Left-right flips, and when they matter

Flip a photo left to right and, on its own, almost nothing happens to the score. That is a deliberate property of how most vision models are trained, not a coincidence, and it is worth separating from everything else that is different about an actual mirror selfie.

Why the flip itself is a non-event

Training pipelines routinely mirror images at random as one of the cheapest ways to double the effective size of a dataset, on the reasoning that a photo and its mirror image show the same content and should be treated the same way. The habit is as old as modern image models: the network that won ImageNet in 2012 was trained on random crops "and their horizontal reflections", which Krizhevsky, Sutskever and Hinton (2012) credit with preventing "substantial overfitting". A model raised on that habit learns representations that are close to flip-invariant: an embedding for a photo and an embedding for its mirror land near each other in the model's space, because nothing during training rewarded telling them apart. So a pure, clean, digitally-flipped copy of a photograph will usually score within noise of the original. That is not understanding; flipping was simply never a signal worth learning to use.

Where an actual mirror photo differs

A selfie taken in front of a mirror is not a flipped photo of the same shot - it is a different shot entirely, and several of the differences matter more than the flip does.

The angle is rarely identical, because holding a phone up to a mirror changes hand position and distance in ways a front camera does not, and camera angle changes apparent shape, and the model follows regardless of which direction the resulting image faces. The light bounces off glass rather than reaching the lens directly, which softens and sometimes yellows it depending on the mirror and the room, a variable closer to what happens with lighting direction generally than to anything about flipping. Mirrors are rarely perfectly flat or perfectly clean, and both compress and smear detail in ways a digital flip never does. A phone case, a hand, or the edge of the mirror frame often enters the shot too, adding objects a direct photo would not have.

None of these differences are about orientation. They are about the photo being a genuinely different capture, and the model is responding to that, not to the fact that it happens to be mirrored.

What this means in practice

If you are comparing two photos and one looks off, checking whether it was a mirror shot is a reasonable first question, but the answer is rarely "the model penalised the flip." It is almost always the glass, the angle, or the light. The wider list of things that move a score between otherwise similar photos covers this alongside crop, distance and framing, and mirror shots are best read as one entry on that list rather than a phenomenon of their own. Penis Rater's notes on reading a score make the same point from the user's side: a single number without that context is easy to over-read.

None of this touches what the mirror shot is actually of. Any question about how a specific tool wants a photo framed belongs to the tool, not to the mechanism - Rate Cock's photo guide covers that ground directly. A human reviewer, for what it is worth, notices a mirror shot immediately and can factor it in explicitly rather than absorbing it into a number - Rate Penis's judge pages describe what a person brings to that kind of read. For anyone trying to get a comparable pair of images across attempts, Measure My Cock's gear notes cover the equipment side of holding conditions steady, mirror or camera, which is the more useful lever than worrying about which way the image faces.