In the DeepSeek Harness (DSH) ecosystem, adding multimodal capabilities to agents often requires manual handling of external services or complex integrations. The dsh-imagegen plugin integrates image generation directly into the conversation flow by registering the generate_image tool, allowing agents to request and display images during conversations.

Plugin Introduction

This is an image generation plugin for DeepSeek Harness 0.2.x, maintained by user little-traincar. It addresses the need for agents to call image generation APIs and display the results in the chat interface.

Core Capabilities

  • Inline rendering: Generated images are attached to the conversation and rendered directly in the chat window, while same-origin /imagegen/<file> links are also provided as an alternative.
  • Flexible configuration: Supports configuring model IDs per provider, aliases, and overrides at invocation time.
  • Open provider table: Built-in Doubao Seedream and Alibaba Cloud Qwen; compatible with any OpenAI-compatible /images/generations endpoint (including gateways, relays, or local deployments).
  • No watermark: Sends watermark: false by default, with a toggle available for models that reject this field.
  • Text embedding: Text in prompts is embedded verbatim into the image without being rewritten.
  • Image-to-image: Supports reference images via URL, local absolute path, or Data URI.
  • Save to disk: All images are saved to outDir by default.
  • Multiple candidates and fault tolerance: Providers supporting supportsN can generate multiple images in a single request; if some requests fail, successfully generated images are retained.

Installation and Activation

In Desktop mode, you must fully quit the application first before running the installation command:

dsh plugin --profile desktop add @little-traincar/dsh-imagegen

Restart the application to load the plugin.

Configuration

After installation, go to Settings → imagegen. You need to configure the API Key for the provider you are using.

  • Built-in providers: The plugin includes configurations for Doubao and Alibaba Cloud Qwen; you only need to fill in the corresponding API Key.
  • Custom providers: Supports any OpenAI-style POST <baseUrl>/images/generations interface. Fill in the configuration in the Custom providers field, for example local Ollama:
{
  "local": {
    "baseUrl": "http://127.0.0.1:11434/v1",
    "model": "x/flux2-klein",
    "apiKeyOptional": true,
    "sendWatermark": false,
    "size": "1024x1024"
  }
}
  • Model ID: Model IDs can be modified individually in Settings → imagegen → Models, or configured centrally in cordis.patch.yml. Aliases are supported, and you can specify provider and model at invocation time.

Usage Examples

Enter image generation requests directly in the conversation. The agent will call the generate_image tool, and the image will be rendered directly in the chat window.

For example:
“Draw a summer coffee-shop poster: retro magazine collage, cream yellow and coffee brown, portrait; headline ‘Summer Ice Cafe Festival’, subtitle ‘Half price for the second cup.’”

Notes and Uninstallation

This plugin requires DeepSeek Harness version 0.2.x (compatible with 0.2.0-rc.2 and 0.2.1-alpha.1).

To uninstall, you must fully quit the application first, then run:

dsh plugin --profile desktop remove @little-traincar/dsh-imagegen

Note: The uninstall command does not delete generated images (located in outDir) or the imagegen section in configuration files. Clean up manually if needed.

dsh-imagegen provides a flexible, watermark-free, and highly compatible image generation solution, suitable for scenarios where images need to be output directly within the conversation flow. For more details, see the GitHub repository.