SD HeartMuLa Tags Builder

The most honest way to describe SD HeartMuLa Tags Builder: it's a string formatter with nice dropdowns. It exists because the Conditioning node's tags input is just a comma-separated string, and nothing stops you from typing that string yourself. What this node adds is not power - it's that you never mistype "r&b" and you never forget the comma, because everything is picked from a list and joined for you.

Mechanically that's the whole story. The node assembles a tag string in a fixed order - primary genre, optional secondary genre, vocal type, mood, tempo, then whatever instruments and extra tags you type - and outputs one comma-separated string. It lowercases your free-text entries along the way. There's no model involved, no tokenization, no network call. It's the cheapest node in the pack in every sense, and it's the only one you could fully replace by hand-typing.

The dropdowns that matter:

• genre - 24 options, from pop and rock through jazz, hip-hop, r&b, metal, ambient, cinematic, edm. Set your primary here.
• vocal_type - female vocal, male vocal, duet, choir, instrumental, vocal harmony, rap, spoken word.
• mood - 16 of them, energetic through melancholic, aggressive, romantic, dreamy, dark, epic.
• tempo - slow, medium, fast, very slow, very fast (default medium).
• secondary_genre - none by default; pick a second genre for fusion styles like "indie rock."
• instruments - a free-text box, e.g. piano, guitar, drums, synth.
• additional_tags - a catch-all comma-separated box for anything not covered.

The one output, tags (STRING), is designed to plug straight into the Conditioning node's tags input - that's the entire purpose of the thing. Wire it up and you never have to remember the exact spelling of "melancholic" again.

Now the honest part, and it's the same caveat as everywhere else in this pack: the tags you build here are suggestions, and the 3B HeartMuLa model is not a great listener. Community testing after the January 2026 release found the model frequently ignored genre tags entirely at the default settings - people described each generation as a gacha roll on whether you'd get the style you asked for. The fixes that actually worked were raising cfg_scale on the Conditioning node (people landed around 4–6) and dropping temperature on the Sampler (0.5 was the common number). Once you've done that, this node's dropdowns become meaningful - before that, "indie rock, melancholic" can come back as ambient pop with a straight face.

My take: use it. It's small, it removes a real class of typos, and when you come back to a workflow weeks later the dropdowns read like documentation of what you intended. Just don't expect it - or the model - to enforce your taste. Treat the tags as a starting negotiation, and let the conditioning and sampler settings do the actual persuading.