Tilt the camera and the geometry the model sees changes

A camera pointed at the same subject from a different angle produces a different silhouette, and silhouette is one of the more direct things a vision model has to work with. Nothing about the subject changed between the two shots. What changed is which surfaces face the lens and how much each one foreshortens, and that alone is enough to move a shape-sensitive score.

Foreshortening, briefly

When a surface tilts away from the camera, its length in the photo shrinks relative to its actual length, an effect called foreshortening. A cylinder photographed dead-on shows its full length; the same cylinder photographed from an angle that tilts it toward or away from the lens shows a shorter apparent length, purely as a function of the viewing angle, with nothing about the object itself different. This is ordinary projective geometry, and it applies to any three-dimensional object photographed from any angle other than the one that happens to show its true proportions clearest.

Why the silhouette is what gets scored

A vision model does not receive a three-dimensional description of a scene. It receives pixels, and the outline and contour within those pixels are among the strongest signals it has for inferring shape, alongside the shading cues covered separately in how lighting direction changes what the model reads. Change the angle, and the silhouette changes - which parts appear longer, shorter, wider or narrower relative to each other - and the model, having no independent way to know the true three-dimensional form, scores what the silhouette shows it. This is the same underlying issue that makes symmetry judgements angle-dependent: a small change in where you stood changes the projected outline more than the subject itself ever could. A model attempting to infer depth from this single flat silhouette runs into the same wall from a different direction, and what monocular depth estimation can and cannot recover from one image explains why guessing the third dimension back out is not a fix.

Downward angles tend to foreshorten length and can make proportion axes read differently than an angle closer to level; angles from the side change which surface is most visible and can obscure or emphasise particular contours. There is no single "correct" angle in the sense of one that produces the true shape and every other angle producing a distortion - every angle is a projection, and every projection foreshortens something. Level, front-on framing is simply the projection that tends to match what most training photos looked like, which is a statement about the training distribution, not about geometric truth.

How this differs from a lens effect

Angle and focal length are separate mechanisms that can compound. Focal length and distance exaggerate near-versus-far regardless of tilt; angle changes which surfaces are visible and how foreshortened they are, regardless of lens. A photo taken close, with a wide lens, from a steep downward angle stacks both effects, and there is no way to look at the resulting image and cleanly attribute how much of the apparent shape came from which cause. Even the direction of the tilt matters on top of that: a downward angle and an upward angle foreshorten different parts of the same subject, so two photos shot from opposite tilts of the same magnitude are not mirror images of each other in what they distort.

What to take from this

This is a description of a mechanism, not instructions for finding a flattering angle - the practical version of that question belongs to the photo guide Rate Cock has already written, and repeating it here would just be a worse copy of a guide that already exists. What matters for understanding a result is that angle is one of the inputs behind the spread you'll see testing the same subject repeatedly, and it is one of the more controllable ones, unlike the model's underlying calibration. Holding angle fixed matters for anyone trying to do an actually controlled comparison across attempts, since an uncontrolled angle change is enough on its own to explain a shift that might otherwise look like the model being inconsistent. A person reviewing a photo brings the same geometric intuition a viewer always has, and what a human judge notices that a model does not includes recognising when a shot was taken from an unusual angle, which changes how much weight they put on it. Reading a single number without accounting for the angle it came from is a habit worth dropping before drawing any conclusion from it.