The advice under the score, read sceptically

The tips-to-improve box is mostly a retention feature, not advice: its job is to get you to resubmit, ideally with a higher score. Its suggestions push toward flattering the photo rather than making the next attempt comparable to the last.

What the tips are actually written to do

A tool wants two things from the box: another submission, and a submission that scores well enough to keep you coming back. Both goals point toward the same kind of advice - change the framing, get closer, catch better light - because those are the levers most likely to move a score upward on the next attempt, not the levers most likely to make the next attempt comparable to the last one.

Those are different projects. Raising the number and standardising the input can point in opposite directions: a tighter crop or a flattering angle can genuinely raise a score while making the result less useful as data, because now it was taken under conditions the last one was not.

Why the advice leans toward flattering the photo

Tips that say "get closer" or "try warmer light" are optimising the input for this specific model's preferences, on this specific attempt, which is a narrower and more self-serving goal than it sounds. Even the direction of such tips is not a given: in Bryan, Perona and Adolphs (2012), face photos taken from within personal space drew lower human ratings of traits including attractiveness than photos taken from further away, with image size and lighting controlled. It is not advice about the tool's own rubric in any documented sense - most boxes are generic, the same handful of suggestions regardless of what actually moved the last score - and it is never advice about holding anything constant, because a tool with a retention incentive has no reason to suggest a change that would leave the score unchanged.

Contrast that with what a genuinely comparable second attempt requires: the same distance, the same angle, the same light, changed one variable at a time so any movement in the number can be attributed to something. The tips box has never once suggested that, because a fixed reference protocol produces fewer surprises, and fewer surprises is a worse retention mechanic than a box promising you can do better.

Why the tips rarely change between attempts

Look closely and the box usually says roughly the same thing regardless of what your breakdown shows, because a genuinely responsive tip generator - one that reads which axis moved and tailors advice to it - is more product than most tools have built. A generic list rotated from a small pool of suggestions is cheaper to ship and just as effective at prompting a resubmission, which is the only outcome the box is actually measured against internally. That genericness is itself a tell: advice that would be identical whether your last score was a 4 or an 8 was never built to respond to your result in the first place.

What this is not about

None of this concerns the subject of the photo, and nothing here should be read as implying it does - the tips box operates entirely on submission and presentation: distance, light, crop, angle. If you want the mechanics of actually capturing a comparable submission, the step-by-step version lives on Rate Cock, which owns that kind of how-to guide; this site's interest is only in what the tips box is for, not in walking you through using one.

Reading it usefully

Read the box as a hint about what the model responds to, not as instruction to follow before every attempt. If a tip mentions light, it is confirming light is a variable the score is sensitive to - useful information - separate from whether you should act on it before your next submission or hold your conditions steady instead. How a submission's conditions get fixed on the far side of the category, where the tool is a tape rather than a model, follows a stricter version of the same logic: control the setup, then compare.

A tip box exists because the tool is automated and needs you to act without a person in the loop; a human judge gives feedback that responds to what you actually sent rather than a templated nudge toward resubmission, which is a structurally different kind of note even when the words sound similar. How the model behind the score is actually built has nothing to do with the tips box either - the advice is a downstream product decision, not an output of the scoring pipeline itself.