DeepSeek can't see images. Now it can — locally, privately, for free.
The missing vision layer for DeepSeek-V4, DeepSeek-R1, o1, and every text-only reasoning model.
| Your setup | The pain | The fix |
|---|---|---|
| DeepSeek-V4 Pro + WorkBuddy | Powerful reasoning, no vision | Send an image → skill describes it → keep coding |
| DeepSeek-R1 via Ollama | Amazing reasoning, zero vision | Local bridge gives it eyes |
| Any text-only model | Must switch models to read a screenshot | Never switch again |
DeepSeek is arguably the best reasoning model on the planet. Its one weakness: it can't see. This skill fixes that.
You: [paste a screenshot of a bug report]
│
▼
DeepSeek: "I can't see images" ❌
│
▼ (with Image Vision Bridge)
qwen3.5:4b (local GPU): "The screenshot shows a 500 error page
with stack trace at line 42: NullPointerException..."
│
▼
DeepSeek: "The NullPointerException is at UserService.java:42.
Let's check — it's likely because the user object isn't
initialized before calling getProfile(). Here's the fix:" ✅
Debugging from a screenshot:
You paste a screenshot of an error page → Skill describes: "React error boundary showing 'Cannot read properties of undefined' at Dashboard.tsx line 156" → DeepSeek: "That's a missing null check. Add
data?.metrics ?? []at line 156."
Reading a UI mockup:
You paste Figma export → Skill describes: "A 3-column dashboard with header nav, sidebar filters, and a data table with 6 columns: Name, Status, Priority, Date, Assignee, Actions." → DeepSeek: "I'll scaffold this as a React component with Tailwind. Here's the structure..."
Extracting text from a PDF screenshot:
You paste a section of a research paper → Skill describes full text content → DeepSeek: "Here's a summary and Chinese translation of this section..."
You're using a powerful reasoning model (DeepSeek-V4, DeepSeek-R1, o1, Qwen-Max...) but it can't read images. Every time you need image understanding, you have to switch to a multimodal model, breaking your workflow. Frustrating.
Image Vision Bridge is a lightweight skill for AI coding assistants (WorkBuddy, CodeBuddy, Claude Code, Cursor) that uses a local Ollama vision model to describe images in text. Your text-only model gets the description injected back into the conversation — it "sees" the image without you ever switching models.
💡 All processing stays on your machine. No API keys. No data leaves your computer.
🧠 Optimized for DeepSeek-V4 Pro and DeepSeek-R1 on WorkBuddy. Works around the "IMAGE_NOT_SUPPORTED" restriction.
- 🔒 100% Local & Private — runs entirely on your hardware via Ollama
- ⚡ Tiny footprint — default model is qwen3.5:4b (~3.4 GB), fits on any modern GPU
- 🌍 Multi-language — Qwen models natively support Chinese, English, and more
- 🎯 Zero config — install the skill, pull one model, you're done
- 🧩 Works with any AI assistant — WorkBuddy, CodeBuddy, Claude Code, Cursor, etc.
- 📝 Customizable prompts — extract text, analyze UI, read code screenshots, whatever you need
# 1. Install Ollama (skip if already installed)
# macOS/Linux: curl -fsSL https://ollama.com/install.sh | sh
# Windows: https://ollama.com/download
# 2. Pull the vision model (~3.4 GB)
ollama pull qwen3.5:4b
# 3. Install the skill into your assistant
cp -r image-vision ~/.workbuddy/skills/image-vision
# or: cp -r image-vision ~/.codebuddy/skills/image-vision
# 4. Done! Send an image and watch it work.┌─────────────────────────────────────────────────────────────┐
│ YOUR AI ASSISTANT │
│ │
│ "What's in this screenshot?" ← text-only model │
│ │ (can't see images) │
│ │ │
│ ▼ │
│ ┌─────────────┐ base64 ┌──────────────┐ │
│ │ describe_ │──────────▶ │ Ollama │ │
│ │ image.py │ │ qwen3.5:4b │ │
│ └─────────────┘ text ◀─── │ (local GPU) │ │
│ │ description └──────────────┘ │
│ ▼ │
│ "The image shows a login form with │
│ email/password fields and a blue │
│ 'Sign In' button in the center..." │
│ │
│ ✅ Text model now reasons about the image! │
└─────────────────────────────────────────────────────────────┘
| Model | Size | Quality | Best For |
|---|---|---|---|
qwen3.5:4b ⭐ |
3.4 GB | ★★★★☆ | Daily use, fast startup |
qwen3.5:9b |
6.6 GB | ★★★★★ | High-accuracy tasks |
minicpm-v |
~5 GB | ★★★★☆ | Strong Chinese support |
llava:7b |
~4 GB | ★★★☆☆ | Legacy compatibility |
gemma3:12b |
~7 GB | ★★★★★ | Latest Google model |
# Extract text from a document scan
python scripts/describe_image.py scan.png \
--prompt "OCR: extract all text from this image, preserving line breaks."
# Analyze a UI design mockup
python scripts/describe_image.py design.png \
--prompt "Describe this UI mockup: layout, components, colors, and spacing."
# Read a code screenshot and reproduce the code
python scripts/describe_image.py code_screenshot.png \
--prompt "Extract the complete source code visible in this screenshot, preserving all indentation."
# Summarize a chart or graph
python scripts/describe_image.py chart.png \
--prompt "Describe what this chart shows, including axes labels, trends, and data points."
# Use a different model
python scripts/describe_image.py photo.jpg --model qwen3.5:9b- Ollama v0.3+ (install guide)
- Python 3.9+ (standard library only — no pip install needed)
- VRAM: ~4 GB for qwen3.5:4b, ~7 GB for qwen3.5:9b
- Works on macOS, Linux, Windows (anywhere Ollama runs)
| Local (this skill) | Cloud API (GPT-4V, Claude) | |
|---|---|---|
| Privacy | ✅ Data never leaves your PC | ❌ Images sent to cloud |
| Cost | ✅ Free | ❌ Pay per image |
| Speed | ✅ ~2-5 sec (after first load) | |
| Internet | ✅ Works offline | ❌ Requires connection |
| Rate limits | ✅ Unlimited | ❌ API quotas |
| 你的配置 | 痛点 | 解决 |
|---|---|---|
| DeepSeek-V4 Pro + WorkBuddy | 推理能力无敌，但读不了图 | 发张图 → skill 帮你读懂 → 继续写代码 |
| DeepSeek-R1 via Ollama | 推理能力无敌，但眼睛是瞎的 | 本地桥接给它装上眼睛 |
| 任何纯文本推理模型 | 看张图就要切模型 | 再也不用切了 |
DeepSeek 可能是地球上最强的推理模型。它唯一的短板：不能看图。这个 skill 补的就是这个短板。
你：[粘贴一张报错截图]
│
▼
DeepSeek："我看不了图" ❌
│
▼ （有了这个 skill 之后）
qwen3.5:4b（本地 GPU）："截图显示一个 500 错误页面，
堆栈信息显示 NullPointerException 在 UserService.java:42..."
│
▼
DeepSeek："这个空指针在 UserService.java:42 行，应该是
user 对象没初始化就调了 getProfile()。修复方案：" ✅
对着截图修 bug：
你粘贴一张浏览器报错截图 → skill 读图："React 错误边界显示 'Cannot read properties of undefined'，位置 Dashboard.tsx 第 156 行" → DeepSeek："这里缺了空值检查，把
data.metrics改成data?.metrics ?? []就行"
对着设计稿写前端：
你粘贴 Figma 导出的设计稿 → skill 读图："三栏 Dashboard 布局，顶部导航、左侧筛选栏、中间数据表格有 Name/Status/Priority 等 6 列" → DeepSeek："我用 React + Tailwind 给你搭一个，结构如下..."
对着 PDF 截图做翻译：
你粘贴一篇英文论文的截图 → skill 读图：完整提取原文 → DeepSeek："以下是中文翻译和章节摘要..."
你正在用 DeepSeek-V4、DeepSeek-R1、Qwen-Max 这类超强推理模型干活，但它读不了图片。每次要看图都得切模型，工作流被打断，烦得要死。
Image Vision Bridge 是一个 AI 编程助手（WorkBuddy / CodeBuddy / Claude Code / Cursor）的技能插件——用本地 Ollama 视觉模型把图片翻译成文字描述，再自动注回对话里。你的纯文本模型就"看见"图片了，全程不用切换。
💡 所有处理都在你本地机器上完成，不需要 API Key，数据不离开你的电脑。
- 🔒 100% 本地运行 — 基于 Ollama，数据不出本机
- ⚡ 极小体积 — 默认模型 qwen3.5:4b 仅 3.4 GB，任何现代 GPU 都跑得动
- 🌍 多语言原生支持 — Qwen 模型天然支持中英文及更多语言
- 🎯 零配置 — 装好 skill、拉一个模型，完事
- 🧩 兼容所有 AI 助手 — WorkBuddy、CodeBuddy、Claude Code、Cursor 等
- 📝 自定义提示词 — 提取文字、分析 UI、读代码截图，随你定制
# 1. 安装 Ollama（如已装则跳过）
# macOS/Linux: curl -fsSL https://ollama.com/install.sh | sh
# Windows: https://ollama.com/download
# 2. 拉取视觉模型（约 3.4 GB）
ollama pull qwen3.5:4b
# 3. 安装 skill 到你的 AI 助手
cp -r image-vision ~/.workbuddy/skills/image-vision
# 或: cp -r image-vision ~/.codebuddy/skills/image-vision
# 4. 搞定！发张图试试效果。┌─────────────────────────────────────────────────────────────┐
│ 你的 AI 助手 │
│ │
│ "这张截图里有什么？" ← 纯文本推理模型 │
│ │ （原生不支持读图） │
│ │ │
│ ▼ │
│ ┌─────────────┐ base64 ┌──────────────┐ │
│ │ describe_ │──────────▶ │ Ollama │ │
│ │ image.py │ │ qwen3.5:4b │ │
│ └─────────────┘ 文字◀─── │ (本地 GPU) │ │
│ │ 描述 └──────────────┘ │
│ ▼ │
│ "图片显示一个白色背景的登录界面， │
│ 中间有邮箱和密码输入框，底部是 │
│ 一个蓝色的'登录'按钮..." │
│ │
│ ✅ 推理模型现在能分析图片内容了！ │
└─────────────────────────────────────────────────────────────┘
| 模型 | 大小 | 效果 | 适用场景 |
|---|---|---|---|
qwen3.5:4b ⭐ |
3.4 GB | ★★★★☆ | 日常使用，启动最快 |
qwen3.5:9b |
6.6 GB | ★★★★★ | 高精度需求 |
minicpm-v |
~5 GB | ★★★★☆ | 中文场景最强 |
llava:7b |
~4 GB | ★★★☆☆ | 经典兼容 |
gemma3:12b |
~7 GB | ★★★★★ | Google 最新模型 |
# 提取文档扫描件中的文字（OCR）
python scripts/describe_image.py scan.png \
--prompt "请逐字提取图片中的所有文字，保留换行。"
# 分析 UI 设计稿
python scripts/describe_image.py design.png \
--prompt "这是一个UI设计稿，请分析布局、组件、配色和间距。"
# 读取代码截图并还原代码
python scripts/describe_image.py code_screenshot.png \
--prompt "完整提取截图中可见的源代码，保留所有缩进。"
# 解读图表
python scripts/describe_image.py chart.png \
--prompt "描述这张图表的内容，包括坐标轴标签、趋势和数据点。"
# 换个模型
python scripts/describe_image.py photo.jpg --model qwen3.5:9b| 本地方案（本 skill） | 云端 API（GPT-4V 等） | |
|---|---|---|
| 隐私 | ✅ 数据不出本机 | ❌ 图片上传到云端 |
| 费用 | ✅ 完全免费 | ❌ 按张计费 |
| 速度 | ✅ 2-5 秒（首次加载后） | |
| 网络 | ✅ 离线可用 | ❌ 需要联网 |
| 调用限制 | ✅ 无限制 | ❌ API 配额限制 |
# From GitHub
git clone https://github.com/duchenyu/image-vision-bridge.git
cp -r image-vision-bridge ~/.workbuddy/skills/image-vision
# Or from ClawHub
clawhub install duchenyu/image-vision-bridgeQ: Does this work with any AI assistant? A: Yes — designed for WorkBuddy/CodeBuddy skill system. The core script is standalone Python — any assistant that runs shell commands can use it. Tested with DeepSeek-V4 Pro, DeepSeek-R1, and Qwen-Max.
Q: How fast is it? A: First call: ~15-30 sec (model loads into VRAM). Subsequent calls: 2-5 sec. qwen3.5:4b is specifically optimized for fast inference.
Q: Do I need a powerful GPU? A: qwen3.5:4b needs ~4 GB VRAM. Works on RTX 2060+, GTX 1660 Ti+, Apple M1+, or any GPU with 4+ GB. Falls back to CPU if VRAM is insufficient.
Q: Why not just use GPT-4V or Claude instead? A: Because you want DeepSeek's reasoning for coding. GPT-4V is great at seeing, but DeepSeek is better at thinking. This skill gives you both — DeepSeek's reasoning + local vision.
Q: Is my data safe? A: Yes. Everything runs locally — the image is read from disk, base64-encoded, sent to localhost, and processed on your GPU. Nothing leaves your machine.
PRs welcome! Areas where help is especially appreciated:
- More model support (vLLM, llama.cpp direct, etc.)
- Streaming output
- Multi-image conversations
- Web UI for testing
MIT — use it, modify it, ship it. No strings attached.
DeepSeek is the brain. This is the eyes. 🧠👁️