What an image model is actually doing when it scores you

The single most common misunderstanding about AI rating is that the model is measuring something. It is not. There is no ruler, no reference object and no depth information anywhere in the pipeline. What actually happens is closer to recognition than to measurement, and once that lands, most of the odd results people report stop being odd.

The pipeline, in four steps

A photo goes in as a grid of pixels. A vision encoder turns that grid into a vector - a few hundred to a few thousand numbers that encode what is in the image, at what scale, in what arrangement, under what lighting. This vector is the only thing anything downstream ever sees. The original pixels are gone.

A scoring head then maps that vector onto whatever scale the system uses. It was trained by being shown many images alongside ratings that humans gave them, and adjusting itself until its outputs matched. That is the entire mechanism. The model has no concept of a centimetre. It has a concept of what images that people scored highly tend to look like, which is a completely different thing that happens to correlate well enough to be useful.

Finally, a language model usually writes the prose - the verdict, the roast, the breakdown - conditioned on the score and the vector. The writing is downstream of the number, not the other way around.

Why it has no idea how big anything is

A photograph contains no scale information. This is not a limitation of AI specifically; it is a property of projecting three dimensions onto two. A phone camera fifteen centimetres away and a phone camera a metre away can produce identical framings of differently sized subjects, and nothing in the resulting image distinguishes them.

Humans get around this using context - a hand, a familiar object, a doorframe. Models do the same thing, which is why a photo containing a hand behaves differently from one that does not. It is also why the number moves when you change your distance to the subject, which most people discover by accident and interpret as the model being broken. It is not broken. It is doing the only thing available to it, which is inferring scale from cues, and you changed the cues.

If what you actually want is a number in centimetres, no image model will give you one reliably, and a tape measure and a repeatable method is the honest answer.

What it is genuinely good at

Consistency, mostly. A model given the same input twice returns close to the same output, which is more than can be said for human raters, whose agreement with themselves a week later is famously poor. It does not get bored, it does not get generous at the end of a session, and it does not know who you are.

It is also good at the properties that survive the encoding: symmetry, proportion, surface condition, framing quality, and the general "does this look like images people rated well" question. Those are real signals, and a system built on them tells you something true about the photograph even though it tells you nothing verifiable about the subject.

The practical consequence

Score a photograph, get a score for a photograph. Every property of the image - the lighting, the angle, the lens distance, the crop, the background - is inside the result, because the vector encoded all of them and the scoring head never saw anything else.

That is why the same subject can span a wide range of scores across a single afternoon, and why the variance is bigger than almost anyone expects before they test it.

Multi-axis systems make this legible rather than fixing it: when a tool reports separate scores for shape, texture and overall impression rather than one aggregate, you can see which part of the judgement moved. Cock Rate is one of the tools that reports the breakdown rather than just the total, which is the property that makes a result worth reading twice. Why that decomposition matters is its own subject.