How many variants is enough, and where more generations start hiding the good one

When each attempt costs almost nothing, the rational-seeming move is to make many and pick the best. In practice past a fairly low number this stops helping and starts hurting, for a reason that has nothing to do with the model.

Choosing between four images is easy. Choosing between forty is genuinely hard — the differences between adjacent candidates become smaller than your ability to judge them, and you end up picking based on order, fatigue, or whichever you saw last.

Worse, a large set trains you to compare candidates against each other rather than against the brief. The best of forty mediocre attempts feels good because everything around it is worse.

Generate four. Look at them against the brief, not against each other. One of three things is true:

• One is clearly right. Take it and move on. This happens more often than people expect and is routinely ignored in favour of 'let me just try a few more'.
• None is right, but one is closer. The prompt needs changing. Change it, generate four more. This is the productive loop.
• All four are wrong in the same way. The prompt is wrong at a structural level. Rewrite a block rather than re-rolling.

The case that does not appear on this list is 'they are all fine and I cannot decide'. That means the brief is not specific enough to choose against, and more generations will not resolve it.

Before generating, write one sentence: what has to be true for this frame to work. 'The impossible element is readable at thumbnail size.' 'The product's label is unchanged.' 'The lower third is empty enough for type.'

Now judge against that sentence. Frames either meet it or do not, and the choice takes seconds instead of minutes.

This also stops the most expensive failure in generative work: producing forty beautiful images, choosing one, and discovering at layout stage that none of them had space for the headline.

There is one situation where forty is the right number: when you need a set rather than a single frame, and you are looking for the handful that happen to be consistent with each other.

Producing thirteen frames for a mosaic might well mean generating sixty and selecting for coherence. That is a different activity — sampling a distribution rather than iterating towards a target — and the economics are genuinely different.

The tell is whether you are asking 'which of these is best' (small batches, iterate) or 'which of these belong together' (large batch, select). Confusing the two is how people end up doing the expensive version of the easy problem.

Reviewing generated images is a specific kind of tiring. Everything is similar, everything is competent, and the differences are subtle. Twenty minutes in, you are no longer evaluating — you are pattern-matching against what you have already seen.

Two habits help. Review in small batches with breaks between them. And view candidates at the size they will actually be used: a frame destined for a small card should be judged small, where half the differences you were agonising over are invisible.

If you find yourself unable to choose, stop. Come back in an hour and the right one is usually obvious immediately.

The stopping rule is simply: the frame meets the sentence you wrote before you started. When it does, you are finished, even if you suspect a better one exists.

A better one always exists. That is true at four generations and at four hundred, and the distance between them is not worth what it costs.

The discipline of writing the criterion first is what makes the stopping rule usable. Without it there is no defined finish, and generative work expands to fill whatever time is available.