Gemini image and transcription models

Image understanding belongs to a completion model. Image creation and audio transcription use separate model factories.

1. Choose an image factory

imageGenerationModel({ api: 'generateContent', modelId }) uses Gemini generateContent.

imageGenerationModel({ api: 'generateImages', modelId }) uses Imagen generateImages.

Both satisfy Anvia's ImageGenerationModel, but send different Google request shapes.

2. Generate with a Gemini-native image model

import { writeFile } from 'node:fs/promises'
import { generateImage } from '@anvia/core/image-generation'
import {
  GEMINI_2_5_FLASH_IMAGE,
  GeminiClient,
} from '@anvia/gemini'

const gemini = new GeminiClient({
  apiKey: process.env.GEMINI_API_KEY!,
})

const imageModel = gemini.imageGenerationModel({
    api: 'generateContent',
    modelId: GEMINI_2_5_FLASH_IMAGE,
})

const result = await generateImage({
    prompt: 'A minimal diagram of an agent calling two tools',
    model: imageModel,
    width: 1024,
    height: 1024,
    providerOptions: {
        config: {
            imageConfig: {
                imageSize: '1K',
            },
        },
    }
})

await writeFile('agent-diagram.png', result.images[0].data)

The adapter requests text and image modalities, extracts inline image data, and returns Uint8Array bytes. Width and height are reduced to an aspect-ratio string.

The adapter merges your config keys in and then writes its generated fields, so responseModalities and the computed imageConfig.aspectRatio always win. aspectRatio is always recomputed from the request width and height, so a user-supplied aspectRatio has no effect; other imageConfig keys such as imageSize pass through.

3. Generate with Imagen

import { generateImage } from '@anvia/core/image-generation'
import { IMAGEN_4_GENERATE } from '@anvia/gemini'

const imagen = gemini.imageGenerationModel({
  api: 'generateImages',
  modelId: IMAGEN_4_GENERATE,
})

const result = await generateImage({
    prompt: 'A clean editorial illustration of a retrieval pipeline',
    model: imagen,
    width: 1600,
    height: 900,
    providerOptions: {
        config: {
            numberOfImages: 2,
        },
    }
})

console.log(result.images.length)

Verify option names and supported aspect ratios (expressed through width and height) against the exact model and API mode.

4. Transcribe audio

import { readFile } from 'node:fs/promises'
import { transcribe } from '@anvia/core/transcription'

const transcriptionModel = gemini.transcriptionModel({
    modelId: 'gemini-3.6-flash'
})

const transcript = await transcribe({
    audio: {
        data: await readFile('uploads/support-call.wav'),
        filename: 'support-call.wav'
    },
    model: transcriptionModel,
    prompt: 'Use the product names Anvia and Acme Cloud.',
    temperature: 0
})

console.log(transcript.text)

The adapter adds a fixed exact-transcription instruction and appends the supplied prompt as domain guidance.

Filename extensions map .wav, .aac, .ogg, .flac, .m4a, and .opus to media types. Other names use audio/mpeg, so validate uploads before the call.

5. Store output deliberately

Persist generated images in object storage. Apply authorization and retention rules to source audio and transcripts, and avoid logging raw provider media responses.