HeartMuLa Tag Builder

ComfyUI has a hundred "tag builder" nodes and ninety-nine of them are for image prompts. HeartMuLa Tag Builder is the one for music - and it exists because the plain tags box on the Music Generator is a leaky way to steer a model that, frankly, already has genre-tag obedience problems. This node keeps your tags structured and clean so you're not debugging "why does it sound like nothing I asked for" when the actual culprit is a typo'd, duplicated, mixed-case tag string.

It's one of the genuinely new pieces in this fork - the pack's own creation, with the concept inspired by the RT-HeartMuLa pack but reimplemented from scratch. The idea is simple: instead of one free-text box, you get categories, because a song prompt is really several dimensions and humans forget the boring ones.

What it does

Seven category fields feed one comma-joined tag string:

• genre, mood, instrument, vocal, production - free-text fields, each with a tooltip full of suggestions (pop/rock/house/techno, uplifting/melancholic/dark, piano/808 bass, and so on). Type anything comma-separated; the suggestions are hints, not a fixed vocabulary.
• tempo and era - single-pick dropdowns (very slow → driving, 1960s → futuristic), default none.
• additional_tags (optional, multiline) - the catch-all for anything the fields above don't cover.

Output is a single tags STRING that goes straight into the Music Generator's tags input. The important behavior is what it does to your text: splits on commas, lowercases, strips whitespace, de-duplicates (first occurrence wins), and drops blanks and none. So a messy "Synthwave, synthwave, 1980s" becomes the clean synthwave, 1980s the model can actually work with.

Why the cleaning matters

HeartMuLa's tag conditioning is literal and comma-delimited. Duplicates and inconsistent casing don't crash anything, but they waste conditioning capacity on noise - and with a model whose tag adherence is already the weak spot (the community's #1 complaint at launch), you want every token earning its keep. The de-dup and lowercase normalization is the difference between "I typed a mess" and "I gave the model a clean spec." One light aside: this won't magically fix the model ignoring your tags - that's a cfg_scale problem, covered in the Music Generator article - but it removes the "my tags were garbage" variable from the debugging session.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Crono141/ComfyUI-JK-HeartMuLa.git
cd ComfyUI-JK-HeartMuLa
pip install -r requirements.txt

Restart ComfyUI and it appears under JK-HeartMuLa. It's a pure string-processing node - no models, no VRAM, no GPU. You can use it on CPU-only rigs and it'll be instant.

Where it fits

In the default workflow it sits at the front, feeding the Music Generator. The nice property is that it's a pure function of text, so you can drop it into any graph and even use it as a general music-tag normalizer independent of HeartMuLa. If you're generating multiple tracks and want consistent output, keep the same tags across runs and change the seed - the Tag Builder doesn't randomize anything, which is exactly what you want from a prompt builder.