Photo or video for family resemblance testing? The science verdict
A frequent question: "will I get a better resemblance result with a video instead of a single photo?" Spoiler: the answer is no, in 95% of cases. Here is why, with data and limits.
What really matters when measuring resemblance
The resemblance score is calculated from facial embeddings: a several-hundred-dimensional vector that summarises a face's geometric "signature". To generate this embedding, the AI needs:
- A front-facing face (or angle limited to 30°).
- Uniform lighting (no harsh shadow zones).
- A neutral or smiling expression (not grimacing).
- Sufficient resolution (min. 200 px between the eyes).
All these criteria can be met in a single good photo. A video provides extra frames, but with very variable angles and often a moving expression — therefore noise that degrades average quality.
Photo vs video: technical comparison
Why Look Like Me only uses photos
Our technical choice: 3 photos (you, dad, mom) instead of a video. Three reasons.
- Consistent accuracy: with a controlled-quality photo, the score is reproducible. Re-running the same test gives the same result. With a video, the score can swing ±5 points depending on the analysed frame.
- Preserved privacy: 3 images is exactly the minimum needed. A 5-second video contains 150 images — 50× more biometric data exposed to the server, for almost zero accuracy gain.
- Speed: 30-second analysis vs 1-2 minutes for a video (upload + frame extraction + multi-analysis). On a consumer site, speed is a major conversion factor.
It is not that video is technically impossible — it is that it brings more complications than benefits.
When video would be useful (the remaining 5%)
As stated, photos win in 95% of cases. But here are exceptions where video would have a marginal advantage:
- Very young babies (<6 months): they move constantly, hard to get a clear front photo. A video would let you automatically extract the best frame.
- Person with particular facial mobility (paralysis, post-stroke asymmetry): multiple angles help the AI converge to an averaged representation.
- Animals: a cat or dog will not pose for a photo — video captures the useful instant.
For these edge cases, the Look Like Me solution is: take 5-6 quick photos rather than just one, and keep the best. Our algorithm accepts a second attempt if the face is not detected.
Tips for the best possible photo
To maximise the accuracy of your test, here is the checklist:
- Soft daylight (window, no direct sunlight).
- Front-facing face, gaze towards the lens.
- No sunglasses, cap, mask.
- Hair pulled away from forehead and ears.
- Neutral or slight smile (not a wide grin that distorts the mouth).
- Distance 30-60 cm for phone, 1-2 m for a camera.
- JPEG format, min. 800×800 px.
With these criteria respected, the score is stable, reliable, and video would add nothing.
❓ Frequently asked questions
Could a video mode be developed in the future?
Technically yes. But adding it would first require proving it significantly improves accuracy on everyday cases — which internal tests have not shown. For now we stick with 3 photos.
Can I extract a frame from my video and use it as a photo?
Yes, this is the best approach for difficult cases (especially babies). Capture a clear frame, export to JPEG, and use it as a photo. Our AI will process it normally.
Does video allow detecting family expressions (identical smile...)?
For that yes, theoretically. But the "resemblance" measured by our service is geometric (face shape), not behavioural (way of smiling). Expressions are another topic, still little explored scientifically.
How many photos do I need to send minimum?
3: one of you, one of your father, one of your mother. More adds nothing — each image is analysed individually, it is not an average.