Stable Diffusion Dirty Hack to run 2.0 (Patreon)
Content
Not really a regular build, this is just a hack so everyone can test the 2.0
It probably broke some options, but at least it render Txt2Img with 2.0
It can work with Xformer OR Attention Slice Disabled (So it wil use more Vram if not using Xformer)
There is only 2 models in this Hack:
Stable Diffusion 2 Base:
- Work just like the older models.
Stable Diffusion 2:
- This model was trained with 768x768 images. So use that if you select this model.
- This model also require a new type of Scheduler. They are the two new schedulers in the list.
- V-Pred: DDIM and V-Pred: Euler Discrete
- This model will only work with those!
A huge shotout to the guys at Diffusers ! They did the heavy lifiting, I'm just implementing then in the GUI.
Dreambooth may still work? Did not test it, but it might.
Link:
https://drive.google.com/file/d/1G3Dt6g-4dM0z-9lha7RvRUjnhwcExO0d/view?usp=share_link
Mirror:
https://drive.google.com/file/d/1ZU_i4aQJPZQh81aFDU7l8U0cdMmm3huc/view?usp=share_link
More Info:
V2 was trained with a new CLIP and new Dataset. Keep this in mind:
There is almost no NSFW images on the new dataset. It might make worse human bodies?
No famous people on the Dataset: It will probably not generate the face of a famous person.
New CLIP don't have that much artists name. It might not be able to simulate your favorite artist style.
Most likely you will need to create new prompts for this new version of SD