🎨全能视觉风格提示词@炮老师的小课堂

Name a task and this node has a field for it. DapaoVisualStylePromptNode ("🎨全能视觉风格提示词@炮老师的小课堂") is the pack's prompt-engineering brain: it does not generate an image, it generates the prompt - and it generates a lot of prompt. Thirty-nine visual style cards, a creation-task selector, a use-scene selector, a target-model selector, and a full bank of advanced visual controls (composition, camera/lens, lighting, color, mood, text strategy, what to preserve, what to avoid). It pipes all of that through an LLM (dapaoAI relay, eleven models, default gemini-3.7-flash) and outputs a finished image prompt plus a machine-readable "Creative Spec." Then you wire that text into whichever generator node - Seedream, FLUX, SD, Midjourney, Banana - matches your target.

What it's actually for

The killer feature set is the 🎯 目标图像模型 dropdown: 通用图像模型 / GPT Image 2 / Banana / Gemini Image / Midjourney / FLUX / Stable Diffusion. Same brief, model-appropriate prompt out. That's the promise of LLM-assisted prompting done right - the KB's llm-in-comfyui essay calls this exact pattern (a second language model writing instructions for the first) the architecture-native way to prompt in 2026.

The practical inputs:

• 📝 原始视觉需求 - the only required text. The tooltip gives the smart workflow: to optimize an existing prompt, paste it here and set 创作任务 to 优化现有提示词.
• 🎨 创作任务 - 8 modes (auto-detect, from-scratch, optimize existing, reference-image subject keep, style transfer, edit/局部修改, product-ad visual, custom).
• 🧩 视觉风格卡 - 39 cards across domains (产品广告/人像摄影/插画叙事/海报封面/科幻奇幻…), or 自动选择 and let the LLM pick a card or declare none fits.
• 🔍 真实提示词联网检索 - here's the interesting one: turn it on and the node queries a bundled read-only database of real prompts (adapted from the NanmiCoder/open-image-prompts project) and validates source-image URLs online. First use downloads ~84.6MB, and it does not download the image pack. Off by default because it's a privacy/bandwidth choice.

The 🚀 展开更多视觉控制 toggle (default on) reveals the deep controls: 构图/镜头/光线/配色/氛围/画面文字/核心保留/避错/风格改造强度/细节密度. It's a lot of knobs. You do not need all of them - set 自动判断 on the ones you don't care about and let the LLM decide.

Outputs

Seven strings: 🎨 最终图像提示词, 🇨🇳 中文提示词, 🇬🇧 英文提示词, 🧬 Creative Spec视觉规范 (the structured spec for API tools), 🔍 检索参考报告, 🏆 使用风格卡, ℹ️ 处理信息. For most workflows you want the final prompt (or EN/CN depending on your target model).

Install and when to use it

Pack-standard: Manager → dapaoAPI, or clone + pip install -r requirements.txt + restart. Then a dapaoAI key for the LLM calls.

Honest verdict: this is a power-user tool with a learning curve hidden under all those dropdowns. If you mostly write prompts fine by hand, it's overkill - a plain LLM node can do 80% of this with one instruction. Where it earns its keep: consistent brand/commercial output, translating one brief across five target models, and the retrieval mode for finding real-prompt grounding instead of hallucinating style. The main trap is cost creep - every run is an LLM call (plus a retrieval call if enabled), so iterate on your workflow before you iterate on the prompt count.