How to prompt Nano Banana 2.1: prose in, settings out, and no transparent background

Nano Banana 2.1 is Google's gemini-nano-banana-2.1, released on October 6, 2026 as the successor to Nano Banana 2. On LUVI it runs in three modes: text-to-image from 80 credits, edit from 83 credits with one input image, and reference-to-image, which turns a video into a still, from 400 credits, all at 1k. LUVI wrote a dedicated prompting guide for 2.1 in October, and the rules below are quoted from it.

What does Nano Banana 2.1 want in a prompt?

Nano Banana 2.1 wants narrative prose, written in a fixed order. LUVI's guide puts it as "register → subject and action → setting → light → camera → in-image text → finish", and adds that the model reads a brief "positively, in prose. There is no negative_prompt, no weighting and no style dial."

Two habits make the biggest difference:

• Open with the operation and the deliverable. "Create an image: a photorealistic product photograph of…" tells the model what it is making before it reads the details.
• Expand only what you wrote. The guide tells LUVI's own rewrite not to add light directions, lenses, props or wardrobe you didn't ask for. Do the same by hand: every extra noun is something the model will try to draw.

Create an image: a photorealistic product photograph for an online shop of a matte sage-green ceramic table lamp with a pleated off-white linen shade. It stands on a pale oak sideboard against a plain light-grey wall. Soft window light from the left, eye-level camera, 50mm lens, shallow depth of field. A small card leaning against the base reads "Sage, 38 cm" in thin black sans-serif capitals.

Why does "no cars" put cars in the picture?

Nano Banana 2.1 has no negative prompt, so an exclusion is just more words describing the thing you don't want. The guide is strict about it: "no people, no cars" becomes "a quiet, deserted street", "never 'no cars' and never 'empty of people and cars'." Even the second version names the cars.

LUVI's prompt check flags a bare "no X" before you generate, but it can't write the replacement for you. Describe the end state instead: a clean counter, an empty beach at dawn, sharp focus throughout.

Where do the size and shape go?

The size and shape of a Nano Banana 2.1 image go in the controls, never in the prose. The guide says it plainly: "Never write '16:9', 'widescreen', '4K' or '8k' into the prose — on a conflict the parameter wins and the words are wasted."

• resolution: 1k, 2k or 4k, in lower case. Text-to-image quoted 80, 120 and 224 credits on October 20, 2026.
• aspect_ratio: fourteen shapes plus auto, including the panoramic 4:1, 1:4, 8:1 and 1:8. On text-to-image, auto gives you a square, so set the shape you want. On an edit, auto keeps the shape of your input image.
• thinking_level: minimal, medium or high, where the default means medium. The setting doesn't change the price. LUVI's guide suggests high for infographics, dense lettering and three or more references, and minimal for quick drafts.

If you generate through Claude or ChatGPT with the LUVI connector, a ratio or resolution written in the prompt is moved into the matching setting when you haven't set it yourself, and the price follows. In our check, a prompt ending in "16:9, 4K" turned an 80-credit quote into 224.

How do you get exact text into the image?

Exact text goes inside double quotes, written exactly as it must appear, with the type and placement described in plain words around it. Three details from LUVI's guide save a reroll:

• Never write "in quotes". The model draws the quotation marks.
• Quote every visible surface or call it blank. Signs, packaging and screens without quoted copy come back with invented lettering.
• An em dash inside the quote is drawn as a separate line. Keep the dash and add, outside the quotes, that the text stays on a single line with a literal em dash between the phrases.

Non-English copy, such as Turkish with İ, ı, ğ and ş, should be large in the frame. The guide warns that small lettering "drops diacritics at any resolution", so raising resolution doesn't fix it.

Can Nano Banana 2.1 make a transparent background?

No. Nano Banana 2.1 has no alpha channel, and LUVI's guide says: "Never ask for a transparent background — the model has no alpha. Ask for a plain seamless white background instead." The 2.1 models on LUVI also have no output-format setting: in the 2.1 results generated on LUVI between October 8 and 16, 2026 we found both JPEG and PNG files, and the PNGs we opened were plain RGB with no transparency.

When you need a cut-out, the guide's workaround takes two generations at the same resolution and shape. First the subject on white:

Create an image: a studio product photograph of a frosted glass perfume bottle with a brass cap, centered on a pure solid white background, shadowless, evenly lit from both sides.

Then an edit of that result:

Change the background of Image 1 from white to pure solid black. Keep the bottle exactly as it is: same position, size, shape, colors, highlights and edges. Where the clear glass is see-through, the new black background shows through it as real glass would. Keep everything else in Image 1 exactly the same.

With the subject on white and on black, an image editor or a short script can calculate the transparency from the difference between the two. At 1k the pair costs 80 plus 83 credits.

What changes when you edit with 2.1?

An edit on Nano Banana 2.1 re-renders the whole frame, so you state the change first and then everything that must stay. The 2.1 edit mode takes up to 14 input images, and the guide asks you to name each one with its ordinal, a noun and its role: "the bottle from Image 2", "the kitchen in Image 1".

What's specific to 2.1:

• The keep clause goes on every edit, "also for compositing edits (place / put / add / insert), not only for change/replace."
• A mark drawn on the input selects a region. "Replace the object circled in red with…" is a valid instruction.
• Name the tonality of a black-and-white input. The guide's wording is "keep the image black-and-white with the same grey tonality"; if you want color instead, say so.
• At most four people whose identity comes from your photos. Google documents "character resemblance of up to 4 characters" for 2.1.

Place the brass desk lamp from Image 2 on the left side of the oak desk in Image 1, lit by the same window light as the desk. Keep everything else in Image 1 exactly the same.

The guide to editing with Nano Banana Pro covers the preservation clause in more depth. The habit is the same on 2.1.

How does the video-to-still mode work?

Nano Banana 2.1 reference-to-image takes exactly one video, plus up to 14 optional images, and returns a single image built from what happens in the clip. Google describes the feature as generating "new images using a video's context as a multimodal reference", for thumbnails, posters and summary graphics. Refer to the clip as Video 1 and name the moment you want:

Create an image: a cover frame for this cooking video, built on the moment the lid opens in Video 1, with steam rising above the pot. Keep the kitchen, the pot and the cook's striped apron as they appear in Video 1.

This mode starts at 400 credits at 1k, five times a text-to-image. On October 20, 2026, clips of 5, 10 and 20 seconds all quoted the same 400 credits.

How we checked

• Models: Nano Banana 2.1 Text-to-Image, Nano Banana 2.1 Edit and Nano Banana 2.1 Reference-to-Image on LUVI.
• Settings: 1k, 2k and 4k; one and two input images on the edit mode; video clips of 5, 10 and 20 seconds on reference-to-image.
• Date and what was read: October 20, 2026. Model schemas and credit quotes through the LUVI connector, LUVI's prompting guide for Nano Banana 2.1, the file type of every completed 2.1 result in LUVI's database (October 8 to 16), and Google's image generation guide and API changelog.
• Results: 80 / 120 / 224 credits for text-to-image at 1k / 2k / 4k; 83 credits for an edit with one image and 86 with two; 400 credits on reference-to-image for every clip length we tried. Every example prompt above passed LUVI's prompt check. No outputs were generated for this post, so it makes no claim about output quality.
• Drawbacks: we didn't run the two-render transparency workaround for this post, and the mix of JPEG and PNG output may change without notice.

In LUVI

Open the Workspace, pick a Nano Banana 2.1 mode and press the feather button beside the prompt. That opens LuviTransLex, the free check that applies the model's prompting guide: it adds the missing keep clause to an edit and flags a bare "no X". Set the ratio and resolution in the controls, not in the prompt.

New accounts start with 200 credits, and verifying a phone and a card adds up to 1,500 more. Create a LUVI account and try a 1k draft before you move to 4k.

Sources

• Image generation with Gemini (Nano Banana) — Google AI for Developers, API documentation. Accessed October 20, 2026.
• Gemini API release notes — Google AI for Developers, entry of October 6, 2026. Accessed October 20, 2026.