The person stays the same. The input changes.
Lucentage detects a face, prepares a small face crop and estimates age from that crop. Moving the camera or changing the light changes what reaches the model. A group photo adds another complication: the service selects the largest face.
NIST illustrated this behavior in its 2024 evaluation: estimates from six algorithms varied between frames of the same person with different expressions and eyeglasses. That is evidence about the evaluated algorithms, not a benchmark result for Lucentage.
Try a small, controlled comparison
Use photos of yourself, or an adult who has agreed to the uploads. Take them in the same session so that the comparison is about photography. These are instructions for your own experiment, not published Lucentage test results.
- Take a baseline photo with even light, a relaxed expression and the camera near eye level.
- Take a second photo with the same position and expression, changing only the light direction.
- Return to the baseline lighting. Take a third photo with a slightly different camera angle.
- Upload each original file and write down its estimate and the condition you changed.
- Keep every result in your notes, including results that do not fit your expectation.
Record conditions alongside the numbers
A result is easier to interpret when you can trace it to a specific input. Copy this plan into your notes and add the estimates you actually receive.
| Photo | Change | Keep consistent |
|---|---|---|
| Baseline | None | Original file, one face, no optional filter |
| Lighting | Direction of the light | Camera position and expression |
| Angle | Camera angle | Baseline light and expression |
Separate consistency from accuracy
A model can return similar estimates for several photos and still be consistently wrong. It can also vary between photos while one result happens to be close to the recorded age. Agreement is not proof of accuracy.
Subtracting the lowest estimate from the highest describes the spread for these particular photos. It is not a confidence interval. Averaging the estimates does not create a verified age, because the errors can share the same bias.
An identical file is a separate comparison from a new photo. Keep the file, browser and model version consistent when investigating repeatability; a screenshot or re-export is no longer the same input.
Use the comparison to improve the photo
If you notice that a poor-quality photo behaves differently, choose a clearer photo for future use. Resist selecting only the youngest result as the best one. The useful outcome is a repeatable setup and a better understanding of the estimate.