evanca/flutter-ai-rules/generate-images-with-firebase-ai

Generating images with Firebase AI Logic

Gemini image models return interleaved text and image parts from one call. The response is a sequence to walk, not a string to read.

A request that comes back empty is usually a configuration problem rather than a bug in your code, so section 1 covers the three settings that cause it.

1. Three things that block the very first call

A first call that returns an error, an empty response, or a 403 is almost always one of these rather than your Dart. Rule them out before you debug code.

Billing. Image generation has no free tier. On a Spark-plan project the image models return limit: 0 for generate_content_free_tier_requests, so the first request fails on quota having made zero requests. Text models do work on Spark, which means "my other Gemini call works" proves nothing. Upgrade to Blaze, then verify the current limits rather than trusting this note:

gcloud services quota list --service=generativelanguage.googleapis.com --consumer=projects/YOUR_PROJECT_ID

App Check. Firebase AI Logic enforces App Check when the project has it turned on. Otherwise the endpoint is open to anyone who extracts your config from the shipped client, and that config is public by design. Anything reachable from a device you do not control needs App Check. Debug builds attest with a debug provider, release builds with a real one. Web has a specific trap that costs an afternoon, described in references/setup.md.

`responseModalities`. Without it the model has no permission to return an image, so you get text describing the picture it would have drawn. Set both modalities, as in the call below.

2. The minimal call that works

final model = FirebaseAI.googleAI().generativeModel(
  model: 'gemini-3.1-flash-image',
  generationConfig: GenerationConfig(
    responseModalities: [
      ResponseModalities.text,
      ResponseModalities.image,
    ],
    imageConfig: ImageConfig(
      aspectRatio: ImageAspectRatio.landscape16x9,
      imageSize: ImageSize.size2K,
    ),
  ),
);
 
final response = await model.generateContent([
  Content.multi([
    TextPart(prompt),
    InlineDataPart('image/jpeg', selfieBytes), // omit for text-to-image
  ]),
]);

FirebaseAI.googleAI() is the Gemini Developer API. Prefer it: it needs no GCP surface of its own, and its free tier covers text. FirebaseAI.vertexAI() requires Blaze regardless of model and buys you GCP-side controls, so reach for it when the project already lives in Vertex rather than by default.

Do not pass appCheck: or auth: to googleAI(). Both parameters are deprecated in current firebase_ai; the instance resolves them from the FirebaseApp on its own.

For model IDs, aspect-ratio and size enums, and when Imagen beats Gemini, read references/models.md.

3. Reading the response

The convenience .text accessor does not capture the full sequence, so walk the parts directly. Pattern-match on the part type, which keeps the switch correct when the SDK adds new part types:

Uint8List? image;
final buffer = StringBuffer();
for (final part in candidate.content.parts) {
  switch (part) {
    case InlineDataPart(:final bytes):
      image ??= bytes;          // first image wins
    case TextPart(:final text):
      buffer.write(text);
    default:
      break;
  }
}

Keep the interpretation of a response in its own pure function taking a Candidate. Candidate, Content, TextPart and InlineDataPart are all publicly constructible, so that function is testable with real SDK types, no Firebase and no test doubles. It is the one seam in this stack that unit tests genuinely reach.

When no image comes back

An empty result surfaces as a blank error and reads like a client bug, which sends people debugging the wrong half of the system. The response carries the reason, so report it. In order:

Image-only responses, with no text at all, also happen intermittently on prompts that reliably return both. If you ask for text alongside the image, treat its absence as normal and degrade instead of throwing.

4. Sending a user photo

Downscale before you send. InlineDataPart.toJson() base64-encodes the bytes synchronously on the main isolate, so a full-size phone photo freezes the UI for seconds while the request is built. Gemini downsamples large images anyway, so you pay for detail that is then discarded. Scale at pick time rather than after:

final file = await picker.pickImage(
  source: ImageSource.gallery,
  maxWidth: 1280,
  maxHeight: 1280,
  imageQuality: 85,
);

Size the cap to your subject rather than to a habit. 1280px on the long edge holds a face or a single figure comfortably, while fine texture, legible text in the source, or a wide scene the model has to read across will want more. Match mimeType to what the picker actually returned.

5. Sizing the result

Use ImageConfig when your ratio is one of the supported enum values, because it is a real constraint. Asking for a ratio in the prompt text is a suggestion the model frequently ignores. If you need a ratio the enum does not offer, 2:1 for instance, measure what came back and lay out from the measurement:

final descriptor = await ui.ImageDescriptor.encoded(
  await ui.ImmutableBuffer.fromUint8List(bytes),
);
final size = ui.Size(descriptor.width.toDouble(), descriptor.height.toDouble());
descriptor.dispose();

Wrap it so a failure returns null instead of throwing. An image you cannot measure is still an image worth showing.

6. Getting text and image from one call

You often want machine-readable data about the image the model just drew, such as a caption or the names it invented. Once that data is pixels your app cannot read it, so ask for it as text in the same call and reconcile the two halves in the prompt: "the title painted into the image must match the JSON character for character."

Parse that text defensively. Asked for a fenced ``json block, the model will across one session return fenced JSON, bare JSON, prose-wrapped JSON, and nothing at all. Try the fence, fall back to the first balanced {...}`, and return null instead of throwing.

7. Prompting

The failures here are not code failures, and no unit test reaches them. The recurring ones have specific fixes worth knowing before you write the first prompt: placeholder words painted literally into the artwork, the source photo's clothing surviving an outfit change, text spelled differently in two places, and panels that do not share a background. Read references/prompting.md.

8. Deciding what to test

The seam is at your boundary. Unit tests protect your interpretation of a response and nothing past it, so a green suite says nothing about whether the app produces a good image. references/testing.md covers the layers and what each one cannot reach.

External documentation

Everything here is a summary that will drift. When a detail matters, confirm it at the source.

• Firebase AI Logic docs, including

get started and generate images with Gemini

• `firebase_ai`, the Dart SDK. Its

source is the fastest way to settle an API question: ~/.pub-cache/hosted/pub.dev/firebase_ai-*/lib/src/

• `image_picker`, which supplies the

pickImage call in section 4. A third-party choice you can swap.

• Patrol, a third-party E2E framework by

LeanCode, discussed in references/testing.md

Reference files

• references/setup.md: Firebase console path, provider choice, billing, and the web App Check debug-token trap
• references/models.md: model IDs, aspect-ratio and size enums, Gemini vs Imagen
• references/prompting.md: the mistakes image prompts actually make, and the phrasings that fix them
• references/testing.md: unit, golden, e2e and eval layers, and what each cannot reach