Nano Banana 2 Prompts: How to Write Them for Better Image Results

Nano Banana 2 prompts work best when you treat them as a short creative brief rather than a list of keywords. A reliable prompt names the subject, the action, the setting, the style, the lighting, and the framing — then stops. If you are editing an existing image instead of generating a new one, the same logic applies but the structure changes: state what to change and what to keep. The guidance below covers both cases, plus how to use reference images for character consistency and how to fix prompts that produce the wrong result.

The six-part prompt structure

Most failed prompts are missing one of these elements, or pile on so many that the model has to guess which one matters.

A complete generation prompt built from the table:

A ceramicist in her 60s shaping a bowl on a wheel, in a sunlit studio with clay-dusted shelves, editorial photography with muted earth tones, soft window light from the left, medium shot slightly above eye level.

That is roughly 40 words. It is specific enough to constrain the output and short enough that no element competes with another.

How long should a prompt be?

Longer is not better. Every added clause is another thing the model can weight differently than you intended. Start with the six elements above. Add detail only when a generation comes back wrong and you can name the specific thing that was wrong — then add one clause to fix it. If you add five clauses at once and the result improves, you will not know which one did the work.

Writing edit prompts: change this, keep that

Editing prompts fail differently from generation prompts. The common mistake is describing the desired final image instead of describing the change. The model then reinterprets the whole scene rather than modifying one part of it.

Write edit prompts as two explicit lists:

• Change: the specific element to modify — "replace the grey background with a warm terracotta wall"
• Keep: the elements that must survive untouched — "keep the subject's face, pose, clothing, and lighting unchanged"

For example, to swap a background without disturbing the subject:

Change the background to a warm terracotta wall. Keep the subject's face, hair, pose, clothing, and the existing light direction exactly as they are.

Naming what to keep is not optional. Without it, models frequently "improve" adjacent details — a face gets subtly re-rendered, a shadow shifts, a fabric texture changes. The keep-list is your protection against that drift.

One change per pass

If you need three edits, run three passes. Combining "change the background, brighten the image, and add a hat" into one instruction makes it hard to tell which instruction caused an unwanted side effect. Sequential single edits also let you stop as soon as the result is right, instead of regenerating everything.

Using reference images for character consistency

Character consistency across multiple images depends on giving the model a stable visual anchor, not on describing the character in more words. A reference image carries far more identity information than any paragraph of description.

The workflow that holds up:

1. Generate or select a clean reference. A clear, well-lit image of the character with an unobstructed face works better than a dramatic or partially cropped one.
2. Attach the reference to every subsequent generation. Do not rely on a text description of the character to carry identity forward.
3. Name the character in the prompt. Give them a consistent label — "Mira" — and use the same label every time. Named elements are easier to reference and easier to keep stable than "the woman in the red coat."
4. Describe only what changes. The reference handles identity; the prompt handles the new scene, pose, or action.

A consistency prompt looks like this:

Mira, standing in a rain-soaked alley at night, neon signage reflecting on wet pavement, cinematic photography, cool blue and magenta lighting, full-body shot.

The reference image supplies Mira's face and build. The prompt supplies everything else. If you instead re-describe Mira's features in text each time, small wording changes will produce small identity changes.

When consistency still drifts

If the character's identity shifts between generations, check these in order:

• Is the reference image attached to this generation, or did it drop out?
• Is the reference itself consistent — same person, same lighting, same angle as your other references?
• Did you add a clause that conflicts with the reference, such as a different hair color or age?
• Is the new scene so different in lighting or angle that the model is reconstructing rather than transferring the face?

Fix one of these at a time and regenerate. Changing several variables at once makes the cause unknowable.

Troubleshooting common prompt failures

The general rule for all of these: isolate one variable, change it, regenerate, compare. Rewriting the entire prompt after a bad result throws away the information you just gained about what the model did with your wording.

Iterating without wasting generations

Treat prompting as a loop, not a one-shot:

1. Write the six-element prompt.
2. Generate.
3. Compare the result against your intent and name the single biggest gap.
4. Change only the element responsible for that gap.
5. Regenerate and compare again.

Two or three passes usually converge. If you are several passes in and the result is not improving, the problem is usually structural — a missing reference image, a contradictory clause, or a prompt so long that no single element dominates. In that case, cut back to the six elements and start the loop again rather than adding more description.

Note that generation consumes credits on this platform, with starter credits on sign-up and paid usage afterward, so converging in fewer passes has a practical benefit beyond speed.