Apparent symmetry is a property of the viewpoint

A symmetry score depends heavily on where you stood, because it judges an outline, not the subject. The outline is a projection, produced by where the camera sat, and a small change in that position changes the projection more than any real asymmetry in the subject does.

What the model is actually looking at

A vision model has no access to the three-dimensional form behind a photograph, only the silhouette and shading that a single viewpoint happened to produce. Symmetry, read off pixels, means left and right halves of that silhouette matching - roughly equal widths, roughly mirrored contours, roughly balanced shading. That is a real and readable property of an image. It is also entirely a function of the angle the photo was taken from, in a way most people do not expect until they see it happen.

Why angle dominates

Photograph the same subject dead-on and the two halves of the frame tend to balance. Rotate the camera even slightly off that axis and one side moves closer to the lens than the other, which under perspective projection makes that side appear larger - not because anything about the subject changed, but because nearer things are bigger in a photograph regardless of their real size. The size of the effect is larger than intuition suggests: modelling faces photographed at about 12 inches, Ward and colleagues (2018, JAMA Facial Plastic Surgery) calculated that the nose, the feature nearest the lens, appears about 30% larger in men and 29% in women than in an undistorted projection. A few degrees of rotation, or a hand or camera held slightly off-centre, applies that same nearer-is-bigger effect unevenly to the two sides, producing an asymmetry in the image that has no correlate in the subject at all. The model cannot tell the two apart, because both look like the same thing once flattened to pixels: an outline whose left and right sides do not match.

Lighting compounds it. A single light source to one side puts more shadow on the far half, which reads as texture and shading asymmetry on top of whatever the silhouette is already doing.

What this means for a single result

One symmetry score, from one photograph, is closer to a report on how square the camera was to the subject than a report on the subject's actual symmetry. This is the general geometry problem that runs through most of what a model cannot verify from a photo, applied to the one axis where it is easiest to demonstrate: take two photos from slightly different angles and watch the symmetry number move on its own. Repeated shots from a consistent, squared-on angle are the only way to get a symmetry read that means something across attempts, which is part of why scores from a single afternoon spread more than people expect.

Rate Cock reports shape and proportion as their own axis rather than folding symmetry into a single blended number, which at least makes it possible to notice when a result looks angle-driven rather than real. A person reviewing the same photo can often tell the difference between an odd angle and an odd subject in a way a score alone cannot, and a fixed, repeatable angle is exactly what a manual method controls for that a single upload does not. Comparing how consistently different tools handle this is a fair question to ask before trusting either one.