Replies: 3 comments 9 replies
-
i appreciate that you're looking how to improve instead of just stating its broken! |
Beta Was this translation helpful? Give feedback.
-
Vlad, a question: do you think it's worth collecting here the issues I file regarding my workflow (for actual bugs, not enhancement requests)? I think my tests have uncovered at least one which I'm currently trying to narrow down before filing. |
Beta Was this translation helpful? Give feedback.
-
Hm, I'm starting to get the hang of it. I'll write down a list of things that may help others, and for some of these I'll file issues:
|
Beta Was this translation helpful? Give feedback.
-
Hello,
I've been testing a bit Control and I've found a few roadblocks towards implementing this workflow. This isn't posted as an issue because a. I might be getting things wrong and b. it is a discussion on how Control works, without saying things "it's broken" or "it doesn't work", but rather "Is Control, in its current state, able to support workflow X?"
My workflow in the original backend with ControlNet
My adapted workflow with Control
The effect of the resize options in the pipeline is different, or possibly I could be using it wrongly, because it results in a single pass + resize, which causes a noticeably drop in image quality at the end (the original backend does two actual passes, the diffusers backend just one), as the resize is done without following up with additional sampling steps.
An idea I've been toying with is to supply the generated image back to Control (feedback) and do a second pass manually, but this is more complex and error prone than having it in a single pipeline.
EDIT:
I've done it and in that case works as expected:
So, should one want to replicate what the ControlNet extension does in the original backend, two steps are required.
Hence the questions:
Thanks.
Beta Was this translation helpful? Give feedback.
All reactions