Deploy and maintain image understanding (OCR + local VLM + cloud VL) and image generation for Codex connected to text-only models like DeepSeek. Use when the user wants to see/read images in the conversation bar with a text-only model, set up automatic OCR or vision-model processing for pasted images, configure image generation via SiliconFlow or local Stable Diffusion/Flux, fix a broken vision proxy, or asks "看图/生图/贴图识别/图片理解/OCR代理" on a Codex+DeepSeek setup. Deploys a local proxy between Cod...
Scanned 9/3/2026
Install to Claude Code
npx -y skills add aiskillstore/marketplace --skill deepseek-vision-bridge --agent claude-codeInstalls into .claude/skills of the current project.
Are you the author of Deepseek Vision Bridge?
Add the live security badge to your README — it updates automatically with every re-scan.
[](https://www.skillsdirectory.com/skills/aiskillstore-deepseek-vision-bridge)More formats (shields.io, HTML) on the badges page.
---
name: deepseek-vision-bridge
description: >-
Deploy and maintain image understanding (OCR + local VLM + cloud VL) and image
generation for Codex connected to text-only models like DeepSeek. Use when the
user wants to see/read images in the conversation bar with a text-only model,
set up automatic OCR or vision-model processing for pasted images, configure
image generation via SiliconFlow or local Stable Diffusion/Flux, fix a broken
vision proxy, or asks "看图/生图/贴图识别/图片理解/OCR代理" on a Codex+DeepSeek
setup. Deploys a local proxy between Codex and the model that converts images
to text before forwarding.
---
# DeepSeek Vision Bridge
为 Codex + 纯文本模型(DeepSeek)部署"看图 + 生图"能力。
核心机制:本地代理拦截 Codex 请求中的图片,用三级引擎(OCR / 本地VLM / 云端VL)
转成文字后再转发给 DeepSeek;生图由 image-gen.js 完成后返回文件路径。
## 部署流程
### Step 1: 环境自检
运行 `scripts/check_env.py`(需要 Python 3),得到 JSON 报告:
```bash
python scripts/check_env.py
```
关注字段:`node`、`ollama_installed`、`ollama_models`、`gpu`、
`proxy_port_open`、`codex_config`、`deepseek_key_configured`。
### Step 2: 确定部署位置与路线
选定代理目录(例如 `~/codex-vision-bridge/` 或用户偏好位置),创建:
```text
<代理目录>/
├── ocr-proxy.js ← 从 assets/ 复制
├── image-gen.js ← 从 assets/ 复制
├── config.json ← 从 assets/config.json.template 复制
├── package.json ← 从 assets/ 复制
├── .env ← 从 assets/.env.template 复制并填写
├── start-proxy.ps1 / .sh ← 从 assets/ 复制(按平台)
└── node_modules/ ← npm install 生成
```
按 `references/model-guide.md` 与用户确认路线:
- 注册了硅基流动 key → 云端 VL 看图 + fast 生图
- 有 GPU 且愿意装模型 → Ollama 本地 VLM + 本地 SDXL/FLUX 生图
- 都不愿意 → 仅 OCR(纯文字图可用)
### Step 3: 复制文件并填配置
复制 assets 下文件到代理目录,然后编辑 `.env`:
```bash
DEEPSEEK_API_KEY=sk-... # 必填,DeepSeek 平台获取
SILICONFLOW_API_KEY=sk-... # 可选,硅基流动获取
LOCAL_VL_MODEL=minicpm-v:8b # 可选,本地 VLM 模型名
OCR_PROXY_PORT=57323 # 默认
```
若用户选本地 VLM,先安装 Ollama 并拉模型:
```bash
# Windows/macOS: https://ollama.com/download
ollama pull minicpm-v:8b
```
### Step 4: 安装依赖并启动代理
```bash
cd <代理目录>
npm install # 安装 tesseract.js
./start-proxy.sh # macOS/Linux
# 或 powershell -ExecutionPolicy Bypass -File start-proxy.ps1 # Windows
```
验证端口监听:57323(或自定义端口)。启动日志在 `<代理目录>/outputs/proxy.log`。
### Step 5: 配置 Codex 指向代理
修改 `~/.codex/config.toml`(Windows 为 `%USERPROFILE%\.codex\config.toml`):
```toml
model = "deepseek-v4-flash"
model_provider = "custom"
[model_providers.custom]
name = "custom"
wire_api = "responses"
requires_openai_auth = true
base_url = "http://127.0.0.1:57323/v1"
approvals_reviewer = "user"
```
关键:`model_provider = "custom"` + `base_url` 指向代理端口。
改配置前先备份原文件。注意 config.toml 用 UTF-8 无 BOM。
### Step 6: 设置开机自启(Windows 可选)
```powershell
powershell -ExecutionPolicy Bypass -File install-autostart.ps1
```
### Step 7: 验证
1. 重启 Codex 会话(让 config.toml 生效)
2. 对话栏贴一张图 → 应看到模型基于图片内容回答
3. 查看代理日志确认路由(OCR / 云端 VL / 本地 VLM)
4. 要求生图 → 应看到 image-gen.js 生成的图片
5. 若纯文字截图 → 提示词带"提取文字",验证走 OCR
## 故障排查
见 `references/troubleshooting.md`,按症状查:
端口未监听 / 400 / 401 / 超时 / OCR 乱码 / Ollama 连接失败 / 生图失败。
## 引擎选型与架构
- 引擎对比与组合策略:`references/model-guide.md`
- 请求链路与缓存设计:`references/architecture.md`
## 重要原则
- 图片二进制不进上下文:代理转文字、生图返回路径,禁止 Base64 内联
- 纯文字图优先 OCR;图文/纯图走 VLM;云端失败自动降级本地
- 每台机器的 key、端口、模型名不同,以用户机器的 .env / config.toml 为准
Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.
No comments yet. Be the first to comment!