Iterate and edit

Intermediate · 6 min · Choose between editing and regenerating, one change at a time.

  • ai images
  • iteration

An edit changes one thing and keeps everything else. Regenerating re-rolls everything. Most disappointment in image work comes from re-rolling when you should have been editing — the composition was right, and the whole thing was thrown away over a jacket colour.

Edit or regenerate?

The test is one question: if everything except this one thing stayed identical, would I be happy?

  • Yes → edit. The foundation is right; only one variable needs moving.
  • No → regenerate. The subject, composition or concept itself is wrong, and no amount of tweaking will rescue it.

Editing is how finished images get finished. Regenerating is how drafts get started.

The edit loop

How it works, step by step

  1. Generate a small batch — two to four candidates, not twenty. You are choosing a direction, not searching for perfection.
  2. Pick the closest candidate as your foundation. Closest, not perfect: the rest of the loop is what makes it right.
  3. Name exactly one change. “Warmer light.” “The jacket, deep green.” One variable per pass, written down.
  4. Apply it as an edit: a whole-image instruction that says what to keep and what to change — or, where your tool supports it, mark just the region and describe its replacement (a region edit like this is called inpainting).
  5. Compare the new version with the previous one and keep the better of the two. If the change spread further than you asked, revert and try a smaller step.

A bad example

Ask: “Same image but make it better.”

Nothing is named, so everything may change. There is no way to tell whether it improved — you only get a different image to react to.

A better example

Ask: “Keep everything exactly as it is; change only the light to soft and warm, and keep the shadows where they are.”

One named variable, everything else pinned. The result is comparable, and the loop can continue.

Why it works

Generation decides every variable at once, so a re-roll resets your progress along with it. An edit holds the variables you do not mention steady while it moves the one you named — which is why edit modes exist in every serious image tool. The side benefit is a trail: accepted versions become the base for the next pass, and “go back one step” becomes possible. Change one thing per pass and you always know what caused what.

An image is 90% right — the composition, subject and styling all work — but the light reads too cold. What is the disciplined next step?

Practice

Order the edit loop

You have four versions of a product shot. One composition is right; the background colour is slightly off and the label area needs a small replacement. Instead of another round of re-rolls, you decide to work by editing — one disciplined loop from foundation to finished image.

Order the steps of the edit loop so each pass changes one thing and keeps everything else intact.

  1. Apply it as an edit: describe the change for the whole image, or mark the region and describe its replacement.
  2. Compare the new version with the previous one and keep the better of the two.
  3. Only if the change spread too far: revert, then take a smaller step from the better version.
  4. Generate a small batch — two to four candidates — and choose a direction.
  5. Pick the closest candidate as your foundation — not the perfect one.
  6. Name exactly one change you want to make.
Hint

A foundation comes before fine-tuning — pick the closest candidate first.

One variable per pass; anything else is a re-roll wearing a disguise.

Transfer

  • Any tool with editing: prefer an edit over a re-roll when most of the image is right.
  • Keep versions numbered or saved — the previous version is your undo.
  • When an edit overshoots, revert and take a smaller step rather than piling on corrections.
  • New projects: use the loop from the start and iteration becomes routine, not a rescue.

Key takeaways

  • Edit the close one; regenerate the wrong one. Ask: would everything else staying identical make me happy?
  • One named change per pass — that is what makes results comparable.
  • The version trail is part of the method: keep the better of each pair and move on.

Next

Next: Consistency across images — keeping a character or brand stable across a whole series. The inpainting entry covers region edits, and the diffusion entry explains why re-rolls differ at all.

Finished this lesson?