👩‍🦱超写实人物提示词@炮老师的小课堂

You do not have a prompt-writing problem. You have a "blank box, no idea what a real photographer would say" problem. This node - 👩🦱超写实人物提示词@炮老师的小课堂 (DapaoPortraitPhotographyPromptNode) - is a prompt director, not an image generator. Feed it a rough idea ("老城区的女店主, 雨后," whatever), and it asks a cloud LLM to write you a full structured portrait prompt plus negative prompt plus a shooting card, the way a photo director would brief a set. It never renders a pixel itself. That framing saves you most of the confusion: wire its outputs into your actual image node and stop expecting pictures here.

The twist that makes it worth the setup: it ships a built-in "生命感" (life-feel) photography skill with real scene and action pools, and it runs in three modes you pick under 🎬 运行模式. A upgrades an existing portrait (原照片) while locking identity - same face, better everything. B writes original shots from nothing, no photo needed. C retakes your photo as an "atmosphere big shot." There's also an auto mode that picks A if you fed it an image, B if you didn't. That split is the node's whole personality: preserve who someone is, or invent a fresh subject, never muddle the two.

How it works

Mechanically it's an API-wrapper node with an LLM inside - the graph-bolted-on language-model pattern, just not local. You paste a 🔑 API密钥 from api.dapaoai.com, pick a 🤖 LLM模型 from a catalogue that mixes gpt/claude/gemini/deepseek names (default gemini-3.7-flash), and the node calls the relay's /v1/chat/completions endpoint asking for one JSON object back: a shot list where every entry bundles positive_prompt, negative_prompt, and a structured shooting_card covering scene, light direction, lens choice, color, hair, wardrobe, emotion, even the one optical effect. The Chinese UI is intimidating until you realize the defaults are sane. Two settings actually matter:

• 📝 人像需求 - your brief. Write it in whatever language; vague is fine, it's a starting point.
• 🎨 目标图像模型 - this decides the prompt dialect, since "GPT Image 2", "Banana / Gemini Image", "FLUX / Stable Diffusion" and "Midjourney" all want different syntax. Same LLM behind it, different writing style.

Flip ⚙️ 展开摄影参数 and you get the fun part: per-slot dropdowns for 场景, 光线 (summer noon vs. 伦勃朗), 景别, 机位, 色彩关系, 发型, 服装. Leave it off and the model picks. 🔢 方案数量 lets one LLM call plan up to 12 shots for a series; 🎯 选中方案 says which one lands on the first two outputs, while the list outputs always carry everything.

Inputs, outputs, and wiring

Optional image inputs: 🖼️ 原照片 (batch, auto-downscaled to a 2048px PNG before upload) and 🎨 风格参考图 - which only borrows photography language, never the face or composition. Two tooltips worth trusting: the 🎲 随机种 only controls ComfyUI's cache, it is not sent to the LLM, and 🚫 出错时跳过 keeps a bad item in a list/batch run from killing the whole queue.

Nine STRING outputs, but you live in the first two: 📝 选中人物提示词 and 🚫 选中负面提示词. Run those into a CLIP Text Encode for local FLUX/SD, or into the prompt field of an image-API node. 📋 完整拍摄方案JSON is the whole shot list if you're scripting; ✅ 提示词质检报告 is the model checking its own homework.

Install

Same story as the rest of this pack: ComfyUI Manager → search dapaoAPI → install → restart. Or manually:

cd ComfyUI/custom_nodes/
git clone https://github.com/paolaoshi/ComfyUI-dapaoAPI
cd ComfyUI-dapaoAPI
pip install -r requirements.txt   # requests, Pillow; torch/numpy come with ComfyUI

No models to download - inference happens on the author's relay. You do need an account: register at api.dapaoai.com, buy a redemption code in the wallet, and apply for an API key in the default group, then paste it into the node.

Where people get burned

• It returns only text. If nothing appears, you forgot to wire the prompt output into an image node. The most common complaint, and it's a feature.
• Every run is a paid LLM call. Plan-count of 12 is one chunky request; still, test with 1 before batching.
• Chinese-only labels. Set 🌐 输出语言 to English (recommended) for prompts that behave in your downstream.
• It's a third-party relay, so your prompt and reference photos leave your machine by design - the standing caveat for every API-wrapper node. Use a key you can afford to see leak: this is arbitrary custom-node Python holding a credential, the exact shape the ecosystem got burned by once. Read it before you trust it.