Media Generation

Dartantic provides a unified generateMedia() API that works across multiple providers. Each provider uses its own underlying mechanism (native image generation, code execution, etc.) but the API surface is consistent.

Provider Support

Provider	Images	PDFs	CSV/Files	Mechanism
OpenAI Responses	Native + Code	Code	Code	DALL-E + Code Interpreter
Google	Native + Code	Code	Code	Nano Banana Pro (`gemini-3-pro-image-preview`) + Code Execution
Anthropic	Code	Code	Code	Code Interpreter (matplotlib, reportlab)

When you request an image, Dartantic routes to the provider’s native image model (DALL-E for OpenAI, Nano Banana Pro for Google). For other file types, it uses each provider’s code execution environment.

Basic Usage

final agent = Agent('google'); // or 'openai-responses', 'anthropic'

final result = await agent.generateMedia(
  'Create a minimalist robot mascot for a developer conference.',
  mimeTypes: const ['image/png'],
);

// Access generated assets
for (final asset in result.assets) {
  if (asset is DataPart) {
    print('Generated: ${asset.name} (${asset.mimeType})');
    File('output/${asset.name}').writeAsBytesSync(asset.bytes);
  }
}

Image Editing with Attachments

All providers support image editing by providing an input image as an attachment. Google uses native Imagen for high-quality editing, while OpenAI and Anthropic use code execution (PIL/Pillow):

import 'package:cross_file/cross_file.dart';

final agent = Agent('google'); // or 'openai-responses', 'anthropic'

// Load the input image
const imagePath = 'input/robot_bw.png';
final imageFile = XFile.fromData(
  await File(imagePath).readAsBytes(),
  path: imagePath,
  mimeType: 'image/png',
);
final imagePart = await DataPart.fromFile(imageFile);

// Edit the image
final result = await agent.generateMedia(
  'Colorize this black and white robot drawing. '
  'Make the robot body blue, the eyes bright green, and add '
  'orange/yellow accents.',
  mimeTypes: const ['image/png'],
  attachments: [imagePart],
);

// Save the edited image
for (final asset in result.assets) {
  if (asset is DataPart) {
    File('output/${asset.name}').writeAsBytesSync(asset.bytes);
  }
}

Common image editing use cases:

Colorization: Add color to black and white images
Style Transfer: Apply artistic styles to existing images
Object Removal: Remove unwanted objects (inpainting)
Background Replacement: Change backgrounds while preserving subjects
Enhancement: Improve quality, lighting, or resolution

Example MIME Types

The mimeTypes parameter tells the provider what kind of output you want, e.g.

image/png, image/jpeg - Images
application/pdf - PDF documents
text/csv - CSV files
Other file types depend on provider code execution capabilities

Streaming Generation

Use generateMediaStream() to receive incremental updates:

final agent = Agent('openai-responses');

await for (final chunk in agent.generateMediaStream(
  'Create a colorful abstract art piece.',
  mimeTypes: const ['image/png'],
)) {
  // Check for partial/preview images in metadata
  final previews = chunk.metadata['image_generation'] as List?;
  if (previews != null) {
    for (final preview in previews) {
      final b64 = preview['partial_image_b64'] as String?;
      if (b64 != null) {
        print('Preview available at index ${preview['partial_image_index']}');
      }
    }
  }

  // Check for completed assets
  if (chunk.isComplete) {
    for (final asset in chunk.assets) {
      if (asset is DataPart) {
        File('final.png').writeAsBytesSync(asset.bytes);
      }
    }
  }
}

Specifying Media Models

Use the model string to specify which media model to use:

// Use Nano Banana Pro explicitly for Google image generation
final agent = Agent('google?media=gemini-3-pro-image-preview');

// Or combine with a chat model
final agent = Agent('google?chat=gemini-3-pro-preview&media=gemini-3-pro-image-preview');

The model string is passed through to the provider’s API, so you can use new models as soon as they’re available without waiting for a Dartantic update.

Provider-Specific Options

Each provider accepts options for customization:

OpenAI Responses

final agent = Agent(
  'openai-responses',
  mediaModelOptions: OpenAIResponsesMediaGenerationModelOptions(
    imageModel: 'gpt-image-1',  // Specific image model
  ),
);

Google

final agent = Agent(
  'google?media=gemini-3-pro-image-preview',  // Use Nano Banana Pro
  mediaModelOptions: GoogleMediaGenerationModelOptions(
    aspectRatio: '16:9',  // Image aspect ratio
  ),
);

Anthropic

final agent = Agent(
  'anthropic',
  mediaModelOptions: AnthropicMediaGenerationModelOptions(
    // Uses code interpreter for all generation
  ),
);

Result Structure

MediaGenerationResult contains:

assets - List of Part objects (typically DataPart with bytes)
links - List of LinkPart for hosted URLs (if applicable)
messages - Conversation messages generated during the run
metadata - Provider-specific metadata (previews, progress, etc.)
isComplete - Whether generation is finished (for streaming)

Examples

Server-Side Tools - Lower-level access to code interpreters
Multimedia Input - Send images and files to models

Getting Started

Features

Chat Widget

Advanced

Server Side Tools

Integration

Provider Support

Basic Usage

Image Editing with Attachments

Example MIME Types

Streaming Generation

Specifying Media Models

Provider-Specific Options

OpenAI Responses

Google

Anthropic

Result Structure

Examples

​Provider Support

​Basic Usage

​Image Editing with Attachments

​Example MIME Types

​Streaming Generation

​Specifying Media Models

​Provider-Specific Options

​OpenAI Responses

​Google

​Anthropic

​Result Structure

​Examples

​Related Topics

Provider Support

Basic Usage

Image Editing with Attachments

Example MIME Types

Streaming Generation

Specifying Media Models

Provider-Specific Options

OpenAI Responses

Google

Anthropic

Result Structure

Examples

Related Topics