Something like "Restore Faces"? Or variable Noise region?

[mr-september]

My workflow often uses Ultimate Upscale at around 0.3 noise after inpainting, etc. This noise is great for accentuating prompts and fixing distortions/bad 'patching' done by inpainting steps.

However, it also almost always changes the faces just enough that they essentially become different people. Not always graphically bad, just not the original identities. Conversely, if the noise is set to ~0.1-0.2, the facial identity gets retained, but so do the other glitches that I was Ultimate Upscale would've fixed at slightly higher noise values.

I tried enabling the normal img2image "Restore Faces" option, but that doesn't seem to help.

Is there any other way to retain facial identities? Or would it be possible as a feature request?