乐于分享
好东西不私藏
当前位置：夜雨聆风 > 技术教程 > 软件教程 > 如何将在英伟达白嫖的模型接入 OpenClaw:文本模型 + FLUX 生图完整教程

如何将在英伟达白嫖的模型接入 OpenClaw:文本模型 + FLUX 生图完整教程

当前时间： 2026-04-25 15:38:00 更新时间： 2026-04-25 分类：软件教程 评论(0)

如何将在英伟达白嫖的模型接入 OpenClaw:文本模型 + FLUX 生图完整教程

这篇文章只讲两件事：

  1. 1. 如何把 NVIDIA 的文本模型接进 OpenClaw（通过 LiteLLM）
  2. 2. 如何把 flux.2-klein-4b 图像模型接进 OpenClaw（通过 skill）

目标很明确：

  • • 读者能看懂
  • • 跟着做能落地
如果还不知道如何在英伟达官方免费领取模型API key请看我上篇文章:《零门槛白嫖194个AI模型！英伟达免费API保姆申请教程》。

1. 先看整体方案

本次接入分两条链路：

1）文本模型链路

OpenClaw → LiteLLM → NVIDIA OpenAI-compatible API

适用模型：

  • • DeepSeek V3.2
  • • Qwen 3.5 397B
  • • Llama 3.3 70B
  • • GLM 5.1
  • • Mistral Large 3

这些模型是 OpenAI 兼容接口，适合进入 LiteLLM，再由 OpenClaw 统一管理。

2）图像模型链路

OpenClaw → flux-klein-image skill → NVIDIA GenAI API

适用模型：

  • • black-forest-labs/flux.2-klein-4b

这个模型不是 chat/completions，而是独立 GenAI 图像接口，且有固定 ratio / size 规则和 base64 输出，所以更适合做成 skill。

3）接入案例（提示词： 帮我生成在森林里的大灰狼图片）

2. 文本模型接入 OpenClaw（LiteLLM）

第一步：验证 NVIDIA API 是否可用

curl -s https://integrate.api.nvidia.com/v1/models \  -H "Authorization: Bearer $NVIDIA_API_KEY"

确认模型列表中包含：

  • • deepseek-ai/deepseek-v3.2
  • • qwen/qwen3.5-397b-a17b
  • • meta/llama-3.3-70b-instruct
  • • z-ai/glm-5.1
  • • mistralai/mistral-large-3-675b-instruct-2512

第二步：把模型加入 LiteLLM

编辑文件：

~/.openclaw/litellm_config.yaml

加入模型配置（示例）：

-model_name:nvidia-deepseek-v3.2litellm_params:model:openai/deepseek-ai/deepseek-v3.2api_base:https://integrate.api.nvidia.com/v1api_key:nvapi-xxxx-model_name:nvidia-qwen3.5-397blitellm_params:model:openai/qwen/qwen3.5-397b-a17bapi_base:https://integrate.api.nvidia.com/v1api_key:nvapi-xxxx-model_name:nvidia-llama-3.3-70blitellm_params:model:openai/meta/llama-3.3-70b-instructapi_base:https://integrate.api.nvidia.com/v1api_key:nvapi-xxxx-model_name:nvidia-glm-5.1litellm_params:model:openai/z-ai/glm-5.1api_base:https://integrate.api.nvidia.com/v1api_key:nvapi-xxxx-model_name:nvidia-mistral-large-3litellm_params:model:openai/mistralai/mistral-large-3-675b-instruct-2512api_base:https://integrate.api.nvidia.com/v1api_key:nvapi-xxxx

说明：

  • • model_name 是本地别名
  • • model 是 NVIDIA 实际模型 ID

第三步：注册到 OpenClaw

编辑文件：

~/.openclaw/openclaw.json

1）添加模型元数据

（示例：Qwen 3.5）

{"id":"nvidia-qwen3.5-397b","name":"Qwen 3.5 397B (NVIDIA Build)","api":"openai-completions","reasoning":false,"input":["text"],"cost":{"input":0,"output":0,"cacheRead":0,"cacheWrite":0},"contextWindow":131072,"maxTokens":16384}

2）添加 alias 映射

"litellm/nvidia-qwen3.5-397b":{"alias":"Qwen 3.5 397B (NVIDIA Build)"}

其他模型按同样方式加进去。

第四步：重启 LiteLLM

kill <litellm-pid>nohup litellm --config ~/.openclaw/litellm_config.yaml --port 4000 > /tmp/litellm.log 2>&1 &

第五步：验证模型是否生效

查看模型列表：

curl -s http://localhost:4000/v1/models \  -H "Authorization: Bearer sk-openclaw-litellm-key-12345"

验证对话请求：

curl -s http://localhost:4000/v1/chat/completions \  -H "Authorization: Bearer sk-openclaw-litellm-key-12345" \  -H "Content-Type: application/json" \  -d '{    "model": "nvidia-qwen3.5-397b",    "messages": [{"role":"user","content":"你好，测试一下"}],    "max_tokens": 50,    "stream": false  }'

跑通即完成。

3. 文本模型怎么选（快速版）

  • • 中文写作：Qwen 3.5 397B
  • • 代码与技术分析：DeepSeek V3.2
  • • 复杂推理：GLM 5.1
  • • 英文内容：Llama 3.3 70B

4. 图像模型接入 OpenClaw（skill）

flux.2-klein-4b 不是标准 chat/completions，而是 NVIDIA GenAI 图像接口。它有固定 ratio / size 规则，返回 base64 图片数据，还需要本地落盘和多图批量能力。

因此采用 skill 来接入：

OpenClaw → flux-klein-image skill → NVIDIA GenAI API

5. flux-klein-image skill 做了哪些事

这个 skill 不是简单的“调接口”，而是把模型的规则一次性封装好：

  • • ratio → size 映射（1:1、16:9、9:16…）
  • • 中文风格别名（写实、电影感、赛博朋克…）
  • • style prompt 扩展
  • • steps / count / seed / width / height 边界校验
  • • base64 图片解码并写入本地文件
  • • 多图批量生成与 preview 拼图
  • • .env 自动加载（保证 NVIDIA_API_KEY 可读）

6. flux-klein-image skill 文件内容

1）SKILL.md

---name: flux-klein-imagedescription: Generate images with NVIDIA FLUX.2-Klein-4B via NVIDIA API. Use when the user wants to create images with FLUX, especially for fast text-to-image generation, aspect-ratio presets, style-guided prompting, multi-image batch generation, and contact-sheet previews.---# FLUX Klein ImageUse this skill when the user wants to generate images with NVIDIA's `black-forest-labs/flux.2-klein-4b` model.## What it does- Text-to-image generation- Aspect ratio presets like `1:1`, `16:9`, `9:16`, `4:5`, `5:4`, `3:2`, `2:3`- Style-guided prompting- Chinese style aliases- Multi-image batch generation via automatic seed variation- Optional 4-grid / contact-sheet preview image- Saves generated images to local files## Runtime requirements- Environment variable: `NVIDIA_API_KEY`- Binary: `python3`- Python package: `Pillow` (optional but recommended for preview grids)## UsageSingle image:python3 <SKILL_ROOT>/scripts/flux_klein.py '{"prompt":"a cinematic mountain landscape at sunrise","ratio":"1:1","style":"photorealistic","seed":0,"steps":4,"output":"./flux-output.jpg"}' Batch generation with preview grid:python3 <SKILL_ROOT>/scripts/flux_klein.py '{"prompt":"a cow grazing in a grassy field","ratio":"9:16","style":"电影感","count":4,"make_preview":true,"output":"./outputs/cow.jpg"}'## Parameters- `prompt` (required): text prompt- `ratio` (optional): one of `1:1`, `16:9`, `9:16`, `4:5`, `5:4`, `3:2`, `2:3`- `style` (optional): English or Chinese style name- `width` (optional): one of 768, 832, 896, 960, 1024, 1088, 1152, 1216, 1280, 1344- `height` (optional): one of 768, 832, 896, 960, 1024, 1088, 1152, 1216, 1280, 1344- `seed` (optional): integer, `0` means random for single image, batch mode auto varies seed- `steps` (optional): 1-4- `count` (optional): number of images to generate, `1-4`, default `1`- `make_preview` (optional): whether to generate a contact-sheet preview, default `false`- `output` (optional): output image path, default `./flux-klein-output.jpg`## Ratio mapping- `1:1` -> 1024x1024- `16:9` -> 1344x768- `9:16` -> 768x1344- `4:5` -> 896x1152- `5:4` -> 1152x896- `3:2` -> 1216x832- `2:3` -> 832x1216## Style presets### English- `photorealistic`- `illustration`- `anime`- `cinematic`- `minimalist`- `product`- `oil-painting`- `cyberpunk`### Chinese aliases- `写实` -> `photorealistic`- `插画` -> `illustration`- `动漫` -> `anime`- `电影感` -> `cinematic`- `极简` -> `minimalist`- `产品图` -> `product`- `油画` -> `oil-painting`- `赛博朋克` -> `cyberpunk`## Output behavior- If `count=1`, returns a single file path- If `count>1`, returns a list of generated files- Batch mode appends suffixes like `_01`, `_02`, `_03`, `_04`- If `make_preview=true`, also returns a preview image path like `_preview.jpg`## Notes- This skill calls NVIDIA's direct GenAI endpoint, not the standard OpenAI images API.- The response contains JPEG image data inbase64.- If both `ratio` and `width`/`height` are passed, `width`/`height` win.- If Pillow is unavailable, generation still works but preview grids are skipped.- If the user only asks for a concept or prompt, help refine the prompt before generation.

2）flux_klein.py（核心结构）

import os, sys, json, base64, random, mathfrom pathlib import PathAPI_URL = "https://ai.api.nvidia.com/v1/genai/black-forest-labs/flux.2-klein-4b"ALLOWED_SIZES = {768, 832, 896, 960, 1024, 1088, 1152, 1216, 1280, 1344}RATIO_MAP = {"1:1": (1024, 1024),"16:9": (1344, 768),"9:16": (768, 1344),"4:5": (896, 1152),"5:4": (1152, 896),"3:2": (1216, 832),"2:3": (832, 1216),}STYLE_MAP = {"photorealistic": "realistic photography, natural light, lifelike details, high realism","illustration": "clean digital illustration, polished shapes, vivid colors, refined composition","anime": "anime style, expressive lines, stylized shading, vibrant character art aesthetics","cinematic": "cinematic lighting, dramatic atmosphere, film still, rich depth and mood","minimalist": "minimal composition, clean background, simple visual language, uncluttered design","product": "product photography, studio lighting, clean composition, premium commercial look","oil-painting": "oil painting, painterly texture, rich brush strokes, classic fine art feel","cyberpunk": "cyberpunk aesthetic, neon lighting, futuristic city mood, high contrast",}STYLE_ALIASES = {"写实": "photorealistic","插画": "illustration","动漫": "anime","电影感": "cinematic","极简": "minimalist","产品图": "product","油画": "oil-painting","赛博朋克": "cyberpunk",}

7. 运行前的关键要求

1）准备 API key

将 NVIDIA API key 写入：

~/.openclaw/.env

内容：

NVIDIA_API_KEY=nvapi-xxxx

2）安装依赖

pip install requests pillow

8. 示例调用

单图生成

python3 <SKILL_ROOT>/scripts/flux_klein.py '{"prompt":"a cinematic mountain landscape at sunrise","ratio":"1:1","style":"电影感","seed":0,"steps":4,"output":"./outputs/mountain.jpg"}'

多图生成 + 预览拼图

python3 <SKILL_ROOT>/scripts/flux_klein.py '{"prompt":"a cow grazing in a grassy field","ratio":"9:16","style":"写实","count":4,"make_preview":true,"output":"./outputs/cow.jpg"}'

9. 最后总结

  • • 文本模型适合走 LiteLLM，进入 OpenClaw 模型体系
  • • flux.2-klein-4b 适合做成 skill，把规则一次性封装好
  • • skill 的价值不是“包一层”，而是把模型规则和输出能力收口成可复用动作

按照这篇文章的步骤操作，读者可以把 NVIDIA 的文本模型和 FLUX 图像模型完整接入 OpenClaw，并能真正投入工作流使用。

如果还不知道如何在英伟达官方免费领取模型API key，看我上篇文章:《零门槛白嫖194个AI模型！英伟达免费API保姆申请教程》。