his name is da’ezho’ . Have fun pronouncing it.

Also available on Matrix and Discord with the same name.

  • 0 Posts
  • 65 Comments
Joined 1 year ago
cake
Cake day: April 8th, 2024

help-circle
















  • Not concepts as we humans operate on, but concepts in the feature extraction sense. You have a gross misunderstanding of how these models work.

    Imagine it as how an image detection model learns how to detect images. It extracts features (that are vectors of data that are completely meaningless to humans). Image generation can be somewhat imagines as the inverse of a feature detector, it creates features from noise.

    These models are not plagiarizing your work. You teach it the shape a dog is expected to have, and you teach it the color black. Without such thing existing in the training set, it can generate a black dog.

    This is fundamental to stable diffusion otherwise it could never work at all.




  • To actually use it properly, install ComfyUI and get in deep. The whole “actually using SD to generate what you want consistently” is a form of art on it’s own, it’s kind of like learning Krita.

    Because generating random bullshit ala ChatGPT is easy. Getting anatomy, object placement, logical consistency in the image and also overcoming model biases or introducing new biases is some complicated shit. For example, there are models that can generate depth layers, models that can dictate object placement, and then since SD output is smol you need to wire it into an upscaler model, etc.

    SDWebUI is much simpler, but also much more limited. Same as Android apps that run SD locally, really nothing matches the level of control that ComfyUI has.

    All you need is any recent Nvidia GPU with 4GB of VRAM although less VRAM was reported to work. For AMD, since AMD hates GPGPU apparently, only their top cards support ROCm. As for Apple and Intel, I do not know.

    If you are stuck with AMD and can’t use ROCm, just use KoboldCPP. It will be significantly shittier in every way but atleast it runs.