A usable frame needs exactly one thing: a face whose eyes, nose and mouth corners are all visible, captured sharply enough that those points can be located precisely. Everything people worry about — lighting, background, resolution of the overall image — matters far less than that.
The reason is mechanical. Face search detects five landmarks, uses them to warp the face into a standard pose, then crops tightly and discards the rest. If the landmarks are visible and sharp, the frame works. If they are not, no amount of image quality rescues it.
What makes a frame good or bad, specifically?
Judge the face, not the picture.
| Factor | Good | Marginal | Fails |
|---|---|---|---|
| Head angle | Front-on to ~30° turn | 30–45° turn | Strong profile, back of head |
| Occlusion | Face fully visible | Hair across one eye | Hand, object or heavy shadow over the eyes or mouth |
| Sharpness | Static or slow motion | Slight blur | Motion blur smearing the features |
| Face size in frame | Face fills a good share of the image | Medium shot | Wide shot, face a few dozen pixels across |
| Lighting on the face | Even, either bright or dim | One side in shadow | Silhouette, blown-out highlights |
| Compression | Native player capture | Re-saved JPEG | Screenshot of a screenshot, heavy artefacts |
| Background | Irrelevant | Irrelevant | Irrelevant |
That last row is not a joke — it is the most useful thing on the table. People discard frames because the room is dark or cluttered. Alignment crops the room away entirely. Judge the face.
What are the steps to capture one?
- Scrub to a moment where the performer faces the camera. Dialogue, opening scenes and reaction shots are usually the richest sources. Moving forward five seconds costs nothing; fighting a bad frame costs a lot.
- Pause fully. A paused frame is sharp. A frame grabbed during playback, especially on a laggy stream, can catch a decoded partial frame.
- Use your device's screenshot function, not a phone camera pointed at the screen. On macOS use the built-in capture; on Windows use the Snipping Tool or Print Screen; on mobile use the hardware button combination.
- Capture at native resolution. If the player is windowed and small, maximise or go full-screen first. You cannot recover detail you never captured.
- Do not pre-crop tightly to the face. Leave the frame as captured, or crop generously. The aligner needs margin around the landmarks.
- Save as PNG if given the choice. PNG is lossless; a low-quality JPEG re-save adds artefacts exactly where the fine facial detail lives.
- Check for other faces worth trying. If the scene has more than one person, you have more than one chance at identification.
Why does a profile view fail when a human can identify it easily?
Because a turned head hides landmarks the aligner needs, and the encoder was trained overwhelmingly on faces that were not turned that far.
At a strong profile, one eye is invisible and one mouth corner is often occluded. Alignment then estimates positions it cannot see, and the warp is wrong. The vector that comes out is a confident description of a face that does not quite exist. It will still return results — it will just return the wrong ones, at scores that look acceptable.
The practical rule: if you can see both eyes clearly, proceed. If you can only see one, find another frame.
Which of several frames should you pick?
When you have options, rank them by how much identity signal survives the pipeline.
| Priority | Frame type | Why |
|---|---|---|
| 1 | Close-up, front-facing, paused, well-lit | Maximum landmark precision |
| 2 | Medium shot, front-facing, sharp | Fewer pixels, but geometry intact |
| 3 | Close-up at a moderate angle | Angle costs accuracy but is recoverable |
| 4 | Cover art or promotional still, if available | Often the exact material an index was built from |
| 5 | Wide shot or steep angle | Last resort |
Row 4 is worth knowing about. Index entries are built from covers and stills — ours holds 241,792 faces drawn from that kind of material across 106 sites (our index, 2026-08 snapshot). If you can find the cover image rather than a video frame, you are searching with material closer to the index's own distribution, and scores tend to reflect that.
How do you know the frame was good enough?
Read the score, not the ranking. Cosine similarity runs 0 to 1, and our calibrated threshold for video stills is 0.40.
If your best result sits well below threshold, the frame is the first suspect — before the index. Vector search returns its nearest candidates whether or not the right person exists in the index, so a list of low scores is ambiguous between "bad frame" and "not indexed". A second frame of the same person resolves that ambiguity for free, which is why it is always the next move.
Boundary worth stating: even a perfect frame cannot find someone we never indexed, and our sources are concentrated — 2,246 of 2,333 named performers have their representative image from a single site. A good screenshot maximises your chances; it does not create coverage.
Related questions
- How to find a video from a single screenshot
- How face search works, step by step
- Why can't I find the title I'm looking for?
- Common mistakes people make when searching for a source
- Is uploading a photo to a search tool safe?
Once you have a frame with both eyes visible and no motion blur, upload it here. Detection runs in your browser, so you will find out immediately whether a face was found — and if none was, nothing was sent at all.