MiniMax Music 3 Prompt & Lyrics Enhancer (T8)

MiniMax Music 3 doesn't want a vibe description. It wants a three-part caption - Global Metadata, Vocal Details, Arrangement - written to an official spec, plus your lyrics fed to a separate input. Get that contract wrong and the track comes back generic, or with the wrong sections. This node does the paperwork: it takes one line of musical intent and produces the lyrics, the official caption, and a ready-to-paste JSON payload, in a single pass.

It's the third node in T8mars' pack that shares the "cloud or local LLM" plumbing with the H3 and Seedance enhancers, but it's deliberately narrower. It has no IMAGE, VIDEO or AUDIO inputs - it reads text, thinks about text, and outputs text. It does not generate audio, it does not listen to reference tracks, and if you run it locally it loads the text GGUF only, never the vision projector. That's a feature, not a limitation: the whole job is prompt preparation.

How it works

The node embeds MiniMax's official music-caption-rewriter skill, frozen at a fixed commit - 18 genre indexes and 1000 templates that live in the repo and get hash-verified before any paid request. The quality_mode decides how deep it goes: "official full" runs a router that picks at most 2 genre indexes and at most 3 templates, disclosed in stages (so it never dumps the whole library into one request), which typically means 2–4 LLM calls. The "fast core" mode skips the template layer and just executes the core three-section contract in 1–2 calls.

Lyrics handling is where the node earns its keep. lyrics_mode offers AUTO, generate new, strictly preserve, polish-to-spec, or instrumental. "Strictly preserve" passes your existing lyrics through verbatim, locally, with zero LLM touching them. Polish mode can target a single section - "only fix verse 2" - and if it can't confirm what you meant it stops before you're billed, rather than guessing and charging you. And by default the caption stage never even sees your lyrics text: semantic_profile_mode defaults to "privacy isolation," so the node sends mood, narrative intensity and energy arc instead of your actual lines. The one light aside I'll allow myself: that's more privacy discipline than most music tools shipping today.

The inputs and outputs that matter

You fill music_idea (the only required field - genre, theme, mood, instruments, intent) and pick lyrics_mode. target_duration_seconds, fixed_bpm, meter and structure_preset are soft constraints, all optional; Music 3 doesn't guarantee exact timing anyway. Wire existing lyrics in via the lyrics STRING input if you have them.

The four outputs are the whole story:

• lyrics → Music 3's input
• music_caption → Music 3's instructions
• music3_payload_json → a ready-made {"input": lyrics, "instructions": music_caption} if you're scripting
• enhancement_report_json → stage timings, request counts, cache hits and warning codes, deliberately stripped of your lyrics, key and template content

Install

Same pack as the H3 enhancer - Manager search "MiniMax H3 / Seedance 2.0 / Music 3 Prompt Enhancer (T8)", or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-prompt-enhancer-T8.git

Restart, Ctrl+F5. Cloud mode needs nothing extra; local mode runs the same python install_local_qwen.py installer and wants a hefty GPU for reasonable speed (24 GB recommended).

Where people get burned

• The API key is for the caption/lyrics LLM (the author's seedance.nz or t8star gateways, or your own OpenAI-compatible endpoint), not for the Music 3 audio generation API.
• "Generate new lyrics" and "polish" are T8's own extensions, clearly labeled unofficial - don't expect the official skill to rewrite your words; the strict-preserve mode is the one that guarantees verbatim.
• The 10-minute stage cache is in-memory only and keyed per credential - it saves you from re-paying for upstream stages after a crash, but it dies with the ComfyUI process.
• Instrumental mode sets [Instrumental] and tells the caption not to introduce a singer; if your result still has vocals, that's upstream, not the node lying about it.