Sd GUI 0.63 (Dreambooth training fix) (Patreon)

This took some time in debug, because I had to train models multiple times to find and fix the code, in the end I had to rewrite quite a lot of the training code.

It probably still have some problems with float32 training, but float16 should work a lot better.

Link [With and without torch folder]:

https://drive.google.com/drive/folders/1gxxaTx0Xth5CCguzdXUEsCI9wV0A_rSq?usp=share_link

Been also working on a new training method that may work better to teach styles to the model, but it still WIP, not in the build.

There is also now the version for each model on its name, so if the naming is wrong, it is probably on the wrong folder:

• models_sd: V1 Models
• models_V2: V2 Models
• models_inpanting: models for in(out)painting
• models_upscale: Models that do upscale
• models_vae: VAE models
• models_hyper: Hypernetwork models