How to Generate and Edit Images in Android Apps with Gemini Nano Banana

Imagen is on the way out. Firebase documents that Imagen models are deprecated and can shut down as early as 17 August 2026. The replacement path is the Gemini Image family — the models Google nicknamed Nano Banana.

You call them from an Android app with the same Firebase AI Logic SDK you already use for Gemini text. No separate image service. One generateContent() call can return a PNG, or a mix of captions and pictures.

This guide follows the official Firebase image-generation docs and the Android Gemini Developer API page. Model IDs and billing notes are those published as of mid-September 2026. Confirm them in the Firebase model catalog before you ship.

What Nano Banana is (and is not)

Google uses the nickname across three current Gemini Image models:

• gemini-3.1-flash-image — Nano Banana 2. Speed and volume. Official docs list this as the default replacement for most Imagen 4 Fast / standard work.
• gemini-3.1-flash-lite-image — Nano Banana 2 Lite. Cheaper, still an image model.
• gemini-3-pro-image — Nano Banana Pro. Higher fidelity, better in-image text, Google Search grounding on the Pro image model. Firebase lists this path as billing required.

These are cloud models. They are not Gemini Nano on the phone. On-device Nano still handles text and some multimodal analysis through ML Kit and hybrid inference. Image generation goes to the Gemini Image endpoints.

Firebase recommends Gemini Image when you want world knowledge, conversational edits, and interleaved text plus images. Use Imagen only if you still have a narrow quality case — and plan the migration, because Imagen is shutting down.

What you need before the first call

1. A Firebase project with Firebase AI Logic enabled.
2. The Gemini Developer API backend or the Agent Platform Gemini API, following Firebase getting started.
3. A Blaze (pay-as-you-go) plan for current Gemini Image models. Firebase’s model page marks Nano Banana 2 and Pro as billing required.
4. Firebase App Check. Enforcement is scheduled to be required from 2 November 2026. Wire the debug provider now so local builds keep working.
5. Current BoM. Android’s Gemini Developer API snippets use com.google.firebase:firebase-bom:34.19.0 plus firebase-ai and firebase-appcheck-debug.

In the app module:

implementation(platform("com.google.firebase:firebase-bom:34.19.0"))
implementation("com.google.firebase:firebase-ai")
implementation("com.google.firebase:firebase-appcheck-debug")

Prototype prompts in Google AI Studio first. The Android docs explicitly tell you to refine Nano Banana prompts there, then paste the generated snippet.

Generate an image from text

Create a GenerativeModel with an image model ID and set responseModalities to IMAGE if you only want pixels.

val model = Firebase.ai(backend = GenerativeBackend.googleAI()).generativeModel(
    modelName = "gemini-3.1-flash-image",
    generationConfig = generationConfig {
        responseModalities = listOf(ResponseModality.IMAGE)
    }
)

val prompt = "Generate an image of the Eiffel Tower with fireworks in the background."

val bitmap = model.generateContent(prompt)
    .candidates.first().content.parts
    .filterIsInstance<ImagePart>()
    .firstOrNull()
    ?.image

That Kotlin shape is the official sample: the first ImagePart carries a Bitmap. Show it in an Image composable or write it to storage.

If you omit TEXT from responseModalities, the model is supposed to return images only. If you include both TEXT and IMAGE, walk every part in the candidate — captions and bitmaps can arrive in one turn.

Interleaved recipe cards and similar layouts

A useful product pattern is a how-to that needs pictures and copy. Firebase’s example is an illustrated paella recipe. You do not call two models.

val model = Firebase.ai(backend = GenerativeBackend.googleAI()).generativeModel(
    modelName = "gemini-3.1-flash-image",
    generationConfig = generationConfig {
        responseModalities = listOf(
            ResponseModality.TEXT,
            ResponseModality.IMAGE
        )
    }
)

val prompt = """
Generate an illustrated recipe for a paella.
Create images to go alongside the text as you generate the recipe.
""".trimIndent()

val response = model.generateContent(prompt)
for (part in response.candidates.first().content.parts) {
    when (part) {
        is TextPart -> appendCopy(part.text)
        is ImagePart -> appendPlate(part.image)
    }
}

Same idea works for workout cards, furniture assembly, or a travel day plan. Keep the prompt specific about style, aspect, and what must stay consistent across frames.

Edit an existing photo

Pass the user’s bitmap plus an instruction. Official docs support text-plus-image prompts for edits and iterative conversation.

val response = model.generateContent(
    content {
        image(sourceBitmap)
        text("Keep the same person and lighting. Change the background to a quiet Tokyo side street at dusk.")
    }
)
val edited = response.candidates.first().content.parts
    .filterIsInstance<ImagePart>()
    .firstOrNull()
    ?.image

For multi-turn edits, use startChat() with the same image-capable model so later turns keep identity and layout. That is the documented advantage over one-shot Imagen calls: you can say “make the type larger” without re-describing the whole scene.

Gemini versus leftover Imagen

Firebase’s migration guide maps old Imagen IDs onto Gemini Image IDs. The short version:

• Imagen 4 Fast → gemini-3.1-flash-image with a low thinking level
• Imagen 4 standard → gemini-3.1-flash-image with a higher thinking level
• Imagen 4 Ultra → gemini-3-pro-image

Do not leave production traffic on Imagen after the published shutdown window.

Aspect ratio and exact pixel size are not first-class SDK fields yet. Firebase says you should put the size you want in the prompt until those parameters ship.

Documented limits worth coding against:

• Output is PNG.
• Flash image models top out around 1024 px in the older 2.5 notes; Pro image supports up to 4K on the preview/Pro path. Recheck the current model table when you pick a resolution.
• Audio and video are not valid inputs for image-generating Gemini models.
• People in generated or edited images are allowed, subject to safety filters.

App Check and shipping hygiene

From 2 November 2026, App Check enforcement is required for Firebase AI Logic. For debug builds:

Firebase.initialize(context = this)
Firebase.appCheck.installAppCheckProviderFactory(
    DebugAppCheckProviderFactory.getInstance()
)

Run the app, copy the debug token from Logcat, and register it under Firebase console → App Check → Manage debug tokens. Production builds should use Play Integrity, not the debug factory.

Image models cost real money. Cap daily generations per signed-in user, show a progress state (calls are slower than text), and never put a raw Gemini API key in the APK — the Firebase AI Logic client is the supported path on Android.

A small product checklist

• Pick Flash Image for in-app stickers, backgrounds, and drafts; pick Pro Image when typography in the picture has to be readable.
• Iterate prompts in AI Studio, then freeze a template in code.
• Always handle blocked candidates and empty ImagePart lists — safety filters will drop some requests.
• Store results in app-specific storage if the user expects to keep them; do not assume the model will reproduce the same pixels tomorrow.
• Watch the model catalog for Gemini 2.5 image shutdown dates. New work should use 3.x Image IDs.

Conclusion

Nano Banana is not a separate Android API. It is Gemini Image, reached through Firebase AI Logic, with responseModalities set so the model is allowed to return pictures. Text-to-image, photo edits, and mixed recipe-style layouts all use generateContent().

If your app still calls Imagen, treat the August 2026 shutdown date as a hard deadline. Switch the model name, require IMAGE in the config, and put App Check on the calendar before November.

Sources

• Generate and edit images using Gemini (Nano Banana) — Firebase AI Logic
• Gemini Developer API on Android
• Supported models — Firebase AI Logic
• Migrate from Imagen to Gemini Image models
• Firebase AI Logic getting started
• Firebase quickstart Android — Firebase AI samples