Chris Saunders

Like to also report this PR is working great in conjunction with Face ID IP Adaptor, greatly improved likeness using multiple source images.

> > I added it back to my own installation. > > Thank you @huchenlei for making the changes it is as simple as overwriting 5 files. > > I'm...

This looks really easy to implement it's just a custom diffusers pipeline. They setup a pipeline for both the small and large models then create a custom tstitch pipeline which...

Self Attention Guidance triggers it pretty much 100% of the time for me now. It was working at some point but in the middle of a batch I started getting...

Exactly the same thing happening to me under Windows just so you know it's not an isolated incident. torch==2.1.0+cu121 with xformers using SDXL. I can provide my logs when I...

> Thanks for adding this @saunderez I think it looks good - do you want to run `make quality` so the code quality checks can pass? I've tried but unfortunately...

All ready to go @yiyixuxu ! Happy to be a contributor for once, hopefully there'll be more to come :)

We have been testing the usage of DEIS as an alternative scheduler for training in the dev branch the Dreambooth extension for Automatic1111 WebUI. It is currently a user selectable...

> To further add color to @saunderez use-case, we've found that using DEIS in place of DDIM/DDPM for the noise sampler improves training speed without any affect to output quality,...

> @williamberman @saunderez As I understand this, the `get_velocity` function should indeed be the same for all `Scheduler` (sampler) classes given the same noise schedule parameters, but must be implemented...