Same faces, different photos: what can change in an AI comparison?
A useful way to understand a face-comparison tool is to keep the people the same and change only the photos. The exercise shows why a score should be treated as an image-based estimate rather than a permanent number attached to two people.
Start with a baseline pair
Choose two clear portraits with similar lighting and near-frontal head positions. Run the comparison and treat that result only as a baseline for the experiment. Do not try to decide whether the percentage is objectively “right.” The purpose is to see what happens when the visual input changes.
Change only one photo
Keep the first person's image unchanged and replace the second image with another clear photo of the same person. If the score moves, the people did not change — the visual information did. This is one of the clearest demonstrations of why a similarity percentage is photo-dependent.
Test angle separately
Compare a frontal portrait with a second image where the head is turned. A side angle changes the apparent spacing and visibility of facial features. The nose profile, jaw contour and one side of the face become more prominent while other information becomes less visible.
Test lighting separately
Next, keep the pose reasonably similar but use a photo with stronger shadows or backlighting. Shadows can hide detail and alter the apparent depth of facial regions. If the result changes again, lighting is contributing to the numerical comparison.
Test expression separately
A broad smile raises the cheeks, changes the eye area and stretches the mouth. Compare a neutral portrait with a smiling one and observe whether the result is as stable as the baseline.
What this experiment actually teaches
The valuable finding is not a particular percentage. It is the amount of variation across reasonable photos. If several good photos give similar results, the comparison appears relatively stable under those conditions. If the result moves substantially, the image choice is playing a larger role.