"Every few months someone on my team asks whether we can ""just run a face search"" on an image. The answer is usually no, and the confusion comes from mixing up two very different things. Face comparison takes two photos and asks whether they show the same person. Face search takes one photo and hunts for matches in a database or across the web. They sound similar. They are not the same problem, and pretending otherwise leads to bad product decisions.
Here are three myths worth clearing up before you write a line of code.
Myth one: it can find someone from a single photo
This is the big one, and movies are mostly to blame. A 1:1 comparison tool has no index. It does not crawl social networks, it does not know who is in the picture, and it cannot tell you where else that face appears. Give it one image and it has nothing to compare against. Give it two and it will tell you how similar they are.
That limitation is also the point. If you are building identity verification, you usually already have both images: the document photo on file and the selfie the user just took. You do not need a haystack. You need a careful look at two straws. Detection — locating a face in the frame — is a separate step from comparison, and a tool can be good at one and indifferent to the other.
Myth two: a similarity score is a yes-or-no answer
Scores are continuous, and thresholds are a product decision, not a technical fact. A comparison engine returns a number. What you do at 78 percent versus 84 percent depends entirely on your tolerance for false accepts versus false rejects, and those tradeoffs differ wildly between an internal admin login and a public signup flow.
I like to test this manually before wiring anything up. Take a set of known pairs — same person in different lighting, siblings, strangers who look vaguely alike — and run them through something like a browser-based Face Comparison App to see how the numbers actually distribute. CaraComp runs in the browser on desktop or phone, so you can build a rough intuition in an afternoon without installing anything or provisioning a service. What you learn is that siblings score higher than you expect and that bad selfies score lower than you expect. Both facts should shape your threshold.
Myth three: photo quality is a minor detail
Resolution, angle, expression, and lighting move scores more than most people assume. A heavily edited image — the kind of thing a faceapp-style filter produces — can smooth away exactly the geometry the model relies on. Sunglasses, a hard side profile, or a face that occupies fifty pixels of a group shot will all degrade results.
So treat input quality as a first-class concern in your UX. Prompt users for a straight-on shot. Reject images where the detected face is too small. If you are comparing two selfies taken years apart, expect a wider spread and say so in your interface rather than presenting a bare percentage as if it were a verdict. Photo research and lookalike comparisons are legitimate uses, but they benefit from the same discipline as verification work.
None of this makes face comparison less useful. It just means the honest version of the feature is narrower than the version people imagine, and narrower features are easier to ship, explain, and defend."
Top comments (0)