MuLaCover

Inputs

Lyrics use section markers and preserved line breaks:

[Verse]
One lyric line
Another lyric line

[Chorus]
One chorus line

Style is one UTF-8 line using named fields:

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

Download

Download the model weights with the Hugging Face CLI. Authentication is not required:

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

Also download HeartCodec and Qwen3-Embedding-0.6B into sibling checkpoint directories. Reference-audio conditioning also needs the documented YourMT3 and ChordNet checkpoints. See the code repository for the complete installation and generation commands.

Usage

Install the mulacover package and prepare all checkpoint directories using the generation guide. Run the following from the code repository root. ./ckpt is the parent directory containing MuLaCover, HeartCodec, Qwen3, and the transcription models.

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",
)

The pipeline writes a WAV file. Use the MuLaCover package API shown above; this model is not a built-in Transformers model class.

Files

• config.json: MuLaCover model architecture configuration.
• gen_config.json: audio/text token identifiers used during generation.
• model-00001-of-00005.safetensors … model-00005-of-00005.safetensors: sharded model weights.
• model.safetensors.index.json: shard index.
• tokenizer.json: lyrics and prompt tokenizer.
• SHA256SUMS: release integrity checksums.

License and use restrictions

The official model weights are licensed under CC BY-NC 4.0 with the additional terms in MODEL_LICENSE. Commercial use of the official weights is prohibited without separate written authorization from MuLa Labs.

Outputs generated using the official weights are restricted to noncommercial use unless separately authorized in writing by MuLa Labs. The Apache-2.0 license covering the source-code repository does not grant commercial rights to these weights or their generated outputs.

Users are responsible for obtaining the rights required for their lyrics, reference audio, performances, voices, and generated works. Do not use the model to impersonate a person deceptively or violate copyright, privacy, or publicity rights.

Community

community support. Reproducible software defects should be reported through the code repository's GitHub Issues.

Join MuLa Labs, Vera Praxis Lab

We are always excited to meet people with a strong interest in audio and music. MuLa Labs has internship openings for candidates who want to build the next generation of music and audio technology. To apply, email your resume or CV, together with a short introduction, to contact@mulalabs.ai.