🎵Music3音乐提示词生成@炮老师的小课堂

Music generation models are like image models circa 2023: the prompt matters enormously and most people write it badly. DapaoMusic3CaptionPromptNode is the dapaoAI pack's answer - it calls an LLM to turn your one-line music idea into a structured MiniMax Music 3 description plus a full set of lyrics, with a frankly absurd number of musical knobs so the output actually matches what you meant.

The knobs (yes, there are a lot)

The required inputs read like a producer's checklist: 🎼 主风格 (20 options: pop, rock, electronic, folk, cinematic…), 🧬 融合风格, 🎯 使用场景, ⏳ 目标曲长, 💫 情绪弧 (the emotional arc - build, release, resolve), ⏱️ 速度 (BPM-ish), 🎹 调性倾向 (key), 🎚️ 拍号/律动 (time signature), 🥁 核心律动 (groove), 🎙️ 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 (17 options), 🧱 歌曲结构, 🎛️ 制作质感, 🌌 空间与混响, 🔥 编曲密度, and 🚫 排除项. Almost everything defaults to 自动识别 (auto), so you can just type a description and let it fill the blanks - or get surgical when you know exactly what you want.

The output-side controls: 🌐 输出语言 (default English, which the tooltip recommends for Music 3), 📏 输出详略 (default "standard 250–450 words"), 📦 输出格式 (structured text default). Plus the standard temperature (0.35), max tokens (4096), Top_P, cache-only seed, and timeout.

Every one of those enum fields has a matching ✍️ 自定义… string input, so when the preset list doesn't have your specific timbre you can type it.

The lyric handling - the part people ask about

📝 歌词 (and its 🔗 外部歌词 twin) has a clean contract: if you provide lyrics, the node preserves them exactly - it won't rewrite your words, it builds the musical description around them. If you don't, it composes original lyrics in the requested language. That's the right behavior for a tool like this: you keep creative control of the words, the LLM handles the sonic scaffolding.

Outputs

The first two are the money outputs - 🎼 音乐描述 and 📝 歌词 - and the tooltip says they can be wired directly into MiniMax's official Music3 Text Encode node, which is the whole point: this is a front-end for the official pipeline, not a standalone. The rest are structure and diagnostics: 🌐 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, ℹ️ 处理信息.

The honest take

Twenty-three musical parameter dropdowns is overkill for "I want a chill lo-fi beat," and the defaults exist precisely so you can ignore most of them. Where this node earns its keep is the specific brief - a client wants "a warm acoustic pop song, intimate female vocal, chorus that builds then resolves softly" and you need a Music 3 prompt that won't come back as elevator music. That's the LLM-as-translator pattern from llm-in-comfyui.md applied to audio, and it works the same way: the model is a translation layer, not a composer, and it can add flavor you didn't ask for. Scan the description before generating if exactness matters.

Gotchas

• Paid calls, no auto-retry. A truncated response means re-running and paying again - check 最大输出令牌 if outputs seem short for the "high detail" setting.
• Seed is cache-only, never sent to the API.
• Lyrics provided = preserved verbatim; leave 歌词 empty to get fresh writing. Both inputs together (📝 歌词 + 🔗 外部歌词) is redundant - use one.
• Chinese-only labels; defaults are sane.

Install: ComfyUI Manager → "dapaoAPI", or git clone https://github.com/paolaoshi/ComfyUI-dapaoAPI.git into ComfyUI/custom_nodes/, pip install -r requirements.txt, restart. dapaoAI key from api.dapaoai.com (register → wallet → redeem code → apply key in the default group).