AI recognition is local. You can totally disable external access and it will run fine (obviously assuming your photo storage is local).
Their recognition model is loose. It works and it works well enough for what this is to be useful but it's just a quick classifier. It will absolutely have misses.
You can optionally aadd in a coral dev board to save yourself some cpu load but I've never found it necessary.
Is the non-accelerated model also a quantised version? The EdgeTPU is very efficient, but you can get a significant performance drop unless you take care when converting. Modern CPUs can do classification fast enough that you could do background processing and unless you have really tens of thousands of images, you'd be done in an hour or two
Their recognition model is loose. It works and it works well enough for what this is to be useful but it's just a quick classifier. It will absolutely have misses.
You can optionally aadd in a coral dev board to save yourself some cpu load but I've never found it necessary.