Classifier free guidance(CFG) on different prediction types and karras style schedulers about diffusers

[maxs-kan]

To me, this step seems useless since we simply multiplied by a constant and added the constant (at this time step) and then did the reverse operation. However, maybe I'm missing something

[sayakpaul]

We take this into account during training in some of our training script. Here is an example:

diffusers/examples/dreambooth/train_dreambooth_lora_sdxl.py

Line 1650 in d2b10b1

if not args.do_edm_style_training:

For inference, this is automatically handled in the step() method of the underlying scheduler being used. Cc: @yiyixuxu if she has any more inputs here.

This is perhaps better asked as a discussion?