Rare features pull the embedding somewhere odd

A tattoo, a piercing, jewellery, a scar - anything visually unusual in a photograph pulls the image toward a part of the model's learned space that had fewer similar examples during training. Fewer examples means less stable scoring, for the same reason any statistic estimated from a small sample is noisier than one estimated from a large one.

What "pulls the embedding" means concretely

Every photo a model processes becomes a vector - a point in a high-dimensional space that the model places it in based on everything visible in the frame. Ink, metal and unusual marks are visually prominent, high-contrast, high-detail features, and they move that point noticeably relative to a similar photo without them. Where the point lands matters, because scoring is a function of location in that space: nearby points, from images the model saw often, tend to get consistent, well-supported scores; points in emptier regions get scores extrapolated from whatever was nearby, which is a weaker basis for a number.

Why this produces instability rather than a predictable direction

It would be a fabrication to claim tattoos or jewellery systematically raise or lower a score - there is no published, checkable basis for that claim, and it would depend entirely on what happened to be in a specific model's training data, which is not public. What can be said honestly is about stability, not direction: an unusual feature moves an image into a less densely sampled region, and scores from less densely sampled regions vary more across near-identical inputs, repeat runs and model versions, than scores from densely sampled ones do. The instability is the finding. The direction is not knowable from outside a model's training pipeline.

How this differs from a distribution-shift claim generally

The broader idea that a model is only calibrated on what resembles its training data is covered elsewhere on this site as a general property of any unusual photo. This note is about a specific, common category of "unusual": a visually salient feature added to an otherwise typical frame, rather than a photo that is atypical in composition, lighting or subject overall. The mechanism is the same; the trigger is narrower and easier to recognise, because you can usually point to the exact feature - the ink, the piercing - that is doing the pulling.

What to do with this

Nothing needs removing or hiding - this is a description of how a model behaves, not advice about presentation. What is worth knowing is that a surprising or inconsistent score on a photo with an unusual feature in frame is not necessarily a signal about anything beyond the model's own thin training support for that feature. Rate Cock reporting an axis breakdown makes it possible to see whether an odd result concentrates on axes near the feature or spreads evenly, which is at least a starting point for reading it. Retaking a photo with the same feature and comparing repeats is a more useful test than a single number either way, and how much repeat variance to expect before a result means anything is worth reading first, alongside the general habit of reading a score rather than just taking it that Penis Rater covers. None of this touches physical measurement, which Measure My Cock handles with a method that does not depend on where an image sits in a learned space at all, and a human reviewer reads an unusual feature consciously rather than as an unlabelled pull on a vector, which is the structural difference Rate Penis covers between the two approaches.