/vt/ - Virtual Youtubers

Alright, here is DPM++ 2m (Karras) v2

A little over a month ago, someone identified a bug in k-diffusion that would allow a certain value to go over 1, which while not causing any visible issues, is a bug which needed to be patched out.

You can find the code and the instructions on how to implement the fix here: https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/8457

Contrast has been affected though with this change (as in it has been lowered), although that may be a good thing depending on your model/vae.

NOTE: The changes between the samplers is very slight, I would recommend zooming in on the full sized image to see differences beyond just the contrast.
https://files.catbox.moe/h78ae5.png

>>47545503
I deleted it because I forgot the image.