🪂全能image提示词生成@炮老师的小课堂

The display name says "全能image提示词生成" and the node description adds a line that saves you five minutes of confusion: 本节点不出图 - this node does not output images. DapaoAllroundImagePromptNode is a prompt writer. It calls a dapaoAI LLM, applies a structured design template built for GPT Image 2-era models, and hands you back a finished prompt plus the analysis behind it. If you've spent an afternoon fighting a text-rendering image model with a rambling English paragraph, you already know why this exists: these models reward a structured, template-shaped prompt, and this node is the translation layer.

What you feed it

The required side is where the value lives:

• 🔑 API密钥 and 🤖 LLM模型 - dapaoAI key, and the same 11-model menu as the rest of the pack (default gemini-3.7-flash).
• 🎛️ 任务模式 - seven modes: auto-detect, new image prompt, optimize existing prompt, create from reference images, image edit prompt, masked local edit, and diagnose & rewrite. Auto-detect guesses from context.
• 🗂️ 设计分类 (18 categories: UI mockups, product commercial visual, maps & routes, posters & campaigns, character design, split comics, grids…) and 🧩 具体模板 (94 named templates like "学术论文配图｜图形摘要" or "图标游戏与素材｜复古拟物图标集").
• 🎯 目标图像模型 - so the prompt's shape matches the destination: GPT Image 2 / OpenAI-compatible (default), generic Image, Banana / Gemini Image, Midjourney, FLUX, Stable Diffusion.
• 📝 原始图像需求 - your rough idea in plain words.
• Plus ratio, prompt format (structured JSON / structured natural language / terse), detail density, temperature (0.35 default), max tokens (6144), Top_P, a cache-only seed, and timeout.

The optional side is where it gets clever: 📝 需要优化的现有提示词 for optimize/diagnose jobs, 🔤 画面准确文字与数据 for text that must survive verbatim (titles, prices, labels - the thing GPT Image is best and worst at), 🔒 必须保留 and 🚫 禁止出现 constraint boxes, a 🎭 蒙版 (MASK) for local-edit prompts where white = editable, and up to 🖼️ 参考图1–9 fed to the LLM for analysis.

What comes out

Six outputs, but really you care about two: 🪂 最终图像提示词 (wire this into your actual image node) and 📑 参数与优化分析 (why it made the choices it did - genuinely useful when the result misses). The others - 🗂️ 识别分类, 🧩 使用模板, 📄 LLM完整响应, ℹ️ 处理信息 - are breadcrumbs for debugging and for building the same prompt deterministically later.

The honest take

This is a "prompt enhancer" in the classic LLM-in-the-graph sense that llm-in-comfyui.md covers, and the usual caveats apply: the LLM can add detail you didn't ask for, and it can't see inside the image model's head. The template system is the real differentiator - 94 structured templates is a lot of scaffolding for prompts that would otherwise come out as generic prose. Where people get burned: the seed only controls ComfyUI caching, so re-running with a new seed doesn't change the prompt - change a setting or flip "每次重新生成提示词"-style controls if the node caches. And the 随机种 here explicitly says "仅控制ComfyUI缓存，不发送给接口" - that's not a bug.

Install

ComfyUI Manager → "dapaoAPI", or:

cd ComfyUI/custom_nodes
git clone https://github.com/paolaoshi/ComfyUI-dapaoAPI.git
cd ComfyUI-dapaoAPI
pip install -r requirements.txt

Restart, get a dapaoAI key (register at api.dapaoai.com, buy a redeem code, apply for a key in the default group), paste it in, and pipe the final-prompt output into the image node of your choice. Chinese-only UI, consistent with the pack.