How many tries should one good AI image take?

For a simple subject, expect a usable frame within a handful of generations. For a specific brief, ten to thirty is ordinary. What separates a long session from a wasted one is changing exactly one thing at a time and keeping the rejects.

People quietly assume they are bad at this because their first three images were wrong. They were not wrong because you are bad at it. Generation is sampling: the same prompt describes a distribution of pictures, and you are drawing from it. Professional workflows are batch-and-select workflows, and always have been — the same as shooting two hundred frames to publish one.

A realistic budget

BriefTypical generationsWhere the time goes
Loose subject, any interpretation3–8Mostly picking a favourite
Specific composition and mood10–30Prompt revisions between small batches
Named character, repeated across a set30+Locking identity, then holding it
Legible text in the image10–40Text is a re-roll problem more than a prompt problem

How to iterate so the attempts compound

  1. 1.Generate a small batch — three or four — from the same prompt before changing anything. One image tells you nothing about whether a prompt is good or you got unlucky.
  2. 2.Change one clause per round. Two changes and a better image teaches you nothing about which change did it.
  3. 3.Keep the rejects and note what you changed. Most people's second hour is spent rediscovering a wording they already tried in the first.
  4. 4.When three consecutive rounds produce the same failure, stop editing words. That failure is a model limitation, and the answer is a different engine, a reference image, or fixing it after the fact.

Knowing when to stop

Two stopping conditions are worth respecting. The first is having a good image — people keep going past this constantly, chasing a mental picture that was never better than what is already on screen, and pick the ninth-best result at 2am. The second is the plateau: when variance between attempts is small and the flaw is stable, more attempts are just paying for the same picture again.

Rejected generations are the cheapest documentation you will ever have. A folder of near-misses with their prompts attached is how a prompt library gets built — and it is how the prompts published here got settled.

More on this

Other questions

← all questions