MuLaCover by HeartMuLa

MuLaCover by HeartMuLa

Makes cover songs and music remixes from a reference track, lyrics and a style prompt, or from melody and chord MIDI. Writes a WAV file. For noncommercial music projects on a CUDA machine.

Made by HeartMuLa on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: MuLaCover is a controllable model for cover songs and remixes. It reshapes a reference piece of music using lyrics you supply and a structured text style prompt. It can also generate from melody and chord MIDI.

The model has about 4.4 billion parameters. The pipeline writes its result as a WAV file. Its weights are public and need no access request.

Who it suits: Musicians, hobbyists and audio developers who want to try AI cover generation and who can set up several checkpoints on a GPU. Commercial use is not allowed.

Industry | Who uses it | How they use it | Example
Music | Hobbyist musicians | Remake a song in a new genre or mood | A country-style version of a reference track
Audio research | Researchers | Study controllable music generation | Comparing outputs for different style prompts
Software | Audio developers | Build a noncommercial music tool | A prototype that turns lyrics and a reference into a cover
Education | Students | Explore how lyrics and style control a generated track | Changing one mood field and listening to the result

Anyone who needs to use the music commercially can skip it, since the weights and their outputs are limited to noncommercial use.

Where to use it:

• Python on a machine with a CUDA GPU, as in the card's example.
• Windows, macOS or Linux with Python, for downloading the files.
• The MuLaCover code repository, which holds the full install and generation steps.

Use case | Who | What they get back
Cover generation | Musicians | A WAV file of the reference reworked with your lyrics
Style remixing | Producers | A remix shaped by topic, genre, instrument and mood
MIDI generation | Composers | Music generated from melody and chord MIDI

How to set it up: You need Python, the mulacover package and several checkpoints.

• Python with the Hugging Face tool
• A CUDA GPU, since the example runs on cuda:0
• The mulacover package, installed as the code repository describes
• HeartCodec and Qwen3-Embedding-0.6B checkpoints in sibling folders
• The YourMT3 and ChordNet checkpoints, if you condition on reference audio

1. Download the MuLaCover weights.
2. Download HeartCodec and Qwen3-Embedding-0.6B beside it.
3. Add the YourMT3 and ChordNet checkpoints for reference audio.
4. Install the mulacover package and follow the generation guide.

hf download HeartMuLa/MuLaCover --local-dir ./ckpt/MuLaCover

How to use it: Write the lyrics with section markers and keep line breaks. Write the style as one UTF-8 line of named fields.

[Verse]
One lyric line
Another lyric line

[Chorus]
One chorus line

topic:[Longing]; genre:[country]; instrument:[Strings,acoustic guitar]; mood:[hopeful]

Then run the pipeline from the code repository root, where ./ckpt holds all checkpoint folders.

import torch
from mulacover import MuLaCoverGenPipeline

torch.manual_seed(42)
pipe = MuLaCoverGenPipeline.from_pretrained(
    "./ckpt",
    device=torch.device("cuda:0"),
    dtype={
        "mulacover": torch.bfloat16,
        "codec": torch.float32,
        "qwen": torch.float32,
        "transcriptor": torch.float32,
    },
    lazy_load=True,
)
pipe(
    {
        "ref_audio": "/path/to/reference.mp3",
        "lyrics": "assets/lyrics.txt",
        "tags": "assets/tags.txt",
    },
    save_path="cover.wav",
)

Tool | What it does
MuLaCoverGenPipeline | Loads the model and its helper checkpoints and generates the audio
ref_audio | Points to the reference track to rework
lyrics and tags | Point to your lyrics file and your style line
save_path | Names the WAV file the result is written to

Example: Take a reference track as an MP3, write a verse and a chorus in a lyrics file, and set the style to a hopeful country mood with acoustic guitar. Run the pipeline. It writes cover.wav, a reworked version of the reference with your lyrics.

For newcomers and experts: If you are new, think of it as a tool that sings your lyrics over the shape of an existing song in a style you describe. Start by downloading the weights and reading the generation guide, because several checkpoints must be in place.

For experts, it is not a built-in Transformers class, so you use the package API. The pipeline takes per-component dtypes, with the main model in bfloat16 and the codec, Qwen embedder and transcriptor in float32. It can load components lazily to save memory.

Prepare lyrics with section markers
Write a one-line style prompt
Provide a reference track
The pipeline generates with the checkpoints
A WAV file is saved

Good to know: The points below cover limits and costs.

• Rights: you must hold the rights to the lyrics, reference audio, performances and voices you use. Do not use it to impersonate someone deceptively or to violate copyright, privacy or publicity rights.
• License: CC BY-NC 4.0 with extra terms. Commercial use of the weights and their outputs needs written authorization from MuLa Labs.
• Cost to run: no API key or paid service is named. A CUDA GPU is needed.
• Data it reaches: it reads your reference audio and lyrics, and needs several checkpoints downloaded.
• Not for: commercial use without written permission, or deceptive impersonation.

On Hugging Face, 13 November 2026

• Task: audio to audio
• Library: mulacover
• Parameters: 4.4B
• Weights: about 8.7 GB in BF16; about 2.2 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 678
• Likes: 49
• License: cc-by-nc-4.0
• Tags: mulacover, safetensors, music, audio, cover, cover-song-generation, music-remix, remix-generation, text-to-music, symbolic-music, audio-to-audio

Links

• Model: https://huggingface.co/HeartMuLa/MuLaCover
• Files: https://huggingface.co/HeartMuLa/MuLaCover/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download HeartMuLa/MuLaCover

Then follow the usage in its model card. Nothing in this kit runs by itself.