Generating Images with Gemini AI

There are two valid ways to generate images, depending on your system design.

Approach A: Standalone Per-image Generation

Good for simple tools and scripts.

Common Models

Basic Call

const response = await genAI.models.generateContent({
  model: 'gemini-2.5-flash-image',
  contents: prompt,
  config: {
    responseModalities: ['IMAGE'],
    imageConfig: {
      aspectRatio: '9:16',
      imageSize: '1K', // or '2K', '4K'
    },
  },
});

Approach B: Pipeline-oriented Generation

Good for multi-slide video systems.

• Use gemini-3-pro-image-preview as primary model.
• Generate composite grids (for example 3x3) instead of one prompt per slide.
• Use deterministic grid/cell dimensions by videoType and imageSize.
• Split grids downstream into slide assets.

const response = await genAI.models.generateContent({
  model: 'gemini-3-pro-image-preview',
  contents: prompt,
  config: {
    responseModalities: ['IMAGE'],
    imageConfig: { aspectRatio: '9:16', imageSize: '4K' },
  },
});

Extract Base64 Image Data

const parts = response.response.candidates?.[0]?.content?.parts;
if (!parts || parts.length === 0) throw new Error('No image data in response');

const imagePart = parts.find((part: any) => part.inlineData?.data);
if (!imagePart) throw new Error('No inline image data found');

const imageBuffer = Buffer.from(imagePart.inlineData.data, 'base64');

Grounding (Optional)

For supported models/modalities, use conditional grounding helper:

config: withGoogleSearchGrounding(modelId, {
  responseModalities: ['IMAGE'],
  imageConfig: { aspectRatio: '9:16', imageSize: '4K' },
})

Prompt Quality Pattern

A practical prompt structure:

• Subject
• Composition
• Lighting
• Setting
• Context
• Colors

Which Approach to Choose

• Choose Approach A for ad-hoc generation and low orchestration overhead.
• Choose Approach B when you need deterministic batching, slide indexing, and lower orchestration complexity per video.