You have one frame and no title, no code, and no filename. The workable route is not to search for the picture — it is to search for the face inside the picture, then work outward from the performer to the title.
That reordering is the whole method. A screenshot is a unique image that exists nowhere else on the web, so nothing can match it as an image. A face, by contrast, appears on hundreds of covers and stills, and can be compared mathematically across all of them.
What do you actually need in the screenshot?
One face, visible, and not turned too far away from the camera. Everything else is optional.
Face recognition works by locating landmarks — the eye centres, the nose tip, the mouth corners — and using them to straighten the face into a standard pose before encoding it. If those landmarks are hidden, there is nothing to align, and the pipeline stops before matching even begins.
| Element in the frame | Effect on a face search |
|---|---|
| Front-facing or slightly angled face | Best case — landmarks all visible |
| Strong profile (over ~45° turn) | Detection may succeed, matching degrades sharply |
| Face partly covered by hand, hair or object | Usually fails detection |
| Heavy motion blur | Landmarks become unreliable |
| Very small face (a few dozen pixels wide) | Too little detail to encode meaningfully |
| Background, room, furniture | Ignored entirely — contributes nothing |
| On-screen text, logos, timestamps | Ignored by the face step, but useful to you later |
The last two rows surprise people. Face search deliberately throws away the background. That is why an otherwise "bad" screenshot — dim room, cluttered frame — can still work perfectly, as long as the face is clean.
What are the steps, in order?
- Pause on a frame where a face is clearly visible. Prefer a moment where the performer is looking toward the camera. Scrubbing forward a few seconds is almost always cheaper than trying to rescue a bad frame later.
- Capture the frame at native resolution. Screenshot the player, not a photo of your screen. Avoid re-cropping down to a tiny square — the detector does its own cropping.
- Upload it and let detection run. Detection happens in your browser here: the image is analysed locally, and only a cropped face region plus five landmark coordinates are sent onward.
- Pick the right face if several are found. Group scenes return multiple detections. Choosing the wrong one silently sends you down the wrong path.
- Read the similarity scores, not just the order. A ranked list always returns something. The score tells you whether it means anything.
- Confirm against the performer's known titles. Once you have a name, match your frame's specifics — hair length, wardrobe, setting, tattoos or piercings, any visible text — against covers and stills.
Step 6 is the one people skip, and it is where most wrong answers come from. Face search gives you a person, not a title. The title is your inference, and it needs a second piece of evidence.
How do you read the similarity score?
Face vectors are compared by cosine similarity, a number from 0 to 1. On our data the threshold calibrated for video stills is 0.40 — above it, two faces are treated as probably the same person.
| Score range | How to treat it |
|---|---|
| Below 0.40 | Below threshold. Treat as no match, however plausible it looks |
| 0.40 – 0.50 | Weak. Needs independent confirmation before you rely on it |
| 0.50 – 0.65 | Reasonable. Still worth checking a second image of the same person |
| Above 0.65 | Strong. Usually the same person unless the frame is degraded |
A result at 0.42 and a result at 0.85 look identical in a list and mean very different things. Any tool that hides the score is hiding the part you need to make the judgement.
Why does the index matter more than the algorithm?
Because a face that was never indexed cannot be returned, no matter how good the comparison is.
Our index holds 241,792 faces with 2,333 named performers, drawn from adult covers and frames across 106 sites (our index, 2026-08 snapshot). That material simply does not exist in a general-purpose image index, which is why general-purpose tools return nothing useful regardless of their engineering.
Boundary worth stating plainly: our coverage is uneven. Of the 2,333 named performers, 2,246 have their representative image sourced from a single site. That is a real skew. A title from a catalogue we have not indexed, or a performer who never appeared on the sources we cover, will not be found — and the tool will still show you a ranked list of near-misses. Reading the scores is what protects you from that.
What should you do when nothing plausible comes back?
Change the input before you change tools. In order of how often it helps:
- Try a different frame. A second screenshot from another moment in the same scene is the single highest-yield move.
- Try a different person in the scene, if there is more than one.
- Check for a code on screen. Many releases burn an ID into the intro, the cover shot, or a watermark. A code turns an image problem into a text lookup.
- Accept the possibility of absence. Amateur uploads, very old releases, and content from unindexed catalogues may genuinely have no entry.
Related questions
- How to take a screenshot that face search can actually use
- How face search works, step by step
- Why Google reverse image search doesn't work for adult videos
- Why can't I find the title I'm looking for?
- Common mistakes people make when searching for a source
If you have the frame open already, upload it here and see what comes back. Detection runs in your browser — the original screenshot never leaves your device, only a cropped face region is sent for matching, and nothing is kept afterwards.