Checking your session…

Sign in to train

Signing in stores your name, email, and lesson progress. Nothing else.

Back to Creator Reps

Can’t verify your session right now. Check your connection and retry.

Back to Creator Reps
0
Track Progress
0day streak
Prev Next

Lesson 2 of 13 · Module 1: Image Generation Craft

The Edit Loop

Correct one specific, named flaw in a first-pass AI image using targeted edit instructions rather than regenerating the whole image from scratch.

Pick: “Edit ANY AI Image With This JSON Trick | Gemini Nano Banana 2

Teacher’s Tech · 9:26

Strong match. Explicitly shows surgical, targeted edits (swap object, fix text, clone style) that keep everything else in the image intact, on the free Gemini/Nano Banana model. This is the lesson’s core skill on screen.

Objective

BehaviorCorrect one specific, named flaw in a first-pass AI image using targeted edit instructions rather than regenerating the whole image from scratch.
ConditionStarting from an image you generated in Lesson 1 (or a new one), using your tool’s edit, inpaint, or “change X, keep Y” prompting capability, in 3 or fewer edit passes.
CriterionThe final image differs from the first pass only in the specifically targeted attribute; everything not targeted (composition, subject identity, lighting, style) stayed recognizably the same, confirmed by placing both images side by side.

Why This Matters

The default beginner move when an image is 80% right is to regenerate the whole prompt and hope the 20% fixes itself. It usually doesn’t, and now you’ve also lost the 80% that was already working. This is prompt roulette, and it’s slow and expensive in attempts. The actual skill is diagnosis: name the specific thing that’s wrong, then fix only that thing. This is the same discipline as debugging anything else you already do daily. You don’t rewrite the whole system because one function is broken.

The Technique

Before touching the prompt, name the flaw in one specific sentence. Not “something’s off,” but “the left hand has six fingers” or “the color grade is too warm” or “the subject is too centered.” If you can’t name it specifically, you’re not ready to edit yet, look longer.

Once named, use targeted edit language: “keep everything the same except [specific attribute],” or use the tool’s region-based edit/inpaint feature if it has one, pointing at the specific area rather than re-describing the whole scene. The instruction should be surgical, not a rewritten paragraph.

Two things to know about how these models handle instructions. First, full re-prompting and true region edits are different operations depending on your tool, know which one you’re using, a full re-prompt with “keep the same” language is not as reliable as an actual inpaint/edit mode when one is available. Second, models are unreliable at negative instructions. “Don’t make the sky orange” often produces an orange sky anyway, because the model latches onto “orange sky” as the salient token. State what you want instead of what you don’t want: “keep the sky pale blue” beats “don’t make it orange.”

Watch For This

Good

  • Side by side, only the targeted attribute changed.
  • Composition, identity, lighting, and style all held steady across the edit.
  • You needed 3 or fewer passes because each pass was aimed at a specific, named problem.

Classic Failure

  • You “edited” by rewriting the whole prompt, and the whole image changed, including the parts that were already right.
  • You used a negative instruction (“no text in the sign”) and the model produced the exact thing you told it not to.
  • You couldn’t name the flaw precisely, so the edit instruction was vague and the model changed the wrong thing.

Your Drill

Take your Lesson 1 image (or generate a new one), identify one specific flaw in writing before you touch the tool, then correct it using a targeted edit in 3 or fewer passes. Submit the before image, the after image, and the one-sentence diagnosis you wrote first.

Done? Paste what you made into the AI coach below for notes against this lesson's pass checklist.

Pass Checklist

Lesson complete

Criterion met: The final image differs from the first pass only in the specifically targeted attribute; everything not targeted (composition, subject identity, lighting, style) stayed recognizably the same, confirmed by placing both images side by side.

Next: Lesson 3: Character and Style Consistency Across a Set

How solid did that feel?

Noted.

0 / 4

Coach Note

If you catch yourself about to hit regenerate on the full prompt because “it’s just easier,” stop and write the diagnosis sentence first. That ten seconds of naming the problem is the entire skill this lesson teaches. Skip it and you’re back to rolling dice.

AI Coach

Conversations clear when you leave the page.

Ask about this lesson, or paste what your drill produced above and get it checked against the list.

The coach comes online shortly.

Resurfaces In

Lesson 3 (Character and Style Consistency), Lesson 5 (Image-to-Video), Lesson 12 (Pipeline).