Photo-condition testing: lighting, angle, blur and expression
The people can stay exactly the same while the input photos change. This test isolates common image conditions so score movement has a practical explanation.
Why the photo itself matters
Face similarity is calculated from images, not from a person standing in perfect studio conditions. That means a score can move even when the people do not change. During manual testing we deliberately vary ordinary photo conditions to understand when a result is stable and when the input image becomes the main source of uncertainty.
Conditions we check
| Condition | Typical effect | Preferred test photo |
|---|---|---|
| Lighting | Deep shadows can hide shape around the eyes, nose and jaw. | Even, natural or indoor light with visible facial detail. |
| Angle | Strong side angles change visible geometry and proportions. | Front-facing or mildly angled face. |
| Blur | Fine landmarks become uncertain. | Sharp image without motion blur. |
| Crop / distance | A very small face contains less usable detail. | Face large enough to inspect clearly. |
| Filters | Beauty filters can change skin, jaw, eyes or nose. | Unfiltered image whenever possible. |
| Expression | A wide smile or unusual expression changes mouth and cheek geometry. | Natural expression for the baseline check. |
What we look for during repeated comparisons
We do not expect every photo pair to produce an identical percentage. Instead, we look for understandable behavior. If a clear frontal pair gives one result and a heavily blurred, side-angle pair moves significantly, that difference has a plausible image-quality explanation. If small harmless changes produce wild contradictory decisions, that is a stronger signal that the comparison needs investigation.
A practical two-step test
- Baseline: compare two clear, reasonably frontal photos with good lighting.
- Challenge: replace only one photo with a harder version — for example a different angle or lower-quality image — and compare again.
This simple method changes one major variable at a time. It is more informative than replacing both photos and then trying to guess why the score changed.
Why we do not “chase” a preferred score
Repeatedly swapping photos until a percentage matches an expectation is not a fair test of any comparison tool. It turns image selection into a way of steering the answer. For development checks, the expected identity or relationship is known first, and the photo changes are chosen to test robustness rather than to manufacture a particular number.