How to edit an image with Nano Banana Pro without re-rendering the whole frame

Nano Banana Pro is Google's gemini-3-pro-image, and it is the image family LUVI reaches for by default. On LUVI it runs as text-to-image and edit, with an edit-ultra tier above them, at 1k, 2k or 4k, from 175 credits per image at 1k and 300 credits at 4k. The edit mode takes up to ten reference images. It is also the mode where most credits get wasted, for one structural reason.

Why did my edit change things I didn't ask about?

Because an edit on this family re-renders the whole frame. There is no masked region and no pixel lock: the model reads your instruction, then generates a new image that should look like the old one plus your change. Anything you did not pin down is free to drift—a face, a shadow, the text on a sign.

The fix is a preservation clause, and Google's own documentation uses one in its multi-turn editing example: "Update this infographic to be in Spanish. Do not change any other elements of the image."

LUVI enforces the same rule deterministically. The prompt check looks for an edit verb—change, replace, swap, remove, delete, erase, recolor, relight—and, if the prompt carries no preservation wording, appends one:

Replace the plain white mug in Image 1 with a matte black ceramic mug, same size and position. Keep everything else in the image exactly the same.

That closing sentence is not decoration. Without it the same instruction is an invitation to redraw the counter, the light and the hand holding the mug.

How do you point at the right reference image?

By ordinal, written as Image 1, Image 2, and so on—capitalized, with a space, and no @ sign. Nano Banana binds references as ordinals rather than tokens, so image 1 in lowercase is rewritten to the canonical form before the prompt is sent.

Up to ten images go into one edit. Google's documentation describes what those slots are good for on this generation: "Up to 10 images of objects with high-fidelity to include in the final image", plus a smaller number for character consistency and for style reference. In practice that means an edit can be a composition job—put the object from one photo into the scene of another—as long as each image is named in the sentence that uses it.

Why doesn't "no cars" work?

There is no Negative Prompt field on Nano Banana, and the family's guide is blunt about what happens if you improvise one: "describe positively ('an empty, deserted street', not 'no cars'). Naming what you don't want can summon it into the frame."

This is the habit most people bring from Stable Diffusion, and it is the one that costs the most here. Every exclusion has to be rewritten as the presence of its opposite.

• "no cars" becomes an empty, deserted street
• "no text" becomes a clean, unlettered surface
• "not blurry" becomes sharp throughout, crisp edge detail
• "no clutter" becomes three objects on a bare counter, generous empty space

The prompt check will not do this rewrite for you. It fixes what it can prove—the reference form, the preservation clause, weighting syntax that does nothing here—and leaves the judgment calls in your hands.

Which settings actually change the output?

Five parameters, and the two most consequential are easy to get wrong.

• resolution, with the values 1k, 2k and 4k in lowercase. The enum is case sensitive, so 2K is not the same string as 2k, and a value that does not match is dropped—leaving you with the 1k default and a bill you did not expect to be smaller.
• aspect_ratio, with ten values: 1:1, 3:2, 2:3, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9 and 21:9. These match Google's published list exactly. Writing the ratio into the prompt text instead does nothing.
• output_format, png or jpeg.
• media_resolution, which controls how your input images are read rather than how the output is written. Lowering it spends fewer tokens per input image and can lose detail from the reference.
• enable_web_search, off by default, which lets the model ground the generation in real-time information.

There is no seed on this family, so two identical submissions are two different pictures. If a result is close but not right, the way forward is another edit pass on the output, not a re-roll of the same prompt.

How we checked

• Models: google/nano-banana-pro/text-to-image, google/nano-banana-pro/edit and google/nano-banana-pro/edit-ultra.
• Settings: 1k, 2k and 4k; the ten aspect ratios in the schema; edit mode with reference images.
• Date and what was read: September 22, 2026. Model schemas and credit estimates from LUVI, the prompting guide LUVI runs for this family, and Google's image generation documentation linked below.
• Results: 175 credits for one image at 1k and 300 credits at 4k. The prompt check returned two rewrites on a short edit instruction: image 1 to Image 1, and the appended preservation sentence. The ten aspect ratios in LUVI's schema match Google's published list.
• Drawbacks: No outputs were generated for this post, so it makes no claim about output quality. Credit estimates move when a provider reprices a model—read the estimate in the Workspace before a 4k batch.

In LUVI

Open the Workspace, choose Nano Banana Pro Edit, drop your image into the reference slot and write the instruction. Before you press generate, press the feather button beside the prompt: that is LuviTransLex, the free check that applies this family's prompting guide.

On an edit prompt it earns its place immediately. Replace the plain white mug in image 1 with a matte black ceramic mug, same size and position. comes back with two rewrites: the reference is normalized to Image 1, and the preservation sentence is added to the end. Neither costs a credit.

You will notice the Negative Prompt field is missing from the panel. That is a property of the model, not a setting you have turned off—when a model has one and when it doesn't is decided by the model. The same edit can be driven from Claude or ChatGPT through the LUVI connector, which is useful when the source image is already in your Library.

New here? Create a LUVI account and run your first edit at 1k before you commit to 4k.

Sources

• Image generation with Gemini — Google AI for Developers, API documentation. Accessed September 22, 2026.