Get the right YuE2 files.

Match the download to your workflow

YuE2-Vae-legacy is for reproducing the published benchmark protocol. Use the current listening VAE for a normal first run. A separate MERT2 download is not required for standard generation.

Install the official Python pipeline

The documented starting point is Linux, Python 3.12 and a BF16-capable NVIDIA GPU with 24 GB VRAM. With those prerequisites available, the repository’s quick start is:

git clone https://github.com/multimodal-art-projection/YuE.git
cd YuE
python3.12 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install .
python examples/generate.py --output outputs/first-song

The first run downloads model files from Hugging Face. Allow time for that download before judging generation speed. These commands follow the authors’ instructions; we have not run an inference benchmark on this site’s server.

Check more than the audio file

Open outputs/first-song/audio.flac. Keep the rest of the output directory too: it records the score, settings and intermediate artifacts needed for later edits and comparisons.

Inspect result.json and its truncation flags. A playable file can still be an unfinished generation that reached a token limit. Save each changed request in a new output directory.

Keep versions together

Pin the repository and model revisions when you want a repeatable experiment. Changing the runtime, decoder or sampling settings can change the result even with the same seed.

The original YuE lives on the repository’s YuE-v1 branch. Its instructions and license should not be substituted for YuE2’s. For a visual interface, continue with the ComfyUI guide.