Unless the AI has a hard-wired cutoff - which would be reasonable, because its engineers don't want it coughing out too many false matches - this will be a gian

Unless the AI has a hard-wired cutoff - which would be reasonable, because its engineers don't want it coughing out too many false matches - this will be a giant mess of probabilities. There are more metrics for image quality than pixel count. And that's before you get to the lighting on the subject, angle the image was taken from, clothing & hair, etc. Add how distinctive the person's face is. And the quality of your match set images. And the size of those match sets - are you asking the AI "is this one of the 5 Pete's Pizza delivery drivers, to unlock the back door for him?" Or asking it "who is this unknown person in NYC?". And how sure do you want this "positive" match to actually be? The back door at Pete's Pizza is (probably) far lower stakes than a murder conviction.