Most "the result is wrong" messages we get come down to the photo, not the model. The neural network behind a face test reads shapes, proportions, and texture, and a photo can hide all three. Here are the rules that matter, in the order they matter, with the reason behind each one.
1. Face the camera, keep your eyes level
A three-quarter turn changes the apparent width of your face, hides one cheek, and moves your nose off center. The character stills we match against are mostly front-facing, so a turned photo is being compared with a shape it doesn't share. Look straight into the lens and keep the line between your eyes horizontal. A slight tilt is fine; a dramatic one isn't.
2. Light from the front, not from behind
Backlight (a window behind you) turns your face into a dark shape with no detail. Strong side light paints half your face in shadow, and the model reads that shadow as a feature. The best light is soft and in front of you: face a window in daytime, or stand under even indoor lighting. Avoid a single overhead lamp, which sinks your eyes and sharpens your nose shadow.
3. Fill the frame, but don't get too close
The face should take up a good part of the image. A tiny face in a wide shot leaves the model working with a few dozen pixels of detail. At the other extreme, a phone held close to your face uses the wide-angle lens at its most distorting: the nose grows, the ears shrink, and the outline of the face becomes rounder than it is. Hold the phone at arm's length, or use the rear camera with a friend, or take a step back and crop afterwards.
4. Neutral or lightly smiling
A wide grin lifts the cheeks, narrows the eyes, and stretches the mouth: it reshapes the lower half of the face. Since most character stills show a default expression, a neutral face or a small smile matches the way the characters are usually drawn or shot. If you want to see how expression changes things, try both and compare; it's one of the biggest single variables.
5. Skip filters and beauty modes
Skin smoothing, eye enlargement, jaw slimming, and "glow" effects remove or invent the exact details the deeper layers of the network rely on. A filtered face is, to the model, a slightly different face. Use the plain camera. Screenshots of an unfiltered photo are fine; the model doesn't care where the pixels came from as long as they're real.
6. Clear the face
Hair over one eye, a hat brim shadowing the forehead, a mask, and dark sunglasses all remove signal. Regular glasses are fine unless they reflect strongly. Tie hair back or push it aside so the outline of your face is visible from forehead to chin.
7. One face only
If two faces are in the frame, the model reads both, and the vector becomes an average of two people. Crop to just yourself.
8. Resolution and focus
Anything above about 500 pixels across the face is plenty. Blur matters more than size: a sharp small photo beats a soft large one, because blur removes texture.
9. Test with two or three photos
Take a couple of photos under slightly different conditions and run each. A match that appears at the top every time is robust. If the winner changes with every photo, the top few characters are simply close together for your face, which is also useful to know: look at the similarity gap rather than the winner alone.
Quick checklist
- Front-facing, eyes level
- Soft light from the front
- Face fills the frame; phone at arm's length
- Neutral face or small smile
- No filters, no beauty mode
- Hair back, no mask or sunglasses
- One person in the photo
- Sharp focus
None of this requires a good camera. A well-lit, unfiltered phone selfie at arm's length is ideal, and it takes ten seconds.