HeartMuLa

HeartMuLa Teams · 2026

An LLM-based song generation model capable of synthesizing high-fidelity music under rich, user-controllable conditions. HeartMuLa supports multi-conditional music creation with inputs including textual style descriptions, lyrics, and reference audio.

Application types: Text-to-songLyrics-to-songStyle transfer
Architecture:Autoregressive transformer

Essential categories

CategoryStatusNotes
Source code~ PartialGitHub repository includes model architecture, and inference. But lacks specific code for traning or fine-tuning the model, and data processing.
Training data~ PartialTechnical report indicates that HeartMuLa was trainied on ~100,000 hours of high-quality music and describes automated filtering pipelines. However, there is no information about the specific music used.
Model weights✔︎ OpenModel weights are available on Hugging Face.
Code documentation✔︎ OpenAll components of the codebase are properly documented in the GitHub repository, including installation requirements, examples of scripts usage and configuration files.
Training procedure✔︎ OpenTraining procedure is well documented in the technical report, including hardware requirements and model configurations (see Section 2.2).
Evaluation procedure✔︎ OpenThe evaluation procedure is well documented in the technical report, including used metrics and test benchmacks.
Research paper~ PartialA technical report is available on arXiv, but the paper is not yet peer-reviewed. The report provides a detailed description of the model architecture, training procedure, evaluation and inference.
Licensing✔︎ OpenApache License 2.0.

Desirable categories

CategoryStatusNotes
Model card⭐ IncludedModel card available in Hugging Face, although it could benefit from more detailed information about the model.
Datasheet∅ Not included
Package∅ Not included
User-oriented application∅ Not includedAccording to the GitHub repository, there is a model demo available in Hugging Face Spaces and in Model Scope. However, neither of the two are currently operational.
Supplementary material page⭐ IncludedThere is a complementary website for the HeartMuLa model, which includes multiple audible examples.

Raw YAML file with complete evaluation.