智绘图片理解：在 Agent 里统一处理识图、OCR 与多图对比
智绘图片理解（@ixhlink/ixhlink-skills-image-explain）由星华智联科技发布，面向 design-media 场景，支持单张或多张图片的内容描述、视觉问答、文字识别与对比分析。用户说出识图、看图、描述图片、提取图片文字、对比图片等意图时可触发。SkillHub 标注为付费 Skill、requires_api_key=false，当前版本 1.0.3。安装方式为按 https://skillhub.cn/install/skillhub.md 安装上述坐标。文章说明能力边界、典型对话用法与使用注意，并给出官方 SkillHub 与 API 地址。
Read MoreDeepSeek Visual Plugin
The DeepSeek Harness plugin intercepts messages containing images and converts images that pure text models cannot process into textual descriptions generated by a vision model, enabling it to interpret images. It allows multimodal models to pass through directly, while pure text models automatically invoke a vision model to translate them. Users must deploy the code to `$DSH_HOME/profiles/node_modules` and configure the model and API key in `cordis.patch.yml`. The plugin covers session logs and tool execution, supports deployment on Windows, Linux, and macOS, and includes a self-check script.
Read MoreDeepSeek Harness Plugin dsh-mindsee: Letting DeepSeek See Images
`dsh-mindsee` is a community plugin for DeepSeek Harness that enables visual capabilities by calling the MindSee service and passing image descriptions to DeepSeek. To install it, add the plugin via the command line and restart the process. To configure it, enter an access token obtained from mindsee.app in Settings. The plugin supports JPEG, PNG, and WebP formats, as well as local file paths. You can paste an image directly or enter a file path when using it, without switching to a visual model. Before use, please review the plugin source code and license.
Read Moredsh-img: Image understanding and local visual tools for DeepSeek Harness
dsh-img is a community visual plugin for DeepSeek Harness designed to provide image understanding capabilities to pure text models. By integrating with a vision backend or local pixel tools, it enables `analyze_image` for image Q&A, OCR, layout understanding, and `vision_ground` for element localization. The plugin supports Zhipu GLM-4V-Flash, Tongyi Qwen-VL, Ollama local models, and OpenAI-compatible endpoints, and includes key-free local tools such as `vision_crop` and `vision_ocr`. Installation requires Node.js >= 20 and the DeepSeek Harness core, and it is installed into a specific profile via `pnpm add`. When using an API backend, the API key must be injected.
Read Moredsh-image-plugins: Integrating Image Understanding and Generation for DeepSeek Harness
The DeepSeek Harness multimodal plugin, dsh-image-plugins, provides image understanding and generation capabilities for the DSH Web interface. The plugin consists of two independent tools, understand_image and generate_image, which respectively support reading workspace images and returning text descriptions, as well as generating and saving images based on prompts. The plugin supports both OpenAI-compatible endpoints and the native DashScope API, though DashScope image generation requires using the native Model Studio API to avoid 404 errors. After installation, the vision and image functions must be enabled separately via the Web settings panel or configuration file, and the corresponding baseUrl, apiKey, and model must be configured. Keys support both plaintext and environment variables.
Read Moredsh-tool-describe-image: Add image understanding capability to DSH
`sala003/dsh-tool-describe-image` is an MIT-licensed plugin within the DSH ecosystem designed to equip text models with image processing capabilities. By connecting to any OpenAI-compatible vision endpoint (such as Bailian, Zhipu AI, etc.), it converts images into text descriptions or structured HTML for use with DeepSeek. Core features include a Web interface for "paste and recognize," tool invocation via path to describe images, querying DeepSeek account balance, and providing a customizable entry point for a "Whale Girl" desk pet. The plugin supports zero-configuration startup, allowing API keys to be configured later via environment variables or the settings panel. Installation requires DSH 0.1.0-rc.6+, installed globally via npm and enabled. It supports PNG, JPG, WebP, and GIF formats, with a single-image limit of 8Mi.
Read Moredsh-dseyes: Providing native image upload experience for DeepSeek Harness text-only models
dsh-dseyes is an MIT-licensed plugin for DeepSeek Harness (DSH) designed to address the limitation where pure text models cannot directly process images. This plugin enables users to paste or drag images into the Web GUI, displaying them as attachments while preserving the original images. Prior to sending the request to DeepSeek, the host automatically invokes the Zhipu GLM free vision model (e.g., glm-4v-flash) to identify image content, generate a textual description, and replace the original image block, allowing the text model to formulate its response based on that description. The plugin supports an automatic fallback chain, result reuse, and diagnostic interfaces; the conversation continues uninterrupted even if image reading fails. Installation requires Node >= 22, configuration of GLM_API_KEY or ZHIPU_API_KEY, and running `dsh plugin --profi...`
Read Moredsh-vision-tools: Adding image understanding capabilities to DSH sessions
`dsh-vision-tools` is a visual plugin within the DSH ecosystem, maintained by moon09300731 (MIT License). It provides a global `vision_understand` tool that uses OpenAI-compatible APIs (supporting Zhipu, Alibaba, SiliconFlow, and OpenAI) to understand local images, enabling image descriptions, text recognition (OCR), and image-based Q&A. Images can be sent via paste, drag-and-drop, or button selection; they are automatically saved to a specified directory and filled into the prompt. The plugin supports rate limiting auto-degradation, real-time application of global and project-level configurations, and provides an "Image Recognition" entry in the Web interface. The installation command is `dsh plugin --profile web add dsh-vision-tools`, and `dsh web` needs to be restarted. Configuration requires...
Read MoreDSH-Vision-Plugin: Adding Image Understanding Capabilities to DSH
dsh-vision-plugin is a plugin developed specifically for DeepSeek Harness (DSH), designed to address the limitation that pure text models cannot directly process images. This plugin leverages OpenRouter's free multimodal models (`:free`) to automatically convert images into text descriptions before sending, thereby endowing text-based models like DeepSeek with image understanding capabilities, all at zero API cost. Key features include: providing a `describe_image` tool to read local images; intercepting and converting images automatically via the `llm/stream` hook when the model does not support image inputs; and automatically writing a `deepseek-vision` provider pointing to OpenRouter into the configuration file on first startup.
Read Moredsh-vision: Adding Image Understanding to DeepSeek Harness's Plain-Text DeepSeek
dsh-vision is a DeepSeek Harness plugin developed by 54xee, designed to add image understanding capabilities to the pure text model DeepSeek. It wraps adapters to convert images into text placeholders and backfills the recognition results for model processing. The plugin defaults to using the Doubao Web channel, requiring only a browser login, with no API key needed and zero cost. It also supports multiple access methods, including Reverse Gravity IDE, Gemini API, Cockpit reverse proxy, aicode direct connection, and any IDE CLI. Core features include visual evidence memory, content hash caching, automatic tier upgrades, and four task modes. Installation is simple, configuration is flexible, and it is suitable for developers who wish to add visual capabilities to DSH at a low cost or those who already have relevant IDE/API quotas. Please note that you need to check...
Read MoreAchieving Native-Level Image Understanding in DeepSeek Harness with dsh-vision
DSH-Vision is a DeepSeek Harness plugin maintained by Oil-Oil, open-sourced under the MIT License, currently at version 0.1.0, and compatible with DSH 0.1.0-rc.6. When the main model can process images, it directly sends the original image. For pure-text DeepSeek interactions, it submits the original image to ZenMux, Bailian, TokenDance or OpenRouter for analysis, then returns the results to the original model as untrusted context for answering. It also supports see-skill and local OCR. This article verifies the installation, configuration and security boundaries based on the table of contents and the GitHub README.
Read More