How face search
actually works
The short version is one sentence: it compares the geometry of a face, not the pixels of a picture. Everything else follows from that, including why it finds photos an ordinary reverse image search walks straight past.
Why the distinction matters
Reverse image search asks “where else does this file appear?” It fingerprints the picture. Crop it, filter it, screenshot it or re-save it at a different quality and the fingerprint changes, so the search goes blind.
Face search asks a different question: “where else does this faceappear?” It measures the face itself, which means a completely different photograph of the same person — different year, camera, haircut and lighting — still matches.
- 01
Finding the face
Before anything can be compared, something has to decide where the face is. A detector scans the image for the arrangement of light and shadow that a human face produces — eye sockets darker than cheekbones, a bridge brighter than its sides — and returns a box around it.
That box is then cropped to a square and straightened, so a head tilted twenty degrees in the original becomes a head facing forward at a standard size. Every face in the index went through the same normalisation, which is what makes them comparable at all.
This is also the stage that fails most often. No face found means no signature, no search, and nothing charged.
- 02
Turning a face into numbers
The normalised crop is converted into a compact numerical signature — a list of numbers describing the relationships between facial features rather than the pixels themselves.
What survives that conversion is geometry: the spacing between the eyes, where the nose sits relative to the mouth, the proportions of the jaw. What gets discarded is almost everything you would notice first — colour, lighting, background, image quality, resolution.
This single property is the whole reason face search finds things ordinary reverse image search cannot. Change the crop, the filter or the compression and the file becomes unrecognisable to a file-matching system. The geometry doesn't move.
- 03
Comparing it against everything
Every indexed face was reduced to a signature the same way. Searching means measuring how close your signature is to each of theirs and keeping the closest.
That is a lot of arithmetic — around 1.43 billion comparisons per search — and it is why this is the expensive step. In our own completed searches it runs at roughly 81 million comparisons per second, which puts a full sweep of the index in the region of eighteen seconds.
It is also why the search is what you pay for. The comparison consumes real capacity whether it finds anything or not, so a search that comes back empty costs exactly the same as one that comes back with three hundred matches.
- 04
Turning distance into a score
The raw output is a distance: how far apart two signatures are. That gets mapped onto the 0–100 confidence score you see on every result, where 100 means the signatures are nearly identical.
The score is similarity, not probability. A 94 does not mean a 94% chance of being the same person — it means the two faces are geometrically very close. Where you draw the line between 'match' and 'coincidence' is a judgement, which is exactly why we show you the number rather than a green tick.
- 05
Assembling the report
Matches come back ranked by score, each carrying a thumbnail and the URL of the page it was found on. We classify each source — social profile, news article, blog, forum, adult platform — so the shape of the result set is readable before you open a single link.
Results are stored for thirty days so you can return to a report you paid for. Your original upload is not: it was the query, not the haystack, and it is deleted once the search completes.
The numbers behind it
Measured from our own completed searches and read live as you load this page — not a specification sheet.
What the index is
An index of faces found in publicly accessible images: social profiles that are set to public, news photography, blogs, forums, archived pages. We do not crawl it ourselves — ARGUS is built on a third-party face index and pays for every search run against it.
That means the boundary is the same one a signed-out visitor faces. Private accounts, friends-only posts, direct messages and anything behind a login are not in it, and no face search anywhere can reach them. Coverage is also uneven: photo-first platforms are represented far more thickly than video-first ones, and anything published since the last crawl is missing.
Where this stops working
Knowing the mechanism is half of it. The other half is knowing the failure modes — sunglasses and reshaping filters, twins, age, and the demographic variation that face recognition has as a technology.
The honest limitations →