Choosing the unit of comparison

Use a single photo if the question is whether a result changed since last time, and a set only if the question is whether it depends on angle. The two are not versions of the same submission - they are different units of comparison, and the choice should follow from the question, not from habit.

What a single photo has going for it

One slot to hold constant is fewer things that can drift. A single photo taken to the same angle, distance and light every time is the simplest protocol available, and simple protocols survive longer than complicated ones - which is the same reason a single well-kept slot beats a partially-maintained set over a long stretch.

A single photo also means the combination question never arises. There is no averaging rule, no dominant-image ambiguity, no uncertainty about which image a multi-photo tool's score actually reflects - the score is legibly about the one image submitted, and nothing else.

The cost is information. One angle, one moment, one draw. It tells you about that slot and nothing about how the subject reads from anywhere else. And image models are touchy about slots: Azulay and Weiss (2019) showed that "small translations or rescalings of the input image can drastically change" a convolutional network's prediction, so a single frame is a noisy sample even before the angle changes.

What a set has going for it

A set of two or three angles shows whether a result is angle-dependent or holds up regardless of viewpoint, which a single photo cannot answer by itself. Per-angle spread is where a set's information actually lives - if the angles agree, that is worth knowing; if they diverge, that is worth knowing more.

The cost is upkeep. Every additional slot in the set is another condition to reproduce identically each time, and a set with one poorly-held slot is not more informative than a single photo - it is worse, because a mixed-condition set produces a total that describes nothing reproducible.

A middle option

There is a third path worth naming: a single primary slot, maintained rigorously as your comparable record, with an occasional second angle added when you specifically want to check for angle-dependence rather than as a routine part of every submission. This gets most of the reliability of a single-slot protocol with an occasional look at the question a full set answers, without committing to the ongoing overhead of holding two or three slots constant indefinitely.

The trade-off with this middle option is that the occasional second angle is not itself a series - one extra photo taken once tells you whether angle-dependence exists at that moment, not whether it is stable, and it should be read as a one-off check rather than folded into the main record.

A rule for choosing

If the question is "did this change since last time," a single well-maintained slot answers it with the least room for the protocol itself to be the source of the movement. Fewer moving parts, easier to trust the result.

If the question is "does this read differently depending on angle, or is a single number from a single tool session actually representative," only a set can answer it, and the set has to be built with defined slots from the start to be worth the extra maintenance.

Most people do not need both running at once. A single comparable photo, repeated, is enough for most of what a series is for. A set earns its cost when the angle-dependence itself is the thing in question - and at that point building the set properly, with named slots documented before the first photo is taken, matters more than which unit was chosen. For anyone taking the set route seriously, the fuller walkthrough of building and maintaining a photo series goes further into slot naming, cadence and what to do when a slot lapses.

This is a comparability question, not a quality one - it has no bearing on which unit scores higher. What actually moves a score between a good photo and a standardised one is a separate axis Rate Cock's public entries make visible by showing per-axis results rather than a single figure, which is the same distinction driving the choice here: more data points versus fewer variables to control.

The underlying geometry is identical either way. A wide-angle lens renders proportion differently at different distances whether the submission is one photo or five, so the reproducibility problem a single slot solves has to be solved inside every slot of a set too, just multiplied by however many angles are in it.

None of this maps onto a physical read. A tape measurement is a single number regardless of how many photos accompany it, so the one-photo-versus-set question is specific to what a rating tool can see in an image, not to anything a measurement owns.

A human reviewer will usually ask for more than one photo regardless of which unit a rating tool prefers - what a judge wants in a submission is closer to the set case, informed by more than one angle, which is a different trade-off from the one a rating tool's comparability protocol is making.