Imagen 4 shut down on 17 August 2026 and Google names gemini-3.1-flash-image — Nano Banana 2 — as its replacement. Google publishes no formula for it, but it does publish rules. This sheet is those rules, quoted.
- Nano Banana is a family, not one model. Nano Banana 2 is the workhorse; 2 Lite is the fast one; Pro is for the hardest jobs.
- Describe, do not list. Google: “Describe a scene in rich detail. The more specific you are, the more control you have over the results.”
- Say what the image is for. Google asks for the purpose, not just the subject.
- There is no negative prompt. Google’s advice is a “semantic negative prompt” — describe the empty street instead of banning cars.
- Text in images works best in two steps — generate the words first, then ask for the image with them.
- Every output is watermarked. Google: “All generated images include a SynthID watermark.”
Every rule on this sheet is quoted from Google’s own image generation documentation, read in a browser on 16 September 2026 because the docs refuse automated fetches.
The formula
Google publishes no formula for Nano Banana. This shape is assembled from its own instructions, quoted below: Nano Banana image generation, read 16 Sep 2026.
Template:
[what the image is for] [subject, described specifically] [setting and light, in detail] [camera language: shot size, lens, angle] [any text, in quotes, and its style]. Then: change one thing per follow-up message.
Worked example:
A poster for an open-studio day. A ceramicist in her fifties centring clay on a wheel, flour-dusted apron, hands in focus. North-facing studio at dusk, window light only, muted earth palette, visible clay texture. Medium shot, 50mm, slightly low angle. The headline reads "OPEN STUDIO — SATURDAY" in a plain condensed sans, top left, small.
Breakdown — what each part does
Purpose
Google: “Explain the purpose of the image. The model’s understanding of context will influence the final output.” Its example is that “Create a logo for a high-end, minimalist skincare brand” beats “Create a logo”.
Subject and scene
Google: “Describe a scene in rich detail. The more specific you are, the more control you have over the results.” And: “Be hyper-specific: The more detail you provide, the more control you have.”
Camera language
Google: “Control the camera: Use photographic and cinematic language to control the composition. Terms like wide-angle shot, macro shot, low-angle perspective.”
Text in the image
Google: “Gemini excels at rendering text. Be clear about the text, the font style (descriptively), and the overall design.” But it also warns about order: “When generating text for an image, Gemini works best if you first generate the text and then ask for an image with the text.”
Iteration
Google: “Don’t expect a perfect image on the first try. Use the conversational nature of the model to make small changes.” Its own examples of follow-ups: “That’s great, but can you make the lighting a bit warmer?” and “Keep everything the same, but change the character’s expression to be more serious.”
Complex scenes
Google: “Use step-by-step instructions: For complex scenes with many elements, break your prompt into steps.” Build the background, then the foreground object, then the hero item.
Does it take a negative prompt?
Google, Nano Banana image generation, read 16 Sep 2026: “Use ‘semantic negative prompts’: Instead of saying ‘no cars,’ describe the intended scene positively: ‘an empty, deserted street with no signs of traffic.’”
Note what that example does: it keeps the word “no”, and simply makes the absence concrete. The thing to avoid is a bare ban with nothing in its place.
On Nano Banana, an exclusion is a description. Name the empty street, the bare wall, the clean frame — not the thing you do not want.
What about editing an existing image?
Editing is the same conversation. Google: “Provide an image and describe your change. The model will match the original image’s style, lighting, and perspective.” For one region only, it describes inpainting as conversational: “Conversationally define a ‘mask’ to edit a specific part of an image while leaving the rest untouched.”
For faces and logos, Google is explicit about the cost of vagueness: “To ensure critical details (like a face or logo) are preserved during an edit, describe them in great detail along with your edit request.” You can mix up to 14 reference images on the Gemini 3 image models.
What not to do
gemini-3.1-flash-image “supports character resemblance of up to 4 characters and the fidelity of up to 10 objects in a single workflow”.✕Which Nano Banana model?
| Model | Use it for |
|---|---|
Nano Banana 2 — gemini-3.1-flash-image | Most work |
Nano Banana 2 Lite — gemini-3.1-flash-lite-image | Speed, 1K |
Nano Banana Pro — gemini-3-pro-image | Hardest jobs |
Nano Banana — gemini-2.5-flash-image | Gemini API: 2 Oct 2026 Agent Platform: 15 Mar 2027 |
Google, Gemini Developer API pricing, read in a browser 18 Sep 2026: “Gemini 2.5 Flash Image (gemini-2.5-flash-image) is deprecated and will be shut down on October 2, 2026; migrate to Gemini 3.1 Flash Image or Gemini 3.1 Flash Lite Image to avoid service disruption.” Until 22 Sep 2026 this row said only “Legacy”.
Google Cloud, Agent Platform release notes, 14 Sep 2026, read in a browser 22 Sep 2026: “Gemini 2.5 Flash Image (gemini-2.5-flash-image): Deprecated and scheduled for retirement on March 15, 2027 (extended from October 2, 2026).” That extension is for Agent Platform (formerly Vertex AI). The Gemini Developer API deprecations page, read the same day, still gives October 2, 2026. From 22 Sep 2026 this row gives both dates; on 18 Sep it gave the Developer API date alone.
Google, Nano Banana image generation, read 16 Sep 2026. Google calls Nano Banana 2 “the most versatile model, generalist workhorse model for all tasks” balancing “speed with state-of-the-art 4K generation, world knowledge, and reliable text rendering”; 2 Lite “Our fastest and cheapest Gemini image model”, which is “Not optimized for multiple reference inputs or multi-turn sequential editing” and “only supports 1K resolution”; Nano Banana Pro “The premium choice for the most complex visual tasks”; and of the legacy model: “we strongly recommend that customers transition to Nano Banana 2 Lite”. Read 16 Sep 2026.
Free or paid? What the free plan allows, who owns the output, and when you need to pay: free vs paid AI image, video and music tools.
Where this model sits among the others: which models take a negative prompt, and what each vendor publishes as its formula.
Imagen 4 is gone and the rest of Google’s 2026 renames are on every Google AI tool in 2026.
The checklist
Before you send it:
Google, Nano Banana image generation · Gemini deprecations (Imagen 4 shutdown 17 Aug 2026, replacement gemini-3.1-flash-image) · Models. All read in a browser on 16 September 2026; these pages refuse automated fetches.
Model versions and vendor documentation both move. Re-read the source before relying on a specific number.
Every other sheet in this set: all model cheat sheets.