Tuning inference strength of a controlnet by region

[skyler14]

Is there anyway to turn off or modulate the inference strength of control net in an area of an image, it seems like it shouldnt be too hard to attenuate regionally if we can already tune it globally for the whole picture. This could intuitively be coupled to something like an alpha channel to make it easily drawable (either alpha channel can be appended to annotator or you could allow drawing a seperate alpha channel to combine right before inference). If possible and would allow for alot of powerful controlnet composability since you could nicely mix and match the same controlnet (or another model) to be dominant in certain regions of the image or vice versa.

[alelordelo]

+1! This + SAM (segment anything model by Meta)