VisionForge: Visual Understanding and Image Generation Plugin for DeepSeek Harness

Introduction¶

One of the core design principles of DeepSeek Harness (DSH) is “everything is a plugin.” In model inference scenarios, developers often need visual understanding and image generation. Some existing visual plugins only provide image-reading capabilities and lack generation and editing features. VisionForge is a DSH plugin built by adapting an open-source plugin. It not only completes the visual-understanding loop, but also adds image generation and editing capabilities, aiming to provide DSH with a lightweight and unified visual processing solution.

Core Features¶

VisionForge provides full-pipeline capabilities from visual understanding to image generation, mainly including:

• Paste to Read: With the official vision model, pasting an image into the input box displays a native preview; with a pure text model, the plugin takes over the pasted content, automatically converts it to a local path, and calls the vision engine to parse it.
• Multi-provider vision engines: Supports multiple vision engines such as Qwen, OpenAI-compatible endpoints, Anthropic, Gemini, Antigravity, Claude, and Kimi. Supports failover chains, so if one key fails, it automatically rotates to the next one.
• Image generation and editing: Provides visionforge_generate_image for image generation and visionforge_edit_image for image editing (such as outfit replacement, background replacement, and multi-image fusion).
• Image zoom and download: Generated images are displayed in the conversation as 90px thumbnails. Click an image to enlarge the original using the system image viewer; click “Save Image N” to download it to the D: drive root directory (if no D: drive exists, save it to the user directory).
• Parsing priority: Supports two modes: “official first” (try the official vision model first, and fall back to the plugin only if it fails) and “plugin first.”
• Engine translator (custom engines): In addition to built-in engines, any vendor engine such as Google Gemini or Imagen can be added without modifying code.
• Image quality and prompt enhancement: Supports resolution vocabulary (4K/2K/1080P). For image-to-image tasks, it automatically appends protective constraints (such as preserving facial features and not adding watermarks).
• Unified directories and cache cleanup: Generated-image caches and pasted-original-image caches are all converged into fixed directories. Generated images and pasted caches older than 3 days are automatically expired and cleaned up.
• Local loopback service: Preview and download use a fixed local machine port and do not depend on temporary OSS links from providers.

Installation and Configuration¶

The plugin is installed via npm. Before installing, make sure the old version of @liustack/modlens has been removed to avoid two plugins registering same-named tools at the same time.

# 安装
npx -y @deepseek-ai/dsh plugin --profile desktop add @lr611/visionforge

# 若存在旧版，先移除
npx -y @deepseek-ai/dsh plugin --profile desktop remove @liustack/modlens

After installation, go to DSH Settings → Built-in Plugins → VisionForge Configuration card. Select an engine (such as Qwen), enter the API key, endpoint URL, and model, and save to make the changes effective.

Usage Examples¶

• Reading images: Copy and paste an image into the input box. Depending on the parsing priority setting, the plugin automatically handles the image path and calls the vision model to parse it.
• Generating an image: Enter a command in the conversation, for example “Generate a scarecrow in the fields.” The model calls visionforge_generate_image, and a thumbnail appears in the conversation. Click to enlarge or save it.
• Editing an image: Paste 1–3 images and enter the command “Change the girl’s clothes into a white JK skirt.” The model calls visionforge_edit_image and generates an edited result with preview and download options.

Notes¶

• Environment requirements: DSH Desktop v2.0.17 / DSH harness 0.2.0-rc.2 (verified in October 2026). Major DSH version updates may change the plugin interface; please re-verify after upgrading.
• Version compatibility: The plugin is only related to the harness version. As long as the harness version is the same, there is no substantial functional difference between the official desktop client and third-party desktop clients, but the official DSH Desktop is recommended.
• Historical versions: Historical versions must match the harness version. For example, directly installing an older version on harness 0.2.0-rc.2 may fail to load due to incompatible peerDependencies.