HeartMuLa
HeartMuLa is a dual-stage generator: an autoregressive music model (MuLa) and a codec flow solver, built for prompt-plus-tag conditioning.
What it does in Miso
Section titled “What it does in Miso”One task, generate.heartmula, on the create page, using the guided prompt builder. Miso turns your chips into the comma-separated tag list that this family wants.
Vocals
Section titled “Vocals”Multilingual singing, conditioned on the lyrics you provide.
Options
Section titled “Options”| Option | Default | Does |
|---|---|---|
text |
High level description of the song | |
tags |
Comma-separated descriptors. The runtime wraps these in tag tokens | |
lyrics |
Lyric text matching the intended structure | |
duration_sec |
120 | Target length in seconds |
infinite_mode |
Bounded chunking for long songs | |
codec_guidance_scale |
1.25 | Codec CFG scale |
guidance_scale |
1.5 | MuLa generator CFG scale |
Prompting
Section titled “Prompting”Keep the description direct and put the detail in the tags.
text: upbeat summer anthemtags: electropop, bright synths, four on the floor, female vocals, energetic, 128bpmTempo goes in the tags for this family rather than in a separate field. Miso handles that translation from the guided form, so 128bpm in the builder lands in the tag list rather than as a bpm option.
Memory
Section titled “Memory”Miso sends heartmula.mem_saver=true, which releases resident graphs between token generation and codec decoding.
The default duration_sec of 120 is high compared to the other families. If you are on a smaller card, bring it down before trying anything else.