Iterating & Refining Systematically

8 min read

Stop rerolling and start engineering: the one-variable rule, seed locking, and an iteration checklist that converges on your vision fast.

Your first generation is a draft, not a verdict. The difference between amateurs and pros is not that pros nail it on try one — it is that every pro attempt teaches them something, because they change things methodically. Random rerolling is a slot machine. Iteration is engineering.

The one-variable rule

Borrowed straight from the science lab: change exactly one thing per run. If you swap the lighting AND the style AND the camera angle, and the result improves — which change did it? You learned nothing. Change only the lighting, and whatever happens, you now know what that lighting phrase does in this scene. Ten disciplined runs build a private phrasebook; ten chaotic runs build frustration.

Seed locking: freeze the randomness

Every generation starts from random noise controlled by a seed number. Same prompt + same seed + same settings = (nearly) the same image. That makes the seed your control group: lock it, and any difference between two runs is caused by your prompt edit alone. Leave the seed random and you can never tell whether the prompt change helped or you just rolled better noise.

Iteration v1: the baseline

A cozy reading nook by a bay window, worn leather armchair, stack of hardcover books, rainy afternoon light, 35mm, muted tones, seed 4211

Model: nano-banana

Run it and judge: nook works, chair works, but let us say the light feels flat. That is your one variable for v2.

Iteration v2: one change, same seed

A cozy reading nook by a bay window, worn leather armchair, stack of hardcover books, golden hour sun striping the floor through half-open blinds, 35mm, muted tones, seed 4211

Model: nano-banana

Only the light phrase changed: rainy flat light became golden stripes through blinds. Same seed 4211, so the room layout barely moves — you are directing, not gambling.

What to iterate on, in what order

  1. Subject & action — is the right thing in the frame doing the right thing? Fix this first; nothing else matters until it is.
  2. Composition & framing — angle, crop, aspect ratio, focal point placement.
  3. Lighting — the biggest mood lever after the subject itself.
  4. Style & palette — stock, movement, render engine, color story.
  5. Micro-details & negatives — hands, text, clutter, artifacts. Polish comes last.
RunChanged variableVerdict
v1baseline prompt, seed 4211layout great, light flat
v2light: golden hour through blindsmood fixed, chair too dark
v3added: soft fill light on the armchairchair reads, keep
v4style: Kodak Portra 400warmth up, done — ship it

When is it done? When the next change makes it different instead of better. Most images converge in 4-7 disciplined iterations — dramatically fewer than the 30+ rerolls of slot-machine prompting. If you are stuck, the problem is usually upstream: revisit the five-part anatomy before iterating harder.

Related glossary terms: Seed, Iteration, CFG Scale, Prompt Engineering

The Prompt Builder keeps your prompt structured so single-variable edits are trivial — swap one dropdown, keep the rest. Iterate in Prompt Builder