Prompting Gemini for Image Generation (Nano Banana Pro)

TL;DR: "Gemini image generation" is not one model. It is four Nano Banana models (Lite, 2, Pro, and a legacy version) reachable through four different surfaces (the Gemini app, Google AI Studio, the Gemini API, and Vertex AI), and each surface changes what you can actually get. Pick the model by the job, pick the surface by who's building, then write the prompt as a specific brief rather than a tag list.

Every fact below is checked against Google's own documentation on September 3, 2026, with the source linked. Where two of Google's own pages disagree, both are quoted rather than picked between.

What Does "Gemini Image Generation" Actually Cover?

Four models, not one, all under a nickname Google itself uses in its documentation: Nano Banana. The umbrella covers Nano Banana 2 Lite (gemini-3.1-flash-lite-image), Nano Banana 2 (gemini-3.1-flash-image), Nano Banana Pro (gemini-3-pro-image), and the legacy Nano Banana (gemini-2.5-flash-image). "Gemini" is the product family; these four are what actually draws the pixels, whichever door you walk in through.

That distinction matters because "Gemini" also names a chat app, a family of text and reasoning models, a live voice mode, and a developer API that does none of those things by default. Type a prompt into any of them and image generation still routes to one of the four models above. Which one you land on, and what it can do, depends on the surface and the settings, which is what the rest of this post maps out.

If you're arriving with prompts written for Imagen, note that Google's own deprecations table lists the Imagen generate models (imagen-4.0-generate-001 and its Ultra and Fast siblings) as shut down on August 17, 2026, naming gemini-3.1-flash-image as the recommended replacement. That is not a future warning; it has already happened. Our Imagen migration guide covers the field-by-field rewrite if you still have a library built on the old syntax.

Which Gemini Image Model Should You Use?

Match the model to the job, not to whichever one sounds most premium. Google's own one-line descriptions for each model, read from its models page, are a better guide than the marketing names suggest:

All four are marked Stable on Google's models page as of this writing, so none of them is a fragile preview build you should hesitate to depend on.

Read past the marketing adjectives and the actual decision is simpler than four options suggest:

• Default to Nano Banana 2. It's the generalist, positioned by Google for "high-volume use cases", and it is what the Gemini app itself uses for a first generation. Most prompts belong here.
• Reach for Nano Banana Pro when the image has to carry text, a complex layout, or a reasoning step. Google's Vertex AI model card for this model states it is designed "to tackle the most challenging image generation by incorporating state-of-the-art reasoning capabilities", and calls it "the best model for complex and multi-turn image generation and editing". That reasoning pass and the parameter-level detail behind it, reference images, Search grounding, thinking mode, are covered in full in our dedicated Nano Banana Pro guide; this post stays at the model-choice level rather than repeating that walkthrough.
• Drop to Nano Banana 2 Lite for volume and speed, not quality. It's the cost floor, and it is capped at 1K resolution where the other three current-generation models reach higher.
• Only stay on legacy Nano Banana if you have a specific reason to. It has no documented resolution control at all, unlike the three Gemini 3-generation models that followed it.

If you'd rather start from a filled-in template than a bare model choice, our free Nano Banana Pro prompt generator has eight ready-made presets plus the full per-aspect-ratio resolution table for Pro, which we're deliberately not rebuilding here.

Where Can You Actually Generate These Images?

This is the part that gets skipped in most coverage, and it's the actual gap behind a search for "Gemini image generation" rather than a specific model name: the same four models behave differently depending on where you call them.

Three things worth knowing about each row before you assume the surface you're used to behaves like the others.

The Gemini app doesn't let you start on Nano Banana Pro. Google's own Gemini Apps help documentation is specific about this: "To use a different Nano Banana model, switch your Gemini model before you create your image request", and separately, that Nano Banana Pro "is available when you have the Gemini model set to Pro and want to redo an image when additional detail is needed." In practice, you generate with Nano Banana 2 first, then hit a "Redo with Pro" action, and only paid subscribers see that option at all. If you exhaust your daily Nano Banana 2 quota, the same help page states plainly that you can't redo any additional images with Nano Banana Pro either, because the redo draws from the same daily allowance.

Neither the API nor Vertex has a free tier for image output. Google's own pricing page reads "Not available" in the Free Tier column for input price, output price, and Search grounding, on all three current-generation image models. What it does offer instead, on the same page, for every one of those models, is a "Try it in Google AI Studio" link. If you want to test a prompt without paying per token, that's the door, not a hidden free quota on the API itself.

Vertex AI and the direct Gemini API don't even agree on Nano Banana Pro's aspect ratios. This is the kind of thing that only shows up when you compare two of Google's own pages side by side. The Gemini API's image generation guide publishes ten supported aspect ratios for gemini-3-pro-image: 1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, and 21:9. The Vertex AI model card for the identical model ID lists fifteen, adding 1:4, 4:1, 1:8, 8:1, and 9:21. Neither page cross-references the other. We're flagging that disagreement rather than resolving it: verify the ratio you need against whichever surface you're actually calling, not the other one's table. (If your reason for caring about resolution specifically is print or large-format output, the pixel math and the DPI conversions live in our 4K image workflows post, which goes deeper on that one axis than this post does.)

How Should You Structure a Gemini Image Prompt?

Write a brief, not a tag list, regardless of which of the four models or four surfaces you're using. That single habit carries more weight than any individual parameter.

Google's own best-practice guidance for these models is short, and every point in it pushes toward specificity over keyword-stacking. In outline, Google recommends: describing a subject with concrete physical detail rather than a vague category; stating the image's purpose, since a stated intent narrows the output; iterating conversationally rather than rewriting from scratch each time; breaking a complex scene into an explicit build order, background first, then midground, then focal subject; describing the intended scene positively instead of listing what to exclude, since none of these models take a negative-prompt parameter; and using photographic language, wide-angle, macro, low-angle, to control framing instead of a numeric field.

That collapses into a brief with five slots that works whether you're calling the app, AI Studio, the API, or Vertex, because the model reads the same kind of sentence regardless of the door you came in through:

A [shot type] of [subject, with specific physical detail] in [setting].
[Lighting description]. Shot from [camera angle] with [lens type]. [Any
text that must appear, quoted exactly, with its font style described].

Fill every slot. An empty one is where the model improvises, and its improvisations are competent and generic, not wrong exactly, just not what you asked for.

Two things this post is deliberately not re-covering, because they belong to a deeper post that already does the job well: reference-image mechanics (how many images, what each type is for, how style references differ from object and character references) and the "thinking" pass Nano Banana Pro runs before it renders. Both are worked through in detail, with Google's own worked examples, in our full Nano Banana Pro prompting guide. If your prompt needs to render text and it's coming out garbled, our garbled-text troubleshooting post covers that failure mode across models, and if the output shape doesn't match what you asked for at all, that's usually a settings mismatch rather than a model failure, which our wrong-size troubleshooting post walks through.

What Trips People Up When Moving Between Models or Surfaces?

Four specific traps, each one caused by treating "Gemini image generation" as a single, uniform thing when it isn't.

Confusing the model ID with the marketing name. gemini-3-pro-image is Nano Banana Pro. gemini-3-pro-preview is a completely different, already-shut-down text model. One character apart, and only one of them still answers. If you're copying a model string from an old tutorial, verify it against Google's current models page rather than assuming the name maps the way it looks like it should.

Assuming a lowercase size value works everywhere. Across the three Gemini 3-generation image models, image_size rejects "1k" and requires "1K". This is a small thing until it's the reason an otherwise-correct request throws an error on its first real call.

Assuming the app and the API give you the same file. They don't, and the gap is the download path, not the model. What the app hands you back after a tap on "Download" is governed by the app's own resolution ceiling, separate from whatever the underlying model generated. If a specific pixel count matters to your deliverable, generate through AI Studio, the API, or Vertex, where you set image_size yourself, rather than trusting the app's download button.

Assuming a spec that's true on one surface is true on the other. The aspect-ratio discrepancy above is the clearest example, but it's a pattern worth generalizing: when a number matters to your build, check it on the surface you're actually shipping against, not on whichever page you found first.

Where Does Prompt Architects Fit Into This?

Nowhere in the image itself, and that's worth saying plainly. Prompt Architects doesn't generate images. Every model discussed above, the surface you call it through, the resolution it hands back, belongs to Google. What we build is the layer before that: a structured brief with role, task, format, and constraints, saved with variables so the same skeleton works across a product catalog, a client roster, or a dozen social formats, instead of getting rewritten from a chat scrollback every time. If you're standing up a prompt template for image work across a team, that's the actual product; the render itself still goes to whichever Nano Banana model and surface fit the job.

The image prompt tools live in the web app and in the Chrome extension, which works directly on the Gemini app among the platforms it supports. The free plan covers 5 prompt enhancements a day, forever, enough to test whether the workflow fits before anything is paid for. Current pricing is on the pricing page.

Sources

All accessed September 3, 2026:

• Gemini API: image generation and prompting guide
• Gemini API: models
• Gemini API pricing
• Gemini API deprecations
• Gemini 3 Pro Image (Nano Banana Pro) model card, Gemini Enterprise Agent Platform
• Google: Generate & edit images with Gemini Apps