Home Artists Posts Import Register

Content

Changelog summary:

  • Fixing a few bugs like using to big of a prompt.
  • Fixing a bug on .ckpt importer that would not correctly import the text converter, if this one was also trained.
  • Adding support to .ckpt importer if it was created using pytorch lighting.
  • .ckpt importer now support custom VAE and Hypernetworks, that some custom models like novelai already use.
  • Negative prompts
  • [Maybe] More schedulers
  • Improved inpainting/outpainting


Changelog not summary:
Between this and the last update I spent quite some time just undestanding the SD code. There was some aspects that was very familiar to me, like the pipeline and UNET, other parts of the code was all new, like the schedulers.

Let's talk a little about schedulers.
It's no secret that there is already quite a few codes that run SD on k-diffusion, but there is a detail about this:
SD was released with a very simple code, just to run the model. Most users work on top of that, including the codes that already have k-diffusion.
I'm developing the GUI on top of Diffusers, that is a high active library that is constantly improving SD library. The downside is that it currently still don't have k-diffusion working.
I don't like just copy+paste code, so I took the chance to implement it myself  on the Diffusor library. This way I can learn how schedulers actually work while contributing a little to the code.
For now my code is running, but the results are bad. I'm currently in contact with a few nice fellows from the project to see if we can get it working today.
If I can't make it work today, i'm sure it will be working in a few more days.

Something about the Weight Modifiers:

Usually used as ((())) to change the weight of a prompt, it's the same problem that schedulers, but this is not such a complex problem, I just didn't had the time to fully understand the code and implement it on Diffusers, with just a little more time I can make it work.

Dreambooth:

Stable Diffusion is evolving so crazy fast, it's already possible that Dreambooth will be replaced by Hypernetworks? Man, you can't even blink without a breakthough lol
For now I'm focusing on dreambooth, if it change to hypernetwork, at least the GUI can already run those.

Other stuff:
I spent quite some time during the weekend just playing around with the GUI, so I have quite some ideas to improve the quality of life of the GUI, lets go =p

Comments

No comments found for this post.