n00mkrad / text2image-gui

Somewhat modular text2image GUI, initially just for Stable Diffusion
GNU General Public License v3.0
935 stars 95 forks source link

OptimizedSD outdated and more #82

Open MaggotHATE opened 1 year ago

MaggotHATE commented 1 year ago

I know that OptimizedSD is of low priority to you, but it was updated a while ago with negatives and more samplers. I can provide ready to use files for the program, but that by itself is only one part.

  1. Since OptimizedSD supports negatives now, it makes sense to show the text field for it. Been using negatives for a while, works really well, makes OptimizedSD actually usable.
  2. Samplers seem to work, and euler_a is especially useful. Again, UI hides selection in Optimized mode.
  3. "state_dict" problem is still present on OptimizedSD by default, but can be resolved the same way as normally (with _transform_checkpoint_dictkey). Models seem to work as intended.
  4. I am not sure about _split_weightedsubprompts - it's weird to use as is, and not splitting at all seems to give better results, for some reason.

Now, there's also safetensors support, which:

  1. can work on OptimizedSD - safetensors.torch is already in modules, and I can rename models to .safetensors.ckpt to make them selectable in settings;
  2. cannot work in normal mode due to safety check.

Does this mean that safety check for OptimizedSD is broken? Anyway, I'm happy being able to use safetensors models without any issues, but it's a bit weird to see a safetensors model being labeled as malicious or corrupted only because it was renamed.

Like I said previously, I can upload ready to use scripts for OptimizedSD if you don't have time to spend on it - all changes are taken from either the original repo, or Automatic, it's nothing complex. I can use it as is, but it would be nice to have it out of the box.