VRGDG Music3 Caption Builder

If you've written a MiniMax Music3 caption by hand, you know the pain: one long string where you try to cram genre, mood, instruments, vocal character, tempo, song arc, and production style, and the model obeys whichever parts landed in the right order. VRGDG Music3 Caption Builder replaces that with a form. Each field maps to a caption sentence, there are 27 genre presets to start from, and whatever you leave blank comes from the preset instead of vanishing.

It's part of the VRGameDevGirl pack (vrgamegirl19/comfyui-vrgamedevgirl), the LTX 2.3 / MiniMax H3 music-video builder's companion node set. In the chain it sits downstream of the Lyrics Prepare node and upstream of Music3 Text Encode: cleaned lyrics + structured caption in, one well-formed caption string out.

How it works

Pick a preset (27 genre/style presets, from "Raw 90s Alternative Rock" to "Outlaw Country / Americana" and beyond) and every text field below fills in automatically. Then edit whatever doesn't fit your song - the preset is a starting point, not a cage. The design rule is simple: blank fields use the preset, non-blank fields override it. So you can take "Dreamy Indie Rock" and swap in your own instruments without touching the mood and tempo the preset chose.

The vocal_profile dropdown (14 options) is the shortcut for the two vocal fields: "Female Lead," "Male Baritone," "Female Rock / Rasp," "Androgynous / Neutral," and so on. Choosing one auto-fills singer_name and vocal_qualities; "Use Style Default" restores the generic vocal description from your chosen style. If you want a specific singer reference, type it into singer_name yourself - it's ordinary caption text and stays fully editable.

The fields, in order of how often you'll touch them: genre_subgenre, mood_themes, instruments, vocal_qualities, production, tempo_range, song_arc, avoid_characteristics, and extra_direction (the "start with voice and dry guitar only" slot for song-specific staging).

The one trap: the avoid field

Music3 has no separate negative prompt, and the author is up front about it in the tooltip. avoid_characteristics is appended as an "Avoid" sentence inside the caption - it's guidance, not a guarantee, so don't expect "avoid saxophone" to work like an SDXL negative embedding. This matches where the broader ecosystem has landed: on guidance-distilled and instruction-following models, the negative box is often inert or nonexistent. Use avoid_characteristics for direction ("keep it sparse, no big drums"), not for bans.

Inputs and outputs that matter

Outputs: caption - connect this straight to MiniMax Music3 Text Encode's caption input. And caption_breakdown, a human-readable audit of which fields went into the caption. That second output is your debugging tool: when a render comes back sounding nothing like what you asked, read the breakdown and see whether your override actually made it into the string.

Installing it

Same one-time install as every node in this pack. ComfyUI Manager → Install Custom Nodes → search vrgamedev, or:

cd ComfyUI/custom_nodes
git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl.git

then pip install -r requirements.txt, restart, hard-refresh. It's a heavy pack - the requirements pull in TTS and LLM wheels (voxcpm, llama-cpp-python) you won't need for caption building, and on Windows portable the README wants Cython and scikit-build-core installed first (Python 3.12 > 3.13 for compatibility).

Troubleshooting

A song that ignores your style override is usually a precedence problem - check the breakdown to confirm your field actually overrode the preset instead of sitting blank next to a preset value. A song that ignores everything is usually a text-encode problem upstream, not a caption problem. And if two presets both produce something close but not right, that's the seed/caption A/B loop: build both, wire them to the pack's Conditioning Blend, and let the blend slider find the middle.