Introduction

The core design philosophy of DeepSeek Harness (DSH) is “everything is a plugin.” For scenarios that require image generation, DSH does not include a built-in model by default; developers need to connect external services through plugins. oebeliever/dsh-image2-draw, as a community-maintained branch, fills this gap. It allows users to configure an OpenAI Images-compatible relay endpoint and directly use the gpt-image-2 model for image generation in DSH.

Plugin Overview

This is a DeepSeek Harness image generation plugin with gpt-image-2 as the default model. It provides text-to-image and image-to-image capabilities and supports automatic failover across multiple endpoints. The plugin is community-maintained based on the original version by JuneLearn and includes gateway compatibility fixes. It is released under the MIT License.

Installation and Enabling

In an environment where DSH is already installed, add the plugin using the following command:

dsh plugin --profile web add github:oebeliever/dsh-image2-draw

After installation is complete, restart DSH (run dsh web), and the plugin will be loaded automatically.

Configuration Instructions

Go to Settings → Plugins → Plugin config and find the Image2 Draw settings card.

  1. Basic Parameters:

    • You must provide the baseURL and API Key. The plugin automatically appends /images/generations to short-form baseURL values and derives the image editing endpoint as /images/edits.
    • The default model is gpt-image-2.

  2. Key Management:

    • The API Key is stored only in DSH credentials and does not appear in plaintext configuration files.

  3. Network and Permissions:

    • Ensure network access to registry.npmjs.org and github.com.
    • If the gateway groups models by Token, the API Key must belong to the gateway’s Image group.

Core Features

The plugin provides a chat tool card and a Chat-dock studio above the input box, supporting the following capabilities:

  • Text-to-image (image2-generate): Supports generating 1–8 images; requests are executed sequentially.
  • Image-to-image (image2-edit): Supports 1–8 reference images (PNG / JPEG / WebP) for image editing.
  • Multi-view portrait mode: In Chat-dock studio, upload 2–8 photos of the same person to generate new images with consistent face, body type, and hairstyle.
  • Automatic failover: Supports configuring up to 8 relay endpoints. When encountering network errors, timeouts, or 4xx/5xx status codes, it automatically switches to the next endpoint.
  • Output handling: Provides thumbnails, zoom viewing, and “Save as” in the chat; files are also saved to the outputs/image2/ directory.

Size and Quality Control

  • Size: Supports adaptive sizes, automatic passthrough, gateway-friendly presets, and custom WIDTHxHEIGHT (must meet specific validation rules).
  • Quality: Supports four quality levels: low / medium / high / auto.

Notes

  • Environment requirements: Node.js >= 20 is required (DSH supports 22.19.x or >= 24). Git and pnpm must be installed.
  • Retry mechanism: HTTP 524 and timeout errors are not automatically retried, to avoid duplicate charges.
  • Claude Code compatibility: The same repository is also a Claude Code plugin and provides a Skill through skills/image2-draw.

Repository and Source Code