The best photos for a face match

This is the single biggest factor. The proportion match measures five ratios, such as face width to height and the gap between the eyes. When a head turns to the side, the face looks narrower in the photo and the eyes look closer together. The ratios shift even though the face has not changed.

Head tilt and looking up or down have a similar effect on the height ratios. The face descriptor is more robust to angle, but it is not immune. A straight-on photo, like a passport shot but relaxed, gives the most consistent numbers.

Face detection and landmarks work best when the whole face is lit. Window light from in front of you is ideal. Avoid a bright window behind you, which turns your face into a silhouette, and harsh overhead light, which hides the eyes in shadow.

Very dim photos add noise. The expression model may read a calm face as sad or fearful, which lowers the mood match.

If a photo has several faces, the tool shows the faces it found and lets you choose one. That works well. Problems start when faces are very small or overlapping, because tiny faces may not be detected.

For the cleanest result, crop to one person or use a close shot. With a group photo, make sure the face you want is reasonably large and clear. For a selfie with both of you in it, the One photo together input picks the two largest faces for you, and Group ranking scores every pair in a photo of up to 8 faces.

Beauty filters can reshape eyes, jaw and nose. That directly changes the landmark ratios. Cartoon and AR filters can stop a face being detected at all.

Sunglasses hide the eyes, which are key landmarks for both proportions and the face descriptor. Masks and hands over the mouth hide the mouth ratios. Clear glasses are usually fine. Light colour edits are fine too.

• Fine: clear glasses, normal makeup, mild brightness or colour tweaks.
• Risky: strong smoothing or face-reshaping filters, heavy shadows.
• Avoid: sunglasses, masks, stickers or hands over the face, cartoon filters.

Mood match is 40% of the blend, the biggest single part. It compares seven expression probabilities from each face. Two big smiles look very similar to the model. A smile next to a straight face does not.

If you want a high score, agree on the expression. If you want a fair comparison across several friends, ask everyone for the same expression, either all smiling or all neutral.

The face should be clearly visible, with sharp eyes. A face that fills a decent part of a normal phone photo is plenty. Big, high-resolution files do not help much, because each photo is downscaled on your device before analysis. That keeps things fast and private.

Very small, blurry or heavily compressed images can fail. Screenshots of screenshots are a common culprit. If the tool says no face was found, a sharper copy of the same photo often fixes it.

A face match compares two photos, so differences between the photos show up in the score. If one photo is a bright outdoor selfie and the other is a dim indoor shot from below, the numbers will reflect that as much as the faces.

Same light, same angle and same distance is the fairest setup. Taking both photos in the same spot, a minute apart, works well.

Because angle, light and expression all feed into the measurements, two different photos of the same pair will give different scores. That is normal. The method is deterministic, so the same two photos always give the same result, but a new photo is new input.

If you are curious, try three or four pairs and look at which signals move. You will usually see mood match change the most, and proportion match move with head angle.

You can drag and drop photos onto the tool, paste one straight from your clipboard, browse your files, or take a photo with your phone camera. Whatever you choose, the photo is read on your device and never uploaded.

Once your photos are ready, the couple photo test, crush photo test and friendship photo test all use the same method. The how it works guide explains the maths behind each signal.