Everything about submitting more than one image

A single photo has one thing that can go wrong with its comparability: the conditions it was taken under. A set or a series has that same problem multiplied by every additional image, plus a handful of problems a single photo never has to deal with - how the images combine, whether their order matters, whether they were even taken under one condition to begin with.

The through-line for all of it is the same: a set is only as comparable as its least controlled slot. One inconsistent image, one undisclosed combination rule, one unlogged device swap, and the whole set inherits that weak point regardless of how carefully everything else was held.

The unit: slots, not photos

The building block of a comparable set is the slot - a defined position with its own angle, distance, light and crop, specified before the photos are taken rather than assembled afterward from whichever images turned out well. Building a set properly means writing each slot down in enough detail that a future set can fill the same slots again and produce genuinely comparable images, not just similar-looking ones.

Two or three slots is usually the practical ceiling. Each additional slot is one more condition that has to be reproduced identically every time the set is retaken, and a slot that cannot be held constant is worse than not having it, because it introduces variance that looks like signal.

The combination rule: what the total is actually summarising

Once a set exists, a tool that accepts multiple images returns one number, and how it got there is rarely disclosed. Four plausible rules - simple average, best-of, weighted by position, or scoring one internally-selected dominant image - all produce different totals from the same set, and a result page gives no indication which one you are looking at.

The test is cheap: submit a set with one image deliberately weakened and see whether the total drops proportionally, barely moves, or drops sharply. Three or four submissions settle the question, and the answer changes how the resulting number should be read - as a genuine average of the set, as a ceiling drawn from its best image, or as a single-photo score in a multi-photo wrapper.

Averaging across angles: the specific case that hides the most

A multi-angle set is the combination-rule question applied to angle specifically, and it deserves its own treatment because what averaging across angles hides is not noise, it is signal. Each angle reads proportion differently - foreshortening compresses length along the line of sight by an amount that depends on tilt, so a front-on and a three-quarter photo of the same subject are measuring genuinely different projections, not two noisy samples of the same one.

Averaged into a single total, the gap between angles disappears and the result looks more stable than the underlying set actually was. The per-angle spread, not the mean, is where the set's information lived, and it is worth asking a tool to surface the per-image results rather than accepting the average as the whole story.

Order: a variable most people never think to control

Some tools weight the first uploaded image, or the last, more than the rest of a set - a plausible product decision made by whoever built the upload flow, and one that is never stated on the upload screen itself. Whether order changes the result is answered by one test: submit the same images twice in different orders and compare the totals. If they match, order is not a variable for that tool. If they diverge, order joins angle, distance and light on the list of conditions a comparable set has to log and repeat.

Mixed conditions: the failure that breaks everything else

None of the combination rule or order questions matter if the set itself was never taken under one condition. A mixed-condition set - one photo level, another tilted; one under window light, another under flash - has no single condition behind it, and its score describes a blend the combination rule cannot untangle and a future set cannot reproduce, because there was no protocol to repeat in the first place.

This is the sharpest version of the through-line: a mixed set is not a weaker comparable set, it is not a comparable set at all. Committing to one condition per set, every slot filled the same way, is the minimum bar a set has to clear before any of the other questions in this guide are worth asking.

Choosing the unit: one photo, or a set

Given all of the above, a reasonable question is whether a set is worth the overhead at all. A single photo versus a set trades reproducibility for information: one slot is simpler to hold constant and answers "did this change since last time" with fewer moving parts to doubt; a set of two or three angles answers "does this read differently depending on viewpoint," which a single photo structurally cannot.

Most comparability needs are served by a single well-maintained slot, repeated. A set earns its extra maintenance cost specifically when angle-dependence is the question in front of you, and even then, only if every slot in it can actually be held to the same standard as a single photo would be.

The lens variables that sit underneath every slot

Two more terms are worth knowing regardless of whether the submission is one photo or a set. Field of view and crop factor describe how much of the scene a given lens captures at a given distance, and phones commonly switch between two or three lenses automatically depending on distance or zoom. A set that mixes lenses between images without noticing has introduced a hidden variable into every slot it touches, on top of whatever else was being held constant.

Picking one lens, usually the main one, and using it for every slot in every set removes this variable at no cost beyond checking which lens a given shot actually used.

Series over time: keeping a set comparable for a year, not a session

A single comparable set answers a question at one point in time. A series held over a year is a different and harder problem, because three more things have to hold across the whole stretch, not just within one session: the protocol, the device, and the tool itself.

Slow drift is the main risk - standing slightly closer than three months ago, a moved lamp, a loosened crop - none of which looks like a problem in any single session, all of which compound over a year. A control shot, one fixed reference image resubmitted alongside every real set, is the cheapest way to separate drift in the subject from drift in the setup: if the control moves, something in the protocol moved with it.

When something breaks outright - a new phone, a tool update, a rubric recalibration - the honest response is to log the break and start a fresh comparable stretch, not to keep appending to the old series and hope the discontinuity averages out invisibly.

Lenses inside a single set, not just across sessions

The lens question is not only a series-over-time concern. Within one set, a phone can switch between its main and wide lenses automatically depending on how far back you stand or how much you pinch to zoom, which means two slots in the same set, taken minutes apart, can end up using two different fields of view without any visible cue in either photo. Field of view and crop factor describe exactly this: how much of the scene a lens takes in at a given distance, and how that compares to a reference focal length. A set built from slots that each specify the lens as well as the angle and distance closes this gap before it opens.

A short worked example

Take a two-slot set - front and three-quarter - submitted monthly for six months. In month one, the front slot scores 6.9 and the three-quarter scores 7.4; the tool reports an average of 7.15. By month four, the front slot has drifted to 6.6 and the three-quarter to 7.6, for the same average of 7.1 - a total that looks essentially flat. Read separately, the two slots tell a different story: one angle trending down, the other trending up, which is the kind of divergence averaging across angles is built to hide. Nothing about the flat average would have prompted a look at either slot individually. The per-slot numbers, logged alongside the conditions that produced them, are what make that divergence visible at all - and whether it reflects the tool's own drift or something in the protocol is exactly what a control shot, checked against the same log, is there to settle.

What holds all of this together

Every section above reduces to the same instruction: define the condition before you take the photo, and hold it the next time. A slot, a combination rule, an upload order, a lens choice, a protocol over a year - each is a place a set can quietly stop being comparable to itself, and each is fixable by writing the condition down rather than trusting memory or habit to reproduce it.

None of this is a claim about which set scores higher. A tool that shows results per axis rather than folding everything into a single figure - Rate Cock does this on public entries - makes several of the failures in this guide easier to catch, because a proportion axis that jumps between two supposedly-matched sets is a visible prompt to check the slot, the lens, or the order before assuming anything about the subject changed. The vision models behind these tools process each frame on its own terms; what they extract from a single image does not know about the rest of a set's history, which is exactly why the comparability work described here has to happen on the photo side, not inside the model.

None of it maps onto centimetres either - a physical measurement is a single, angle-independent read that a photo set cannot approximate or replace, and every technique in this guide is specific to what a camera and a rating tool can see, not to what a tape can confirm.

A human reviewer sits somewhat outside all of this. A judge working from a submitted set can usually say which angle or which image informed an impression, which is the one piece of information an automated combination rule most often withholds - and it is a reasonable reminder that the entire discipline in this guide exists to recover, mechanically, some of what a person can just tell you by looking.