Open CCpt5 opened 2 months ago
hmm..
It seems override_settings
feature has been removed by forge itself.
you can compare original process_images()
https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/master/modules/processing.py#L825-L839
and forge's process_images()
https://github.com/lllyasviel/stable-diffusion-webui-forge/blob/main/modules/processing.py#L797-L819
forge abandon the original override_settings feature. override_settings
will be simply ignored.
(please see https://github.com/lllyasviel/stable-diffusion-webui-forge/commit/6921420b3fa6c7e69f3f83fdba88b2c071a321a9#diff-1fe5b15d29310b37af0c4560da365990f4e6eee7b99bf3f041cfa18ebb11610fL780 )
Ahh so basically this feature will no longer be possible?
Ahh so basically this feature will no longer be possible?
override_settings feature is supported by webui. so It's up to the forge's decision.
Ahh so basically this feature will no longer be possible?
override_settings feature is supported by webui. so It's up to the forge's decision.
Understood. I explained on the open discussion thread about this issue that's on the Forge repo here: https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/976
I'm ok w/ you closing this if there's no workaround. Thx!
So to be clear, this only breaks the ability to use an alternate checkpoint in Forge, but it still works as a good alternative to adetailer for flux in Forge if using the same checkpoint as the base image?
yes. some basic features will work as expected.
@wkpark There was a new commit to Forge this morning that suggests it restored the override settings. I gave udetailer a try where the 2nd pass is supposed to use an SDXL model to inpaint the face and still didn't see a result, but this may be my error.
Can you confirm whether or not this update -should- fix udetailer so it can once again use a different checkpoint during the second pass? Thanks!
Re:
Hello again!
I'd love to use your extension w/ Flux/Forge to use some of my older Dreambooth trained likeness models to clean up faces on Flux generations. Unfortunately it seems that feature broke on both your extension and adetailer during all the massive updates to Forge the past couple months.
Is there any chance you could try this feature out on your side and if you can replicate perhaps try to come to up w/ a fix? At the moment (as mentioned in this thread on the adetailer issues page) the extension goes through the motions of masking the image and seemingly running SDXL inferencing, but the image remains the same.
If you need any specific information from me lmk. I'm using the latest version of Forge and uddetailer.
Thanks!!
Issue on addetailer page (closed due to inactivity but can confirm it is still an issue w/ that extension as well):