什么是 图像生成技能？

该技能为包括 Midjourney、Flux 1.1 Pro 和 Ideogram 在内的最先进图像生成模型提供统一接口。它通过根据用户意图（无论是电影感写实还是矢量风格图标）自动选择最佳引擎，简化了 Openclaw Skills 的创作流程。该技能利用来自 Legnext.ai 和 fal.ai 的专业级 API，确保了高稳定性和高速度。它旨在单个交互轮次中处理异步轮询和自动放大等复杂工作流，为开发者和最终用户提供无缝体验。

下载入口:https://github.com/openclaw/skills/tree/main/skills/wells1137/image-q

安装与下载

1. ClawHub CLI

从源直接安装技能的最快方式。

2. 手动安装

将技能文件夹复制到以下位置之一

优先级：工作区 > 本地 > 内置

3. 提示词安装

将此提示词复制到 OpenClaw 即可自动安装。

图像生成技能 应用场景

• 使用 Midjourney 集成创建电影感和艺术感视觉效果。
• 使用 Flux 1.1 Pro 生成照片级的产品照或人像。
• 通过 Ideogram v3 设计徽标和文字密集的图形。
• 使用 Recraft v3 制作矢量艺术和扁平化设计资产。
• 使用 Flux Schnell 等超快模型进行快速原型设计以实现快速迭代。

1. 分析用户请求以识别主体、风格和特定的模型要求。
2. 使用针对光影、细节和构图的模型特定描述符增强用户提示词，以确保高质量输出。
3. 通过 CLI 执行带有适当参数（如宽高比和模型 ID）的生成脚本。
4. 对于 Midjourney 任务，执行同轮次轮询以监控任务状态直至完成，无需额外的用户输入。
5. 直接向用户界面返回高质量、永不过期的图像 URL。

图像生成技能 配置指南

在配置文件中安装必要的依赖项并配置环境变量，即可开始使用 Openclaw Skills 的这部分功能。

npm install @fal-ai/client

将您的 API 密钥添加到 openclaw.json 文件中：

{
  "skills": {
    "entries": {
      "image-gen": {
        "enabled": true,
        "env": {
          "FAL_KEY": "your_fal_key",
          "LEGNEXT_KEY": "your_legnext_key"
        }
      }
    }
  }
}

或者，您可以通过设置 IMAGE_GEN_PROXY_URL 使用代理模式，以避免手动管理 API 密钥。

图像生成技能 数据架构与分类体系

该技能利用结构化的 CLI 参数系统，并返回元数据以进行持久访问和跟踪。

结果以经过验证的 URL 数组形式返回，特别关注 displayImageUrl 或 imageUrls，以避免过期的临时链接。

name: image-gen
description: Generate images using multiple AI models — Midjourney (via Legnext.ai), Flux, SDXL, Nano Banana (Gemini), and more via fal.ai. Automatically picks the best model based on user intent, or lets the user specify one explicitly.
homepage: https://legnext.ai
metadata: {"openclaw":{"emoji":"??","primaryEnv":"FAL_KEY","requires":{"env":["FAL_KEY","LEGNEXT_KEY"]},"install":[{"id":"node","kind":"node","package":"@fal-ai/client","label":"Install fal.ai client (npm)"}]},"proxy":{"url":"https://image-gen-proxy.vercel.app","freeLimit":100}}

Image Generation Skill

This skill enables you to generate images using a variety of state-of-the-art AI models. It supports:

• Midjourney (via Legnext.ai) — Best for artistic, cinematic, and highly detailed images. Faster and more stable than other MJ providers.
• Flux 1.1 Pro (via fal.ai) — Best for photorealistic images and complex scenes.
• Flux Dev (via fal.ai) — Fast, high-quality generation for general use.
• Flux Schnell (via fal.ai) — Ultra-fast generation (<2s), great for quick drafts.
• SDXL (via fal.ai fal-ai/fast-sdxl) — Fastest SDXL endpoint, great for stylized art and LoRA support.
• Nano Banana Pro (via fal.ai fal-ai/nano-banana-pro) — Google Gemini-powered image generation and editing.
• Ideogram v3 (via fal.ai) — Best for images with text, logos, and typography.
• Recraft v3 (via fal.ai) — Best for vector-style, icon, and design assets.

Model Selection Guide

When the user does not specify a model, use this guide to pick the best one:

How to Use This Skill

Basic Usage

When a user asks to generate an image, follow these steps:

1. Understand the request: Identify the subject, style, and any specific requirements.
2. Select a model: Use the guide above, or honor the user's explicit model choice.
3. Enhance the prompt: Expand the user's prompt with relevant style, lighting, and quality descriptors appropriate for the chosen model.
4. Call the generation script: Use the exec tool to run the generation script.
5. Return the result: Present the image URL(s) to the user.

User Experience Rules (important)

• Same-turn polling for Midjourney: After submitting a Midjourney job, do not reply "已提交，完成后通知你" and end your turn. The bot cannot push a message later — the user would have to ask "还没好?" to trigger the next turn. Instead, in the same turn, keep calling --poll --job-id every ~15s until status: "completed", then send the result in that same turn. For multiple parallel jobs, poll all job_ids until all are completed, then send one message with all results.
• Links for Midjourney (Legnext): When sending the result, use only displayImageUrl or imageUrls from the script output. Never send imageUrl (the grid) — it is cdn.legnext.ai/temp/... and expires (shows as broken). Use only cdn.legnext.ai/mj/... links.

Calling the Generation Script

Use the exec tool to run the Node.js script at {baseDir}/generate.js:

node {baseDir}/generate.js r
  --model  r
  --prompt "" r
  [--aspect-ratio ] r
  [--num-images <1-4>] r
  [--negative-prompt ""]

Parameters:

• --model: One of midjourney, flux-pro, flux-dev, flux-schnell, sdxl, nano-banana, ideogram, recraft
• --prompt: The image generation prompt (required)
• --aspect-ratio: Output aspect ratio, e.g. 16:9, 1:1, 9:16, 4:3, 3:4 (default: 1:1)
• --num-images: Number of images to generate, 1-4 (default: 1, Midjourney always returns 4)
• --negative-prompt: Things to avoid in the image (not supported by Midjourney)
• --mode: Midjourney speed mode: turbo (default, 10-20s, requires Pro/Mega plan), fast (30-60s), relax (free but slow)
• --auto-upscale: (Midjourney only) After imagine completes, automatically upscale all 4 grid images and return them as 4 individual single images. The output images array will contain 4 separate upscaled URLs instead of a single grid image.

Example:

node {baseDir}/generate.js r
  --model flux-pro r
  --prompt "a majestic snow leopard on a mountain peak, golden hour lighting, photorealistic, 8k" r
  --aspect-ratio 16:9 r
  --num-images 1

? Midjourney Workflow — Submit Then Poll in the SAME Turn (REQUIRED)

Why the user must not have to ask "还没好?": The bot cannot "push" a message by itself. It only replies when you (the agent) produce a response in the current turn. If you reply with "已提交，完成后通知你" and then end your turn, the user will never get the result until they send a new message (e.g. "还没好?"). So you must not end your turn after submitting. In the same turn, keep polling until the job completes, then send the result in that same turn.

Step 1 — Submit job (returns immediately with job_id)

node {baseDir}/generate.js r
  --model midjourney r
  --prompt "" r
  --aspect-ratio 16:9 r
  --async

You get a job_id. Do not reply to the user yet with "已提交" and stop. Continue to Step 2 in the same turn.

Step 2 — In the SAME turn, poll until completed

node {baseDir}/generate.js r
  --model midjourney r
  --poll r
  --job-id 

• If status: "completed" → go to Step 3 and send the result in this turn.
• If status: "pending" or "processing" → call poll again after ~15s (same turn). Repeat until completed (up to ~5 polls, ~75s).
• If still pending after ~75s, then you may reply once: "还在排队，稍后发「还没好」我帮你查"，and end the turn.

Multiple parallel jobs: Submit all jobs first, then in the same turn poll each job_id in turn until all are completed, then send one message with all results. Do not end the turn with "3 个任务已提交" and expect to "notify later" — the user will have to ask "还没好?" to get anything.

Step 3 — Send result in this turn (only imageUrls / displayImageUrl)

When poll returns status: "completed", send one message with the image links. Use only displayImageUrl or imageUrls (never imageUrl — temp link, expires).

Critical — do NOT send fake/expired links: Use only displayImageUrl or imageUrls from the script output (cdn.legnext.ai/mj/...). Never send imageUrl (cdn.legnext.ai/temp/...).

?? 你的图片生成完成了！图1 图2 图3 图4
想要放大哪张？(U1-U4) 或变体？(V1-V4)

Summary: Same-turn polling

1. Submit → get job_id. Do not reply "已提交" and end the turn.
2. In the same turn, poll every ~15s until status: "completed" (or timeout ~75s).
3. When completed, send the result in that same turn. The user must not need to ask "还没好?" to see the result.

Midjourney-Specific Notes

Midjourney is powered by Legnext.ai (faster and more stable than TTAPI). Turbo mode is enabled by default (--turbo), which reduces generation time to ~10-20 seconds (requires a Midjourney Pro or Mega plan). The --aspect-ratio is automatically appended to the prompt as --ar . The model always generates 4 images in a grid. After generation, you can:

• Use --auto-upscale to automatically upscale all 4 images in one command — this is the recommended default for most use cases.
• Ask the user if they want to upscale (U1-U4) or create variations (V1-V4) of any image.
• Use --action upscale --index <1-4> --job-id to upscale a specific image.
• Use --action variation --index <1-4> --job-id to create variations.
• Use --action reroll --job-id to re-generate with the same prompt.
• Add --async to any action to make it non-blocking.

Upscale types (via --upscale-type):

• 0 = Subtle (default): Conservative enhancement, preserves original details. Best for photography.
• 1 = Creative: More artistic interpretation. Best for illustrations.

Variation types (via --variation-type):

• 0 = Subtle (default): Minor changes while preserving composition.
• 1 = Strong: More dramatic variations with significant changes.

# Upscale image 2 from a previous Midjourney generation (async, non-blocking)
node {baseDir}/generate.js r
  --model midjourney r
  --action upscale r
  --index 2 r
  --job-id  r
  --upscale-type 0 r
  --async

# Create a strong variation of image 3 (async)
node {baseDir}/generate.js r
  --model midjourney r
  --action variation r
  --index 3 r
  --job-id  r
  --variation-type 1 r
  --async

# Reroll (regenerate with same prompt, async)
node {baseDir}/generate.js r
  --model midjourney r
  --action reroll r
  --job-id  r
  --async

Prompt Enhancement Tips

• For Midjourney: Add style keywords like cinematic lighting, photorealistic, --v 7, --style raw, --ar 16:9. Legnext.ai supports all MJ parameters.
• For Flux: Add quality boosters like masterpiece, highly detailed, sharp focus, professional photography
• For Ideogram: Be explicit about text content, font style, and layout
• For Recraft: Specify vector illustration, flat design, icon style, SVG-style

Environment Variables

This skill requires the following environment variables to be set in your OpenClaw config:

Configure them in ~/.openclaw/openclaw.json:

{
  "skills": {
    "entries": {
      "image-gen": {
        "enabled": true,
        "env": {
          "FAL_KEY": "your_fal_key_here",
          "LEGNEXT_KEY": "your_legnext_key_here"
        }
      }
    }
  }
}

Example Conversations

User: "帮我画一只在雪山上的雪豹，电影感光效" Action: Select midjourney, enhance prompt to "a majestic snow leopard on a snowy mountain peak, cinematic lighting, dramatic atmosphere, ultra detailed --ar 16:9 --v 7", run script with --auto-upscale --proxy. This will automatically imagine + upscale all 4 images and return them as 4 individual single images in the images array. Present all 4 to the user.

User: "用 Flux 生成一张产品海报，白色背景，一瓶香水" Action: Select flux-pro, enhance prompt, run script with --aspect-ratio 3:4. (Flux is fast ~5s, no async needed)

User: "快速生成一个草稿看看效果" Action: Select flux-schnell for fastest generation (<2 seconds). No async needed.

User: "帮我做一个 App 图标，扁平风格，蓝色系" Action: Select recraft, use prompt with flat design icon, blue color scheme, minimal, vector style.

User: "把第2张图片放大" Action: Run with --model midjourney --action upscale --index 2 --job-id --async, then poll for result.

?? Proxy Mode (Zero API Keys)

If IMAGE_GEN_PROXY_URL is set (or --proxy flag is used), the skill routes all requests through a proxy server instead of calling fal.ai / Legnext.ai directly. This means users don't need any API keys — the proxy handles authentication server-side.

How It Works

User's Agent → generate.js --proxy → Image-Gen Proxy → fal.ai / Legnext.ai
                                         ?
                                    Token Auth
                                  (100 free uses)

Token-Based Authentication

The proxy uses a Token-based authentication system to manage free usage:

1. First use: When you run generate.js with --proxy for the first time, it automatically registers a free token from the proxy server. The token is saved locally at ~/.image-gen-token.
2. Subsequent uses: The token is automatically loaded and sent with every request. No manual action needed.
3. Free quota: Each token has 100 free image generations (all models combined, including Midjourney).
4. One token per IP: Each IP address can only register one token. This prevents abuse.
5. Quota exhausted: When all 100 uses are consumed, you will see a clear message. Upgrade to Pro for unlimited access.

Important: The token file (~/.image-gen-token) persists across sessions. Clearing your AI agent's context will NOT reset your free quota.

Usage

# Via environment variable (recommended — set once in OpenClaw config)
IMAGE_GEN_PROXY_URL=https://image-gen-proxy.vercel.app node {baseDir}/generate.js r
  --model flux-schnell r
  --prompt "a cute cat"

# Via CLI flag
node {baseDir}/generate.js r
  --model flux-schnell r
  --prompt "a cute cat" r
  --proxy r
  --proxy-url https://image-gen-proxy.vercel.app

Proxy Mode for Midjourney

# Submit and get grid (4 images in one)
node {baseDir}/generate.js --model midjourney --prompt "a dragon" --proxy --proxy-url https://image-gen-proxy.vercel.app

# Submit and auto-upscale all 4 images (RECOMMENDED — returns 4 single images)
node {baseDir}/generate.js --model midjourney --prompt "a dragon" --auto-upscale --proxy --proxy-url https://image-gen-proxy.vercel.app

# Poll (does not consume quota)
node {baseDir}/generate.js --model midjourney --poll --job-id  --proxy --proxy-url https://image-gen-proxy.vercel.app

Free Tier Limits (via Proxy)

After the free tier is exhausted, users receive a 402 response with upgrade instructions.