NotaGen
ElectricAlexis/NotaGenReviewed by AI
Checked by RepoPicks Radar on
What is NotaGen?
NotaGen is a research model that composes classical sheet music. You give it a period, a composer's style and the instruments, and it writes the score in ABC notation, which can be turned into MusicXML, PDF or audio. The quickest way to try it is the makers' page on Hugging Face, where you type the prompt and listen. Running it on your own computer means Python with conda, PyTorch with CUDA, downloading the model file by hand and about 8 GB of graphics memory. The code also includes everything used to train and fine-tune it on your own music.
NotaGen is free and open source (MIT) and can replace Suno and Udio.
How easy is it to install NotaGen?
A · download and open · Web
- Windows D
- Linux D
- Web A
On the web it is an A: the makers' own page on Hugging Face Space runs NotaGen-X, you type "Period-Composer-Instrumentation" and "preview the audio / pdf scores, and download them", with nothing installed. On your own Windows or Linux machine it is a D, with no release to download: `conda create --name notagen python=3.10`, PyTorch 2.3.0 with CUDA 11.8, four `pip install` steps, the model file fetched by hand from Hugging Face and roughly 8 GB of graphics memory, following a second set of instructions in `gradio/README.md`. Macs are not covered, since the install line asks for an Nvidia card.
NotaGen on GitHub
- Stars
- 1,246on Nov 10, 2026
- Downloads
- —all releases
- Growth
- —no gain in the last reading
Stars over time
Facts
- Runs on
- Windows, Linux, Web
- You need to install
- Nothing extra
- To build from source
- Python
- Disk space
- 150 MB–450 MB · estimate
- Accounts or keys
- None
- Cost
- Free · MIT code, model files published openly on Hugging Face, and the online demo on Hugging Face Space asks for nothing.
- License
- MIT (Permissive)
- Last change on GitHub
- Apr 21, 2025
NotaGen risk signals (8)
- AttentionDownload links to file-sharing sites
README links to https://drive.google.com/drive/folders/1iVLkcywzXGcHFodce9nDQyEmK4UDmBtY?usp=sharing
automatic check
- AttentionDependencies with known vulnerabilities
Pinned dependencies with known vulnerabilities (OSV): transformers: GHSA-29pf-2h5f-8g72, GHSA-37mw-44qp-f5jm, GHSA-4w7r-h757-3r74; gradio: GHSA-39mp-8hj3-5c49, GHSA-5cpq-9538-jm2j, GHSA-6655-8ph2-63j3
automatic check
- Attentionthe model file comes from elsewhere
Nothing that makes music is in the repository: the README links five `.pth` files on Hugging Face (`weights_notagen_pretrain...`, `weights_notagenx_p_size_16...`) plus two more from a different account for CLaMP 2 (`huggingface.co/sander-wood/clamp2`). PyTorch `.pth` files are loaded by unpacking saved Python objects, which is a way of running code, so what you download decides what happens on your machine. The checks also flag a dataset link on Google Drive.
AI review
- Attentionwhere the music it learned from came from is not said
The overview states "Pre-training on 1.6M musical pieces" and "Fine-tuning on ~9K classical compositions" without naming the source or its licence; only the small example dataset is credited (Schubert lieder from OpenScore Lieder). The project's own filter shows the model can repeat what it learned: in `RL/data.py`, "Sheets that may plagiarize from ground truth data (ld_sim>0.95) are excluded". Who owns a score it writes is an open question.
AI review
- Infoa strong graphics card is the real requirement
The README: "Deploying NotaGen-X inference locally may require 8GB of GPU memory", and for training "The use of NotaGen-large requires at least 24GB of GPU memory for training and inference", with the smaller models offered as the way out. The install line pins `pytorch==2.3.0 pytorch-cuda=11.8`, so it is an Nvidia card or nothing.
AI review
- Infotraining reports to an online service
`requirements.txt` pins `wandb==0.17.2`, the Weights & Biases service, which the pre-training, fine-tuning and reinforcement-learning scripts use. If you train with your own music, the run's data goes to that service unless you turn it off.
AI review
- Infoold pinned libraries and no commits for a year and a half
The last commit is from 2025-04-21, 563 days before this check, with 26 issues open and no releases. `requirements.txt` pins `transformers==4.40.0`, `gradio==5.17.1` and `numpy==1.26.4` — the versions the vulnerability check flagged — so a current Python environment may not reproduce it.
AI review
- Inforesearch code configured by editing files
There is no interface for setting anything up: the instructions tell you to edit `pretrain/config.py`, `finetune/config.py`, `inference/config.py`, `RL/config.py`, `clamp2/extract_clamp2.py` and `RL/data.py`, pasting long weight file names such as `weights_notagen_schubert_p_size_16_p_length_1024_p_layers_20_c_layers_6_h_size_1280_lr_1e-05_batch_1.pth` into each. The handiest versions — a Colab notebook and a ComfyUI node — are other people's contributions, kept in their own repositories.
AI review
Evidence is shown in its original language.
What Radar could not check about NotaGen
- Whether the install still works, with PyTorch 2.3.0, CUDA 11.8 and the pinned libraries
- How long one score takes on an 8 GB card, and how much disk the model files take
- Whether the music is pleasant or useful, which only listening can tell
- Who owns the scores it writes, which depends on training data the repository does not describe
- What the online demo on Hugging Face keeps of the prompts you type