How to prompt NanoBanana

NanoBanana is Google’s Gemini image line, and the bot offers it as NanoBanana 2 for everyday work and NanoBanana Pro for the renders that have to be right the first time. Both take the same kind of prompt, so learning one teaches you both and the only real decision is which tier you spend credits on. What sets the family apart is that it reads a prompt for material, surface and mood instead of working through it like a checklist.

The NanoBanana prompt formula

Six slots cover almost every prompt: subject, action, setting, light and lens, framing, style. Here is one prompt carrying all six:

A woman in her thirties in a grey wool coat sitting at a cafe window on a rainy afternoon, one hand wrapped around a ceramic mug with steam rising off it, the other forearm resting on the wooden counter, soft daylight from the left falling through rain streaked glass, 85mm close portrait with the background dropping out of focus, natural documentary photography

Note that it reads as a description, not an inventory. Tags stacked with commas (woman, cafe, rain, moody, 85mm) get averaged into a mood; connected clauses keep the relationships, light falling through the glass, steam rising off the mug. Say it out loud, then check all six slots are there.

The failure mode that follows does not look like failure. GPT-Image-2 misses literally, so the gap is visible and you know which slot to add. NanoBanana lands in the middle: a thin prompt still returns a competent, plausible image, the average of everything your words could have meant, with nothing pointing at what you never wrote.

Texture and material is the fastest way out of that middle. Ribbed knit, glazed ceramic, weathered brass, condensation on glass: each names a surface the model can resolve, and each narrows the average. Adjectives about quality (“beautiful”, “masterpiece”, “8k”) narrow nothing, since every render was already trying to be those.

The two pictures further down the page show the difference. The first came from three words, “a woman in a cafe”, and the model answered with a pleasant wide shot: a smiling stranger, a pastry, a whole cafe of people behind her. The second names her age, the coat, the mug, the rain on the glass, the light and the focal length, and the crowd, the pastry and the smile all disappear, because none were asked for.

Attaching photos: @image1 and @image2

Up to 4 photos go in one message with the instruction as its caption, and you then address them in the prompt as @image1, @image2 and so on; the step-by-step guide covers the mechanics with screenshots.

Keep the face and hair from @image1 exactly as they are, dress her in the grey wool coat from @image2, place her at a rainy cafe window, soft light from the left, keep the background of @image2 unchanged

What matters here is that NanoBanana does not paste a reference face, it re-renders it, so a likeness survives only as far as your prompt protects it. “Keep the face from @image1 unchanged, same bone structure, same nose, do not restyle the skin” holds up across a change of clothes or light where “keep the face” does not. Ask for new light and it relights the face too, and a relit face is a slightly different person. A dim or heavily filtered source also gives it less to hold onto.

Blending is where the family is genuinely strong. A face from one attachment, wardrobe from another, a location from a third comes back as one scene rather than a collage, without the cut-out edges such edits usually leave. Give every attachment a job: a reference with no stated role is absorbed into the mood instead of used. The more you attach, the more the tier matters.

NanoBanana 2 or NanoBanana Pro

The whole nano banana 2 vs pro question comes down to this: both are good, and only one of them is cheap.

Pro’s cheapest render costs the same as a 4K image on 2, and that pricing is the whole workflow. Iterate on NanoBanana 2, render the keeper on Pro. Run the prompt at 1K for 2 credits, change one slot, run it again, and spend Pro credits only once the wording works.

The difference you are paying for shows up on complexity. One subject and clean light looks near identical on both tiers, so reaching for Pro there wastes credits. Load the prompt up with several people, a busy background, specific wardrobe and a particular light, and NanoBanana 2 starts quietly dropping details: a described garment turns generic, a face at the back of the frame smears. Pro holds the list together and keeps its edges clean at 4K, which is why it is also the one for hero images and close-range skin. The same gap shows on attachments: with three or four references in play, 2 serves one well and lets the rest blur, while Pro keeps each doing its own job.

Text inside images is the weak spot

Every other section here is about getting more out of the model. This one is about knowing its ceiling, which is the most useful gemini image model prompt tip on the page. NanoBanana 2 manages a short label often enough to be tempting and misspells it often enough to be frustrating, and it is the least reliable model here once a string runs past a word or two. NanoBanana Pro is markedly better, and on a short headline usually fine, but “usually” is the operative word.

To give nano banana text in image its best chance: put the exact wording in quotation marks, keep it short, name the surface it is printed on, and re-run rather than rewrite when a letter comes out wrong, since lettering varies between runs of the same prompt.

When the wording has to be exactly right, switch models. GPT-Image-2 is the strongest for text and the cheapest to iterate on, and the GPT-Image-2 guide covers how to get a string to land the first time.

What the filter refuses

Expect roughly GPT-Image-2 levels of strictness, over the same four categories: real named people, brand marks and logos, violence, and explicit content. The way out is identical, so rather than repeat it here, the GPT-Image-2 guide works the rewrite pattern through with examples.

What is particular to this family is the refusal that arrives while editing rather than generating. Once you attach a photo of a real, identifiable person, the limit stops being your wording and becomes the face in the frame: no phrasing unlocks an edit of a public figure, and none unlocks a picture placing someone in a situation they were never in. Re-running it or softening the sentence changes nothing, which is worth knowing before you spend credits finding out.

The workflow the reference tools exist for stays wide open: your own face, friends who agreed to it, restyling, relighting, a change of wardrobe or setting. Expect the strictest reading on children, on anything intimate, and on edits that would read as a claim about something that really happened. Seedream is looser on style, not on subject, so the Seedream guide is worth reading for stylized and ultrawide work, not as a detour around a refusal you earned.

Common fixes

Ready-made prompts

Reading about materials teaches you less than watching a render change when glazed ceramic becomes hammered copper. The NanoBanana prompt library has ready-to-paste prompts next to the images they produced, and the model pages for NanoBanana 2 and NanoBanana Pro list the resolutions, aspect ratios and credit costs.

Open the bot, pick a NanoBanana tier, and paste one in.