| name | niu-image-gen |
| description | Generate or edit images using the Niu Image Gen plugin. Trigger when the user wants to create, draw, generate, or edit images via Niu Image Gen, wants batch image generation, needs AI-generated images saved to disk, or wants to modify an existing image (change background, remove object, change style, etc.). Do not use for the built-in image_gen tool or SVG/vector work. |
Niu Image Gen 🎨
AI image generation with persistent config, zero-confirmation quick mode, and batch mode.
Script location
SCRIPT="$HOME/plugins/niu-image-gen/scripts/generate.mjs"
🔴 输出规则（最高优先级，所有 Branch 都必须遵守）
- 脚本输出 = 最终展示内容：运行脚本后，将脚本的 stdout 输出原样展示给用户。不要改写、省略、翻译、合并或重新组织脚本的输出。
- 禁止代码块包裹：脚本输出是格式化好的纯文本，绝对不要用
``` 代码块包裹。直接作为普通消息文本展示。
- 交互提示 = 原样输出：本文件中所有
「📋 原样输出」 标记后面的引用块（> 开头），必须逐字输出给用户，包括所有 emoji、表格、分隔线。不得用自己的话重新表述。
- 禁止纯文本降级：不要去掉 emoji，不要把表格改成逗号列表，不要把多行合并成一行。
⚠️ ENTRY LOGIC — ALWAYS START HERE
Every time this skill is triggered, run this first:
node "$SCRIPT" --get-config
The output is JSON:
{
"hasKey": true,
"keyPreview": "sk-1aa1...aa4d",
"quickMode": { "quality": "2K", "ratio": "square", "count": 1 },
"batchMode": { "quality": "2K", "ratio": "landscape", "concurrency": 3 }
}
Fields can be null if not yet configured.
Decision tree — pick the FIRST matching rule, top to bottom:
| # | Condition | Go to |
|---|
| 1 | hasKey is false | → Branch A: First-time Wizard |
| 2 | hasKey is true AND quickMode is null | → Branch A2: Quick Mode Setup |
| 3 | User intent is to modify config (keywords: 修改配置、设置、更改参数、配置、settings) | → Branch C: Modify Config |
| 4 | User intent is batch generation (keywords: 批量、batch) | → Branch D: Batch Mode |
| 5 | User intent is to edit an existing image (keywords: 编辑、修改图片、改一下、换背景、去掉、加上、变成、改成、edit、remove、replace — AND references an image or a previous generation exists) | → Branch F: Edit Image |
| 6 | quickMode exists AND user message contains a prompt | → Branch B: Quick Mode |
| 7 | quickMode exists AND user message has no clear prompt | → Branch E: Help |
IMPORTANT: Rules are ORDERED. Rule 3 beats Rule 6 — if the user says "修改配置", go to Branch C even though quickMode exists. Rule 4 beats Rule 6 — "批量生成猫" goes to Branch D, not Branch B. Rule 5 beats Rule 6 — "把背景换成海边" after a generation goes to Branch F, not Branch B.
Branch A: 🆕 First-time Wizard
The user has never configured the plugin. Walk them through everything.
W1: Welcome + API Key 🔑
「📋 原样输出」
👋 欢迎使用 Niu Image Gen！首次使用需要快速设置一下
整个过程只需 30 秒，之后每次 @我 + 描述就直接出图 ⚡
🔑 第一步：请提供你的 API Key
把你的 Key 粘贴给我，我会安全保存在本地，不会上传到任何地方 🔒
When the user provides their key, run:
node "$SCRIPT" --set-key <USER_KEY>
直接展示脚本输出，不要改写。 脚本会输出格式化的确认信息。
Then proceed to W2 below (same as Branch A2).
Branch A2: ⚡ Quick Mode Setup
Used when:
- Continuing from Branch A after key is saved
- hasKey is true but quickMode is null (key set manually but wizard never completed)
W2: Choose Quality 🎨
「📋 原样输出」
⚡ 第二步：设置快速模式 — 以后 @我 + 描述就按这个配置直接出图！
🎨 选择默认 画质：
| 选项 | 分辨率 | 速度 | 适合场景 |
|---|
| 1K 🚀 | ~1百万像素 | ~12s | 草稿、缩略图 |
| 2K ✨ (推荐) | ~4百万像素 | ~20s | 日常使用 |
| 4K 💎 | ~8百万像素 | ~28s | 高清大图、印刷 |
Wait for user to choose.
W3: Choose Ratio 📐
「📋 原样输出」
📐 选择默认 比例：
| 选项 | 比例 | 1K | 2K | 4K |
|---|
| ⬜ 正方形 | 1:1 | 1024×1024 | 2048×2048 | 2880×2880 |
| 🖼️ 横版 | 3:2 | 1536×1024 | 2048×1536 | 3840×2160 |
| 📱 竖版 | 2:3 | 1024×1536 | 1536×2048 | 2160×3840 |
Wait for user to choose.
W4: Choose Count 🔢
「📋 原样输出」
🔢 每次默认生成 几张？（1~4 张）
多张 = 同一描述生成不同变体，选 1 张最快 ⚡
Wait for user to choose (default 1).
W5: Save Config 💾
Run:
node "$SCRIPT" --set-quick-mode --quality <Q> --ratio <R> --count <N>
Where <Q> is 1K/2K/4K, <R> is square/landscape/portrait, <N> is 1-4.
直接展示脚本输出，不要改写。 脚本会输出完整的确认信息，包含配置摘要、使用提示和批量模式提醒。
W6: Handle original request (if applicable)
After wizard completes, check the user's original trigger message and route correctly:
- Contains batch keywords (批量/batch) → Tell user: "✅ 快速模式设置完了！你刚才提到了批量生成，需要我帮你设置一下批量模式吗？" → If yes, go to Branch D (batch setup + collect prompts).
- Contains a single image prompt (e.g., "画一只猫") → Execute it via Branch B (quick mode). The user shouldn't need to repeat themselves.
- Purely about setup (e.g., "帮我设置"/"配置一下") → Stop here, wait for next message.
Do NOT blindly route everything to Branch B. Batch intent must go to Branch D.
Branch B: ⚡ Quick Mode (Daily Use)
This is the zero-confirmation fast path. Do NOT ask the user to confirm or re-select parameters.
-
Extract the image prompt from the user's message.
-
Check: does the user's message explicitly mention a quality or ratio?
- "1K" →
--quality 1K
- "2K" →
--quality 2K
- "4K" →
--quality 4K
- "正方形" / "square" →
--ratio square
- "横版" / "landscape" →
--ratio landscape
- "竖版" / "portrait" →
--ratio portrait
- No mention → omit these flags (script uses saved quickMode config)
-
Run:
node "$SCRIPT" --prompt "<extracted prompt>" [--quality Q] [--ratio R]
Only pass --quality / --ratio if the user explicitly requested them. Otherwise the script reads the saved quickMode config automatically.
- 直接展示脚本输出，不要改写，不要用代码块（```）包裹。 脚本输出本身就是格式化好的纯文本，直接作为普通消息展示。然后提取图片路径并内嵌展示：单张时直接用 📍 后面的完整文件路径；多张变体时将 📍 后面的目录与 ①②③④ 行的文件名拼接成完整路径。展示顺序：先脚本文本，再图片，最后单独一行追加编辑提示：
✏️ 想编辑？换背景 / 改风格 / 去元素 — 说一声就行
Do NOT add a "需要调整参数吗?" prompt at the end of every generation. Keep it clean. The user knows they can say "修改配置" if needed.
Branch C: ⚙️ Modify Config
The user wants to change settings. First run --get-config if not already done this turn.
「📋 原样输出」（用 --get-config 的数据填充 [占位符]）
⚙️ 当前配置：
| 模式 | 画质 | 比例 | 参数 |
|---|
| ⚡ 快速模式 | [Q] | [R] | 每次 [N] 张 |
| 📦 批量模式 | [Q or 未设置] | [R or —] | 并发 [N or —] |
| 🔑 API Key | [preview] | | |
要修改哪个？
1️⃣ ⚡ 快速模式
2️⃣ 📦 批量模式
3️⃣ 🔑 API Key
Based on user choice:
- 1️⃣ Quick Mode: Run W2 → W3 → W4 → W5 from Branch A2. This overwrites the existing quickMode config.
- 2️⃣ Batch Mode: Run the batch setup flow from Branch D (first-time section). This overwrites or creates batchMode config.
- 3️⃣ API Key: Ask for new key, run
--set-key <NEW_KEY>. 直接展示脚本输出。
After saving, the script output already includes confirmation — do not add extra summary.
Branch D: 📦 Batch Mode
Batch mode = multiple different prompts, parallel generation.
Step 1: Check batch config
Run --get-config (if not already done). Check the batchMode field.
If batchMode is null (never configured):
「📋 原样输出」
📦 批量模式 — 先快速设置默认参数 ⚡
Then walk through:
D-W1: Quality — 「📋 原样输出」same table as W2 in Branch A2, wait for choice.
D-W2: Ratio — 「📋 原样输出」same table as W3 in Branch A2, wait for choice.
D-W3: Concurrency
「📋 原样输出」
⚡ 选择 并行数（1~10，默认 3）
数字越大越快，但可能触发 API 限流
| 并行数 | 适合场景 |
|---|
| 1~2 | 稳定优先 🛡️ |
| 3 (推荐) | 速度与稳定平衡 ⚖️ |
| 5~10 | 大批量快速出图 🚀 |
D-W4: Save
node "$SCRIPT" --set-batch-mode --quality <Q> --ratio <R> --concurrency <N>
直接展示脚本输出，不要改写。 脚本会输出完整的确认信息。
Then continue to Step 2 below.
If batchMode already configured:
Skip setup, go directly to Step 2.
Step 2: Collect prompts 📝
First check: did the user's trigger message already contain prompts?
- "批量生成 赛博猫、水墨山水、太空狗" → prompts already provided (split by 、/，/逗号/换行). Skip to Step 3 with these prompts.
- "批量生成" (no prompts) → ask user to provide:
「📋 原样输出」
📝 请提供 prompt 列表：
📌 方式一：每行一个 prompt，写完后说「开始」
📌 方式二：提供一个 JSON 文件路径（内容为 prompt 字符串数组）
Wait for user to provide prompts.
Step 3: Confirm
「📋 原样输出」（用实际值填充 [占位符]）
📦 批量生成确认：
| 项目 | 值 |
|---|
| 📝 Prompt 数量 | [N] 条 |
| 🎨 画质 | [Q] |
| 📐 比例 | [R] ([WxH]) |
| ⚡ 并发 | [C] |
| 📁 输出 | ~/Pictures/niu-image-gen/ |
✅ 确认开始？
Batch mode DOES require confirmation (unlike quick mode), because it involves multiple images and longer execution time.
If user says no or wants to adjust:
- "调整参数" / "改一下" → ask which parameter to change (quality/ratio/concurrency), update only that parameter, re-show confirmation.
- "修改 prompt" / "换一下" → go back to Step 2.
- "取消" / "不了" → abort and return to waiting state.
Step 4: Execute
Inline prompts:
node "$SCRIPT" --batch-inline "<p1>" "<p2>" "<p3>" [--quality Q --ratio R --concurrency N]
Only pass explicit flags if user overrode parameters during this session. Otherwise the script uses saved batchMode config.
File prompts:
node "$SCRIPT" --batch <file.json> [--quality Q --ratio R --concurrency N]
Step 5: Report
直接展示脚本输出，不要改写，不要用代码块（```）包裹。 脚本输出本身就是格式化好的纯文本，直接作为普通消息展示。然后从脚本输出中提取所有 .png 文件的完整路径（📁 后面的路径），将每张图片读取并内嵌展示给用户。批量模式下每张图片展示在对应的 prompt 结果下方。全部展示完毕后单独一行追加编辑提示：✏️ 想编辑其中某张？告诉我哪张和要改什么
If any failed, offer to retry the failed ones.
Branch E: 💡 Help
User @ the plugin but didn't give a prompt or a command.
「📋 原样输出」（用 --get-config 的数据填充 [Q] [R] [N]）
🎨 Niu Image Gen — 使用指南
⚡ 生成图片：@我 + 图片描述（例：一只赛博朋克风格的猫）
✏️ 编辑图片：生成后说「换个背景」「去掉XX」「改成XX风格」
📦 批量生成：跟我说「批量生成」
⚙️ 修改配置：跟我说「修改配置」
📊 当前快速模式：[Q] [R] ×[N]
Branch F: ✏️ Edit Image
User wants to modify an existing image (change background, remove object, change style, add element, etc.).
Trigger keywords: 编辑 / 修改 / 改一下 / 换背景 / 去掉 / 加上 / 变成 / 改成 / remove / replace / change / edit — AND references an image or a previous generation exists in this conversation.
Step 1: Determine image source(s)
- User specified file path(s) (e.g., "编辑 ~/Desktop/photo.png") → use directly. If multiple paths mentioned, collect all.
- User references multiple previous generations (e.g., "把刚才那几张都改成..." / "这些图全部..." / "每张都...") → collect all previous generation paths from this conversation.
- User references a specific previous generation (e.g., "改第②张" / "编辑第二张") → use that specific path.
- Last generated image exists in this conversation (from a prior Branch B/D/F run) → use that image's path. No need to ask.
- No image available → show this prompt:
「📋 原样输出」
🖼️ 你想编辑哪张图？
发送文件路径，或者把图片拖进对话。
Wait for user to provide a path, then continue.
Step 2: Extract edit prompt
Extract the edit instruction from the user's message. The edit prompt is what the user wants to change (e.g., "把背景换成海边" → prompt is "把背景换成海边").
Step 3: Determine edit mode and run
Check the collected image(s) and user intent to pick one of three modes:
Mode A — Multi-source batch edit (multiple different source images, same edit prompt):
Trigger: Step 1 collected more than one image path.
node "$SCRIPT" --edit --image "<path1>" --image "<path2>" --image "<path3>" --prompt "<edit instruction>" [--quality Q] [--ratio R] [--concurrency N]
Mode B — Multi-variation edit (one source image, multiple edit variations):
Trigger: user wants multiple versions of the same edit. Keywords: 几个版本 / 多个版本 / 出N个 / N个变体 / 多出几个 / 给我N个不同的 / N variations. Extract count from user intent (default 2, max 4).
node "$SCRIPT" --edit --image "<image_path>" --prompt "<edit instruction>" --count <N> [--quality Q] [--ratio R]
Mode C — Single edit (default, one source image, one result):
node "$SCRIPT" --edit --image "<image_path>" --prompt "<edit instruction>" [--quality Q] [--ratio R]
Only pass --quality / --ratio / --concurrency if the user explicitly requested them. Otherwise the script uses saved config.
All modes are zero-confirmation, same as Branch B. Do NOT ask "确定要编辑吗?". Just run it.
Step 4: Show result
直接展示脚本输出，不要改写，不要用代码块（```）包裹。 然后提取图片路径并内嵌展示：
- Mode C (单张编辑): 用 📍 后面的完整文件路径。展示后追加：
✏️ 继续改？直接说下一步要改什么
- Mode B (多变体编辑): 将 📍 后的目录与 ①②③④ 行的文件名拼接。展示后追加：
✏️ 继续改？直接说下一步要改什么
- Mode A (批量编辑): 将 📍 后的目录与 ①②... 行的文件名拼接（忽略 ← 后的原图名）。展示后追加：
✏️ 想继续编辑？告诉我哪张和要改什么
The user can continue editing a result by saying another edit instruction — in that case, use the NEWLY edited image as the source (not the original), and loop back to Step 2.
Parameter override rules
These rules apply to ALL branches:
- Explicit flags override saved config: if user says "4K横版画一只猫", pass
--quality 4K --ratio landscape. The script prioritizes explicit flags over saved config.
- Saved config overrides hardcoded defaults: when no explicit flags, the script reads quickMode or batchMode config. Only if those are also absent does it use hardcoded defaults (2K, square).
- Quick mode and batch mode configs are independent: changing one does not affect the other.
Error handling
| Error | Action |
|---|
| 503 "No available compatible accounts" | Wait 30s, retry once. If still fails, tell user the API is temporarily busy. |
| 400 with size error | Fall back to closest valid size and retry. |
| Timeout (generation 120s / edit 180s) | Report and offer to retry. |
| Missing API key | Guide through --set-key setup (Branch A W1). |
Hard constraints
- API base URL (
https://api.iiiiitoken.com/v1/images/generations) and model (gpt-image-2-x) are hardcoded. Never change them.
- Never display the user's full API key in chat.
- Maximum batch size: 20 prompts per run.
- Maximum concurrency: 10.
- Maximum count (quick mode / edit variations): 4.
- Maximum batch edit images: 10.
- Generation timeout: 120s. Edit timeout: 180s.
- Pixel budget: ≤8,294,400 px. Longest edge ≤3,840. Dimensions divisible by 16.
