Stable Diffusion 0.60 (Patreon)

Here we go. There was quite some changes to the code for this one.

Most of the changes are self explanatory, some bug fix or more options to select.

The main change is that this version does not contain models. You need to download them from here:
Models:
https://mega.nz/folder/2x1CBRJL#MzMfyijznMs-2t-VqBXYPQ
Driver Mirror:
https://drive.google.com/drive/folders/1MqjzOisTaV0Y1fBfYRVRS9fG6UrFOkK7?usp=share_link

GUI:
https://mega.nz/file/v5k3QRKY#HM4zRswCx8UR-Y1odj8IoNjOiuYS4FWhsfFTD46jEAE

Driver Mirror:
https://drive.google.com/file/d/12h4bxCZjbrRoRXkswEzACvIvI1nc9Ixm/view?usp=share_link

Edit: Well, thanks a lot Mega. You suck. I though that by having a paid account it would give a good upload rate, but nop... Need to find a host that actually work.

That way you can select what models you want and I can easly add more models in the future.

In the future I also plan to split the torch library, that is 90% of the GUI size. So is easier to download.

Every single .rar file already have the correct folders. You can just drop them inside the "./models" folder inside the GUI folder.

You can also inside the GUI go into configuration to select the models folder, that way you don't need to cut/paste the models on each new update.

Another big change is in the dreambooth tab. There is a few more options.
I will update soon the Tutorial with the new options, but mostly are self-explanatory.

The most complex option is LR Scheduler that give two options.
It control the way that the Learning Rate updates. There is no better option, like most models trainings, you just need to test what work better for you.
Here the links if you wish to undestand each scheduler.
https://pytorch.org/docs/stable/generated/torch.optim.lr_scheduler.ReduceLROnPlateau.html
https://pytorch.org/docs/stable/generated/torch.optim.lr_scheduler.OneCycleLR.html

For now I uploaded the more used models. 1.4 1.5  2.1 anything and the upscalers.

Edit: Oh yeah, another big change:
The .ckpt files used to require some code from sd_1_5.pt to work.
I don't want this to happen anymore, since now models are optional.
So now, the first time you want to use a .ckpt file, it may download some models from huggingface. It should only happen once. If possible in the future will make that this download don't happen.

Edit 2: It seen to have some problem with kuda kernes around the xattention code? still looking into it to find the problem

Edit 3: Found the problem, someone made a few new tests to see if xattention is working, but if it does not work, it break the script

Edit 4: I think I fixed the error for some users. There is a new link for the GUI, please test it out if you are having a problem with the previous build.

Edit 5: Older .PT files will not work on this version. But now the .PT is very generic and should not break ever again. Another thing, some new schedulers require the double of steps. So if you select 30 steps, some new schedulers may do 60 steps.

Edit 6: Loading trained dreambooth .pt files seen to result in a error, already fixing it, don't worry, you will not lose your trained model, it will work once the GUI is fixed.