🎵Music3音乐提示词生成@炮老师的小课堂

MiniMax Music 3 expects a structured music description - genre, tempo, emotion arc, instrumentation, vocal style - and if you've ever stared at that Text Encode node with nothing to say, this node is the fix. DapaoLocalMusic3CaptionPromptNode ("Music3音乐提示词生成") writes the whole thing locally: a structured music description plus complete lyrics, with no API call. The first two outputs (🎼 音乐描述 and 📝 歌词) are designed to wire directly into the official Music3 Text Encode node.

It's another of the pack's seven API-tool ports: original was a dapaoAPI node, the local version keeps the prompt assembly and swaps the API client for a llama.cpp payload on your GGUF. Same output contract, zero cloud.

What you can specify

The required inputs are a music producer's checklist, almost all defaulting to 自动识别 so you can also just type a sentence and let the model decide:

• 📝 原始音乐需求 - your idea. The default is a perfectly good example: "温暖的原声流行歌曲，亲密女声，副歌逐步扩张并在结尾温柔收束."
• 🎼 主风格 and 🧬 融合风格 - 20 style slots from 东亚现代流行 through 现代R&B, 电影感抒情, 电子合成器氛围, 爵士摇摆, 金属重型摇滚, 国风民乐…
• 🎯 使用场景 - 流媒体完整歌曲, 影视剧情/片尾曲, 电影或游戏配乐, 品牌广告, 短视频背景音乐, 现场舞台…
• ⏳ 目标曲长, 💫 情绪弧 (克制铺陈→温暖释放→余韵收束 and friends), ⏱️ 速度 (60–180 BPM bands), 🎹 调性倾向, 🎚️ 拍号/律动, 🥁 核心律动, 🎙️ 人声配置, 📈 人声音域, 🗣️ 人声音色, 🎤 演唱方式, 👥 和声/伴唱, 🎻 核心乐器编制, 🧱 歌曲结构, 🎛️ 制作质感, 🌌 空间与混响, 🔥 编曲密度, 🚫 排除项.
• Output language: English (Music 3 recommended), Chinese, or bilingual. 📏 输出详略 (180–250 / 250–450 / 450–650 words).

The lyrics logic

The optional 📝 歌词 / 🔗 外部歌词 inputs are where it gets clever: provide lyrics and the node keeps them verbatim - no rewriting, no "improving." Leave them empty and the LLM writes lyrics from scratch. The author's framing ("有歌词时原样保留") is exactly right for a model that's sensitive to word changes. ➕ 补充约束 adds extra guardrails.

Inputs and outputs in practice

The eight outputs mirror the Music 3 contract: 🎼 音乐描述, 📝 歌词 (both → official Text Encode), plus 🌐 全局音乐信息, 🎙️ 人声细节, 🧱 编曲结构, 📑 音乐结构分析, 📄 语言模型完整响应, and ℹ️ 处理信息. The 📦 输出格式 control exists only for old-workflow compatibility - the current version always outputs the wire-ready description and lyrics, per its tooltip.

Where people get burned

The token budget. Music descriptions with full lyrics hit the default 📝 最大输出令牌 of 4096 quickly; if the lyrics get truncated mid-song, raise it. And note 🎲 随机种 is only a ComfyUI cache control, not sent to the LLM - if a re-roll comes back identical, flip 每次重新生成 style behavior by clearing cache or changing the seed on the model side, not this one.

Installing it

Standard pack install:

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

then requirements + restart. Pack-wide 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/. A text GGUF is enough here unless you're feeding reference images.