T8 Music Creative Lab（歌词/Caption候选与软QA）

The main Music 3 node in this pack is a serious piece of machinery - it embeds MiniMax's official music-caption-rewriter skill, 18 genre indexes and 1000 templates, and it's laser-focused on producing the lyrics and music_caption pair that MiniMax Music 3 expects. T8 Music Creative Lab（歌词/Caption候选与软QA） is the lighter, more playful companion: a standalone node for generating lyric and caption candidates, surgically editing lyrics, and running a soft QA pass on the words. Each operation is exactly one LLM call, and it never touches the Music 3 node's own behavior.

The honesty framing is what makes it worth a look. There are four operations:

• 歌词候选（T8非官方） - lyric candidates (an unofficial T8 extension, clearly labeled as such, never passed off as an official MiniMax skill).
• Music Caption 候选（官方结构） - caption candidates that follow the official contract: exactly ### Global Metadata, ### Vocal Details, ### Arrangement, in that order.
• 歌词定向修改（T8非官方） - surgical lyric editing, only the lines you name, everything else kept byte-for-byte.
• 歌词可唱性软 QA - a cautious text review: language consistency, line length, rhyme regularity, hook recurrence, section balance, breath density. It explicitly refuses to claim actual singing or audio evidence.

That last point is the pack's signature move, carried over from the README and the acceptance notes: a text-based node will not pretend to hear audio, detect BPM, or judge real singability. The QA adds a local text-level check on top - line counts, repeated lines, script-scripture warnings (asking for 中文 but getting mostly Latin letters, for example) - and marks the whole thing as structure-level, not musical quality.

The inputs

• music_idea (required) - style, theme, mood, use, arrangement hopes.
• source_lyrics / source_caption - needed for directed edit and QA modes (the node refuses to edit/QA without source lyrics).
• lyrics_language - 中文 / English / 日本語 / 한국어 / Custom; pick Custom and you must fill custom_language.
• candidate_count (2 or 3), edit_request, locked_text (must stay verbatim), rewrite_mode, seed, plus api_key / provider_config.

There's a privacy nicety worth mentioning: the Caption-candidate mode won't copy your lyric text into the caption - it only passes bracketed section tags and the overall musical intent, which matches how the main Music 3 node keeps lyrics out of captions.

What comes out

selected_result (the best candidate or the QA result), candidates_json, song_titles_json, soft_qa_report_json, and version_diff (a local diff for the edit mode). From here, feed the winner into T8 Creative Version Stack and the chosen lyric+caption pair into T8 Music-to-Video Beat Sheet - that's the pack's own recommended music flow.

Installing

In the comfyui-minimax-h3-prompt-enhancer-T8 pack. 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

Restart and Ctrl+F5. The single call rides the pack's providers - Seedance NZ by default, or a T8 LLM Provider Config wired into provider_config for AI Workshop / OpenAI-compatible / local Qwen. Text-only, so any of them work; the local Qwen path won't even load a vision projector.

Gotchas

Both required-field checks are up front (no music_idea → error; edit/QA without lyrics → error), so you won't waste tokens on an impossible job. And remember: this is a text companion - the actual MiniMax Music 3 generation still happens in the pack's main Music 3 node (or a Music-01-style API elsewhere). This node just makes the words better before they get there.