YuE2-hum-to-song

YuE2-hum-to-song

Turns a hummed melody of 10 to 30 seconds, plus a style line and lyrics, into a full song that follows your tune. An add-on for the YuE2-3B music model, for songwriters and audio researchers.

Made by Mothersuperior 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: YuE2-hum-to-song takes a hum of 10 to 30 seconds, a line naming the style, and some lyrics, and returns a complete song built on your tune. It is built on m-a-p/YuE2-3B and adds two mechanisms to it. One continues a score written from your hum, and the other is a trained adapter that carries the pitch and timing of your hum into the decoder.

Who it suits: It suits people who start songs by humming and want a full draft, and researchers who want to see how a hum can steer a two-stage music model.

Industry | Who uses it | How they use it | Example
Music | Songwriters | Turn a hummed idea into a full song draft | A hum of about 20 seconds becomes a song of 2 to 3 minutes with an ending
Creative tools | ComfyUI users | Run the hum-to-song flow inside ComfyUI | A hum, a style line and lyrics fed into the workflow
Research | Audio machine learning researchers | Study score continuation and the prosody adapter | Compare the three melody modes on one hum

If you need a model for commercial work, you can skip it, because the license is non-commercial.

Where to use it: The card says you run it with the scripts in its repository or inside ComfyUI. Its files download to Windows, macOS or Linux with Python.

• The scripts that come with the model files
• ComfyUI
• Windows, macOS or Linux with the huggingface_hub package

Use case | Who | What they get back
Hum to full song | Songwriters | A song that keeps your melody and adds new sections and an ending
Hum only | Musicians | A song that is exactly your melody and nothing more
Ignore the hum | Researchers | The planner's own score, with the adapter still shaping phrasing

How to set it up: You need Python, the YuE2-3B base model that this builds on, and a way to run the card's scripts or ComfyUI. The card does not list hardware needs in the part shown here.

• Python on Windows, macOS or Linux
• The Hugging Face hub package
• The base model m-a-p/YuE2-3B

1. Install the Hugging Face hub package.
2. Download the files for this model.
3. Follow the model card to run the scripts, or load the files in ComfyUI.

pip install -U huggingface_hub
hf download Mothersuperior/YuE2-hum-to-song

How to use it: Record a hum of 10 to 30 seconds, write a line naming the style plus lyrics, choose one of three melody modes, then run the card's scripts or the ComfyUI setup. The rest of the song is written around your tune.

Tool | What it does
Continue from hum | Writes new sections, transitions and an ending after your melody, in the same key and meter
Hum only | Ends the score straight after your hum, which gives a song made of your melody alone
Ignore hum | Lets the planner write its own score while the adapter can still shape phrasing

Example: You hum a melody for about 20 seconds, add a style line and a few lyrics, and choose continue from hum. You get back a song of 2 to 3 minutes. In the authors' tests the hummed phrase came back two to four times, usually as the hook.

For newcomers and experts: If you are new to this, think of it as a way to start a song from your voice. A first small step is to download the files with the two commands above, then read the model card for how to run the scripts.

If you know the field, YuE2 works in two stages. An AR planner writes an ABC score and then semantic tokens, and an NAR decoder turns them into VAE latents by flow matching before the VAE decodes 48 kHz audio. The melody is decided in the planner, so the hum enters twice. A score prefix with no end token lets the planner keep writing. A prosody adapter then feeds the decoder a sine tone that follows your pitch and timing, so the decoder never sees your timbre or words.

Hum is transcribed into an ABC score
Planner continues the score
Planner writes semantic tokens
Decoder with prosody adapter adds your pitch
VAE decodes 48 kHz audio

Good to know: The adapter is a rank-64 LoRA plus a few projections, and the combined release merges it with the authors' real-audio decoder LoRA into one rank-96 LoRA. The first mechanism needs no training.

• License: cc-by-nc-4.0, which allows non-commercial use only
• Cost to run: no API key, account or paid service is named, and the card shows no hardware needs in the part shown
• Data it reaches: the download comes from Hugging Face, and the model works on the hum audio, style line and lyrics you give it
• Not for: commercial use, since the license is non-commercial

On Hugging Face, 13 November 2026

• Downloads in the last 30 days: 0
• Likes: 47
• License: cc-by-nc-4.0
• Based on: m-a-p/YuE2-3B
• Tags: audio, music, yue2, lora, hum-to-song

Links

• Model: https://huggingface.co/Mothersuperior/YuE2-hum-to-song
• Files: https://huggingface.co/Mothersuperior/YuE2-hum-to-song/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Mothersuperior/YuE2-hum-to-song

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