FL HeartMuLa Conditioning

If the Model Loader is the door, FL HeartMuLa Conditioning is where you write the song. This is the node you'll spend your actual creative time in: you paste the lyrics, drop in style tags, and it turns all that text into the tensors the Sampler will later consume. It's the closest thing this pack has to a prompt box.

How it works

The node takes your lyrics and tags, lowercases them, and tokenizes them into HeartMuLa's text tokenizer. Tags get wrapped in special <tag> / </tag> markers, and both halves get BOS/EOS tokens tacked on. There's also a placeholder MuQ embedding slot - it's wired for future reference-audio support, so you can ignore it today.

The interesting bit is how CFG works here. HeartMuLa uses classifier-free guidance, so when your cfg_scale isn't 1.0 the conditioning tensors get duplicated into a batch of two - one guided pass, one unguided - and the Sampler blends them. Set it to 1.0 and you get single-batch "no CFG" mode, which is faster but noticeably less obedient to your tags.

The inputs that matter

• lyrics - multiline text with section markers. These aren't cosmetic; the model actually uses them to structure the song. The supported markers are [Intro], [Verse], [Prechorus], [Chorus], [Bridge], [Outro], and [Instrumental]. Blank-line your verses and keep each section short - the model sings what you write, and community testing confirms it nails lyric-following far better than style-following.
• tags - comma-separated style tags like pop, female vocal, energetic. Genre, vocal type, mood, tempo, instruments - all free text, all passed through. This is also the input you wire FL HeartMuLa Tags Builder into if you want dropdowns instead of typing.
• cfg_scale - default 1.5, range 1–10. This is the one knob people actually fight over.

Where people get burned

Tag adherence is this model's weak spot, full stop. Real-world reports across the HeartMuLa threads agree: it follows lyrics brilliantly but treats genre tags as suggestions. The community sweet spot that emerged is cranking cfg_scale to ~3 - the README's default of 1.5 is too timid, and people who never touched it concluded the model "ignores tags entirely." Also, don't expect metal. A bunch of us tried; it wants to make pop.

Output is a single conditioning output (HEARTMULA_COND) that wires straight into FL HeartMuLa Sampler.

Installation

Same pack as the rest of FL HeartMuLa:

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

Or use ComfyUI Manager → search "FL HeartMuLa" → Install → restart. This node needs a loaded model from FL HeartMuLa Model Loader (first-run download is ~6 GB into ComfyUI/models/heartmula/), and it doesn't work without one.

Troubleshooting

• Songs ignore your style - raise cfg_scale toward 3 and reroll; rerolling is expected here, so batch several runs.
• Output sounds unguided - you probably left cfg at 1.0, which disables guidance entirely.
• Weird timing on markers - keep sections short and formatted exactly like the default example. The model reads structure from those bracket labels.

One tip: wire the lyrics input from a text node if you're iterating on many song ideas, so you don't keep reaching into the widget. This node is where the music actually gets written - the rest of the graph just renders it.