Editing as a score variable

Yes, a filter or edit can move a rating score while the subject stays exactly the same, because the tool scores the file, not the intention behind it. A smoothing pass or a colour grade can shift the number in a direction nobody can predict in advance, since the tool cannot separate an edit from its subject.

What an edit actually changes

A smoothing filter removes texture the same way a compression pass does, just deliberately rather than as a side effect. A colour grade shifts tone and contrast, which changes how any axis reading surface detail or presentation reads the image. A brightness or saturation boost changes exposure and vibrancy in ways a rubric can register as real movement, even though nothing about the underlying subject has moved at all. None of these edits are invisible to a model; a tool does not know the difference between texture that was never there and texture an edit removed, or between a colour the sensor recorded and one an app added afterward. Image models are measurably sensitive to this kind of change: Dodge and Karam (2016) tested classifiers against blur, noise, contrast and compression, and found them "susceptible to these quality distortions, particularly to blur and noise." People are moved by edits too: in a 2024 study in Royal Society Open Science, Gulati and colleagues had 2,748 raters judge faces of 462 people before and after an AI beauty filter, and perceived attractiveness rose for 96.1% of the individuals, by a median of 1 point on a 7-point scale.

Why the direction is not predictable

It would be convenient if editing reliably pushed a score one way, because then at least the bias would be knowable and could be corrected for. It does not work that way. A smoothing pass might raise a score on an axis that rewards a cleaner-looking surface and lower one that reads texture as a positive. A grade that increases contrast might help one axis and hurt another, depending on what each axis in a given rubric is actually built to reward. The honest statement is directional uncertainty, not a magnitude anyone can quote with confidence: edits move the number, and which way depends on the specific edit and the specific rubric, not on any rule worth memorising.

Why an edited result cannot be compared with an unedited one

This is the part that actually matters for anyone tracking a result over time. Two submissions are only comparable if everything that is not the subject was held constant between them, and an edit applied to one photo but not the other is exactly the kind of uncontrolled difference that ruins the comparison. It does not matter whether the edit was applied to flatter the result or by accident, through an app's default auto-enhance setting - either way, the two files are no longer the same kind of object, and a change in score between them cannot be attributed to the subject. The honest fix is the same one that applies everywhere else on this list: pick a setting - edited or unedited, and if edited, exactly which edit - and hold it, rather than switching between the two and expecting the numbers to still line up.

Why "just a little smoothing" is not a small edit

The instinct behind a light touch-up is usually to remove something that looks like a flaw rather than to change the whole image, and that instinct misunderstands what the edit actually does to the file. A smoothing filter does not selectively erase a specific blemish while leaving the rest of the surface untouched the way a person's eye would visually skip over a small mark - it applies a blur or a frequency-based softening across a region, and that region includes all the texture in it, not only the part someone wanted gone. The result is a surface that reads as uniformly smoother to a model than the source ever was, on an axis that may specifically be reading texture as informative rather than as a flaw. A "light" edit and a "heavy" edit differ in degree, not in kind - both change what the tool receives, and neither one is small in the sense of being invisible to a rubric built to notice exactly that kind of surface detail.

What this does not cover

A specific and common edit, correcting or shifting colour cast through white balance, has its own separate treatment worth reading if that is the edit in question. And the broader claim that a better-looking photo scores higher but is not a more accurate reading of anything is a distinct argument on its own, one this post assumes rather than makes from scratch.

Where the mechanism sits and where it does not

What a model does with an edited frame follows the same process as an unedited one - it has no separate pathway for detecting edits, it just encodes whatever arrived. None of this touches an actual measurement, which comes from a tape and a method rather than an image at all, edited or not. The same caution applies if the photo goes to a human judge instead - a person can sometimes spot an obvious filter where a model cannot, but a subtler edit fools both, and Rate Cock's photo guide asks for an unedited submission for exactly that reason.