Is trying more tools a good strategy when a search fails?

Only if the tools belong to different families. Running five near-duplicate matchers on a self-captured frame repeats the same structural failure five times.

Why is the top result sometimes obviously wrong?

Because ranked retrieval always returns its nearest candidates. When the right person is absent from the index, the top result is simply the closest stranger available.

How many attempts are reasonable before giving up?

Two or three different frames, plus a code check if any code is visible. Beyond that, additional attempts with the same inputs do not change the outcome.

Common Mistakes People Make When Searching for a Source

Most failed searches come from a handful of repeatable errors: using near-duplicate image search on a self-captured frame, reading rank order instead of similarity scores, retrying the same bad frame, and treating a performer identification as a title identification.

Failed searches are rarely bad luck. They cluster into about eight recurring errors, and most of them come from applying a method to a problem it was never built for — then repeating it harder when it does not work.

Below is each mistake, why it feels reasonable, and what to do instead.

Which mistakes cost the most time?

Ordered by how much effort they waste before the person gives up.

Mistake	Why it feels right	What actually happens	Do instead
Using Google/TinEye on a self-captured frame	It's the obvious tool	Your frame never existed as a published file, so there is no copy to find	Use face-vector search
Reading rank order, ignoring scores	The top result looks confident	Vector search always returns something, even when the answer is absent	Read the score against threshold
Retrying the same bad frame	It feels like persistence	The pipeline fails identically every time	Capture a second frame
Cropping tightly to the face before uploading	Seems to help focus the search	Removes the margin alignment needs, lowers effective resolution	Upload the frame as captured
Treating "found the performer" as "found the title"	Half the answer looks like the whole one	You end up asserting the wrong release	Confirm against filmography and scene details
Ignoring visible on-screen codes	The image feels like the primary evidence	You do image forensics when a text lookup would have resolved it in seconds	Check the first thirty seconds for a code
Trusting a filmography without visual verification	Records look authoritative	Name collisions and merged records produce wrong lists	Compare faces attached to the record
Photographing the screen with a phone	Faster than finding the screenshot key	Moiré, glare and blur destroy landmark precision	Use the device's screenshot function

Why is "just try another tool" usually wrong?

Because tools within the same family fail for the same structural reason, and switching between them changes nothing.

Near-duplicate matchers — Google Images, Bing Visual Search, TinEye, Yandex — all look for copies of your whole picture in a web crawl. A frame you captured from a player has never been published as a file, so none of them has a copy. Running all four is one experiment repeated four times.

Switching families is worth doing. Going from near-duplicate matching to face-vector matching changes the question being asked, and therefore the failure conditions. That is the only tool switch that has any expected value.

Why does the top result look convincing when it is wrong?

Because ranked retrieval has no concept of "not found". Ask for the ten nearest faces and you get ten, whether or not the right person is in the index.

This is the single most common misreading, so it is worth stating flatly: position in the list means nothing on its own. On our data, cosine similarity runs 0 to 1 with a calibrated threshold of 0.40 for video stills.

Score	What it means	Common misreading
0.31	Below threshold — treat as no match	"It's ranked first, so it must be them"
0.44	Weak; needs a second piece of evidence	"Confirmed"
0.72	Strong; probably the same person	Usually read correctly

If a tool shows you a ranked list without scores, it has removed the only information that distinguishes those three rows.

Why isn't identifying the performer the same as identifying the title?

Because a face appears in every title that performer made, and nothing in a face vector encodes which one you are looking at.

Face search answers who. Getting to which release requires a second step: narrowing by scene details — setting, wardrobe, hair length, tattoos or piercings, any visible on-screen text — against the performer's known works, then confirming with a code. People skip this step because the first answer arrives with such specificity that it feels complete.

What should you assume about coverage before starting?

That no index is complete, including ours, and that absence produces plausible-looking wrong answers rather than error messages.

Our index holds 241,792 faces across 106 sites, with 2,333 linked to a named performer (our index, 2026-08 snapshot). Two figures set the expectations honestly:

2,246 of those 2,333 named performers have their representative image from a single source. The coverage is heavily concentrated, so material from catalogues we do not index is invisible.
Our similar-performer data contains 7,004 references to identities with no name attached. A strong match with no name is a normal outcome — it means the face is indexed but the identification step never happened.

Assuming coverage is complete is itself a mistake, because it turns "not in the index" into "must be one of these results".

What does a disciplined attempt look like?

Four moves, in order, then stop:

One good frame — both eyes visible, no motion blur, paused player capture.
Read the top score against the 0.40 threshold. Above it, proceed; below it, assume the frame or the coverage, not the person.
A second frame, or a different person in the scene.
A code check, if any code appears on screen or in a filename.

If all four fail, the honest conclusion is that the material is not reachable from the sources available right now. Re-checking in a few months is more productive than a fifth attempt today.