[Bug]: ControlNet breaks outpainting real images #290
Description
What happened?
working on a project with the imagenet dataset, where I try to collage them next to one another and inpaint in the gaps, so as to make it sort of believable. It kind of works with the standard stable diffusion inpainting model, where the constraints set by the original image are kind of respected. But once I try to use ControlNet with a model like "epicrealism" for example, the outpointing gets completely broken. The ControlNet model I'm using is Kataragi_inpaintXL-fp16 [d20c87b4] with inpaint_only+lama set as the preprocessor. It doesn't seem to work with outpainting a generated dream either. Or am I doing something wrong? Please let me know.
Steps to reproduce the problem
if you want to reproduce it download the Kataragi_inpaintXL-fp16 [d20c87b4] ControlNet model with realisticVisionV60B1_v51HyperVAE.safetensors as the base stable diffusion model and try to make a dream and then extend it with ControlNet outpainting.
What should have happened?
ControlNet should extend the original image/dream by outpainting, yet it completely ignores it. It crates a harsh divide, which is basically a new image with no relation to the previous one.
Commit where the problem happens
latest main branch
What platforms do you use to access openOutpaint?
MacOS
What browsers do you use to access the UI ?
Google Chrome
Browser Extensions/Addons
grammarly, my bib
AUTOMATIC1111 webUI Commandline Arguments
--api --cors-allow-origins=https://zero01101.github.io
Additional information
No response