InstantX/InstantID · Lend me a hand

I'm having an idea. It may be dumb, or it may be pretty good: Could this software be modified to fix hallucinated fingers? You all know those: extra fingers or toes that usual T2I models spit out.
It would take an image as input with extra fingers or otherwise wrong anatomy and preserve the identity of everything else BUT for the hallucinated fingers. It could be a great hit IMHO, potentially being extended to other kinds of hallucinations as well, or even embedded as a processing stage of more complex text-to-image systems. The necessary tweaks don't seem that hard once you can do zero shot on faces: just changing the focus from faces to unhallucinated images. You'd probably need a large dataset of hallucinated images, but other than that, it sounds feasible. ¿Is it a very dumb question to ask? IDK how the focus is set on faces/identity so it would perhaps be hard to shift it, but perhaps not that hard. May want to use some of this too: https://huggingface.co/spaces/datasciencedojo/Finger-Counting-Right-Hand