4 ms·

Be sure to try the styles as well. Thats actually a seperate input than the prompt for SDXL, and most other UIs dont implement the style prompting.

• dragonwriter 3y ago
  > Be sure to try the styles as well. Thats actually a seperate input than the prompt for SDXL. No, its not. There are two text encoders, but they aren't really “prompt” and “style” inputs. > and most other UIs dont implement the style prompting. Most UIs default mode of operation sends the same input to both text encoders, but at least comfy has nodes that support sending separate text to them. OTOH, while there may be some cases where sending different text to the two encoders helps in a predictable way, AFAIK most of the testing people has done has shown that optimal prompt adherence usually comes from sending the same to both.
  • brucethemoose2 3y ago
    Hmm well that was a massive misunderstanding on my part.
    • dragonwriter 3y ago
      I’m not sure the origin, but using ViT-L (the encoder shared with SD1. x) for what you might call the main prompt and ViT-G (the new SDXL encoder, and also a successor to the encoder used as the single encoder in SD 2.x) for a style prompt was a common idea shortly after SDXL launched, so its understandable.
  • antman 3y ago
    What are the two text encoders?
    • dragonwriter 3y ago
      OpenCLIP ViT-G and CLIP ViT-L. The latter is the same encoder used in SD 1.x, OpenCLIP ViT-H was used as the encoder in SD 2.x, and ViT-G is, as I understand it, a successor and improvement on ViT-H.