Android AI Samples

The /samples directory contains various implementations of generative AI features on Android. These samples cover several key patterns:

Hybrid & Cloud Inference

• Hybrid Inference: Demonstrates using both on-device (Gemini Nano via ML Kit) and cloud-based (Gemini via Firebase AI SDK) models, with a fallback mechanism to the cloud when on-device capabilities are unavailable. [Code: samples/gemini-hybrid]
• Gemini Chatbot: A chatbot using the Gemini Flash model. Supports customizing system instructions to modify model tone or persona. [Code: samples/gemini-chatbot]
• Gemini Live API To-do App: Uses the Gemini Live API for voice-based interaction to manage a to-do list. [Code: samples/gemini-live-todo]

Multimodal & Image Generation

• Gemini Image Chat: Uses the Gemini 3 Pro Image model ("Nano Banana Pro") for image generation and iterative tweaks via conversation. [Code: samples/gemini-image-chat]
• Gemini Multimodal: Leverages Gemini Flash for text and image-to-text prompting. [Code: samples/gemini-multimodal]
• Nanobanana: Uses the Gemini 3.1 Flash Image model ("Nano Banana") to generate artistic landscapes, objects, and people. [Code: samples/nanobanana]
• Magic Selfie: Combines the ML Kit subject Segmentation SDK to remove backgrounds with Nano Banana to generate new ones. [Code: samples/magic-selfie]

On-device AI (Gemini Nano via ML Kit)

• On-device Summarization: Uses the GenAI Summarization API. [Code: samples/genai-summarization]
• On-device Image Description: Uses the GenAI Image Description API. [Code: samples/genai-image-description]
• On-device Writing Assistance: Uses the GenAI Rewriting API for proofreading and rewriting. [Code: samples/genai-writing-assistance]

Video Analysis

• Gemini Video Summarization: Uses Gemini Flash to summarize videos with large file support. [Code: samples/gemini-video-summarization]
• Gemini Video Metadata Creation: Uses Gemini Flash to generate thumbnails, descriptions, hashtags, and chapters by providing a YouTube video link in the model context. [Code: samples/gemini-video-metadata-creation]

The Gemini Live Todo sample demonstrates real-time, voice-based interactions using the Gemini Live API. It allows users to manage a to-do list (add, remove, update tasks) through hands-free, conversational commands. The implementation relies on the Firebase AI SDK for Android to interact with the gemini-2.5-flash-native-audio-preview-12-2025 model.

val generativeModel = Firebase.ai(backend = GenerativeBackend.vertexAI()).liveModel(
    "gemini-2.5-flash-native-audio-preview-12-2025",
    generationConfig = liveGenerationConfig,
    systemInstruction = systemInstruction,
    tools = listOf(
        Tool.functionDeclarations(
            listOf(getTodoList, addTodo, removeTodo, toggleTodoStatus),
        ),
    ),
)

try {
    session = generativeModel.connect()
} catch (e: Exception) {
    Log.e(TAG, "Error connecting to the model", e)
    liveSessionState.value = LiveSessionState.Error
}

This sample demonstrates on-device text manipulation using ML Kit GenAI APIs powered by Gemini Nano. It provides two primary capabilities:

1. Proofreading: Correcting grammar and spelling errors.
2. Rewriting: Changing the style of the input text.

The core implementation uses Proofreader and Rewriter clients. When performing inference, the APIs may return multiple results ordered by confidence; the highest quality result is typically the first element in the results list.

// Example of running proofreading inference
private suspend fun runProofreadingInference(textToProofread: String) {
    val proofreadRequest = ProofreadingRequest.builder(textToProofread).build()
    
    // Results are returned in descending order of quality/confidence.
    // We use the first result (index 0) for the highest quality.
    val results = proofreader.runInference(proofreadRequest).await()
    val polishedText = results.results[0].text
}