🪂全能image提示词生成@炮老师的小课堂

Some image models are picky about prompts, and GPT Image 2 is the pickiest of the lot - it wants structured design intent, not a paragraph of vibes. DapaoLocalAllroundImagePromptNode ("全能image提示词生成") is a local LLM that writes, optimizes, and diagnoses prompts in that GPT Image 2 idiom. The node's own description is refreshingly honest: it calls the LLM, builds from a structured design template, and does not generate an image. It's a prompt-authoring node that lives upstream of whatever model you actually use.

The origin story matters for how you read it. This is one of the seven "大炮API常用工具" (dapaoAPI common tools) ported to local inference: the original node called an OpenAI-compatible API, and the local version swaps that client for a llama.cpp payload running your GGUF. Same inputs, same template library, same structured validation - but the request goes to your GPU instead of a cloud endpoint. No key, no per-call cost, uncensored (the usual local-LLM pitch from the KB).

What it can do

The 🎛️ 任务模式 dropdown is the heart of it - seven modes from auto-detect through 新建图像提示词, 优化现有提示词, 参考图创作提示词, 图像编辑提示词, 蒙版局部编辑提示词, and 提示词诊断与重写. Then 🗂️ 设计分类 (UI界面样机, 产品商业视觉, 地图与路线, 海报与Campaign, 人物与角色设定 …) and a huge 🧩 具体模板 menu - 94 templates covering everything from 学术论文配图 to 图标游戏与素材. And 🎯 目标图像模型 lets you aim the prompt style at GPT Image 2, 通用Image, Banana/Gemini, Midjourney, FLUX, or Stable Diffusion - the prompt changes register, not just the label.

Inputs that matter

• 📝 原始图像需求 - your idea, multiline. The one creative input you must fill.
• 📝 需要优化的现有提示词 - for optimize/diagnose tasks.
• 🔤 画面准确文字与数据 - "必须逐字保留的标题、价格、数据、标签或公式." This is the GPT-Image-2-specific superpower: exact text in the final image, guaranteed by the prompt contract.
• 🔒 必须保留 and 🚫 禁止出现 - the guardrails.
• 🎭 蒙版 - MASK input for the local-edit mode (white areas = editable).
• 🖼️ 参考图1 through 🖼️ 参考图9 - source images for the LLM to analyze.
• Sampling: 🌡️ 温度 (default 0.35 - deliberately low for structure), 📝 最大输出令牌 (6144 default), Top_P, seed, timeout.

Outputs are the full diagnostic stack: 🪂 最终图像提示词 (the one that matters), 🗂️ 识别分类, 🧩 使用模板, 📑 参数与优化分析, 📄 LLM完整响应, and ℹ️ 处理信息.

Where people get burned

The low temperature is the clue: this node wants determinism, and if you crank the temperature up you'll get the same prompt written five different ways - pointless for an image that needs a specific layout. Keep it in the 0.3–0.5 range. And remember it's a prompt writer, not a generator: people wire the final prompt straight into a GPT Image 2 node and then complain "nothing came out" because they skipped the generator. The description calls this out for a reason.

Installing it

Standard pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/paolaoshi/ComfyUI-llama_Dapao

then requirements + restart. The pack-wide dependency gotcha: llama-cpp-python pinned to the JamePeng fork at 0.3.47+ (stock PyPI 0.3.35 fails on Qwen3.8 GGUFs), NVIDIA users overlay a GPU wheel from the JamePeng releases, models in ComfyUI/models/LLM/. For image analysis you need a vision model plus its mmproj; for pure text-prompt work, a text GGUF suffices.