HeartMuLa Studio

What it is

A Suno-shaped studio for HeartLib — the MuQ, MuLan and HeartCodec stack. Full songs with vocals, instrumental mode, style tags for genre and mood and tempo, seed control when you want the same take twice, and a queue that chews through a batch while you write the next prompt.

The reference-audio part

This is the piece I care about. You drop in a track, the waveform renders, and you drag a ten-second region over the part that has the feel you want. That region conditions the generation. It turns a vague prompt into a decision you make with your ears.

Lyrics come from whatever LLM you have: Ollama locally, or OpenRouter if you would rather spend a few cents than VRAM.

Getting it running

Clone and run start.sh, or pull the container. Images are published to Docker Hub as ambsd/heartmula-studio:latest and to GHCR as ghcr.io/fspecii/heartmula-studio:latest, so docker compose up is the whole install.

Ten gigabytes of VRAM is the floor, twelve is comfortable. That was the design constraint the whole way through — it had to run on a card people already own.