Docker with NVIDIA support and a CUDA 12.9.1-compatible driver. Qualified on RTX 5070 Ti 16 GB; other GPUs are not qualified. Download the two GGUF models separately.
Local song generation Development snapshot · rolling latest
YuE2
Your lyrics. Your musical direction. Your own GPU.
Generate songs locally from lyrics and musical direction, with stereo 48 kHz output, editable ABC scores, a browser library, replayable takes, and native HTTP jobs.
Download the models first, start Docker, then open Create. Enter musical direction and lyrics and select Generate. Library stores songs in this browser; export the tracks you want to keep.
Generation runs locally after image/model downloads. Models use CC BY-NC 4.0, separately from MIT application code. Use lyrics you have rights to process. Browser storage is not a backup.
What is included
A complete local workspace, not just a model.
- Style tags and lyrics to stereo 48 kHz songs
- MP3 and WAV playback and download
- Editable ABC scores, tokens, and latent replay
- Browser library, favorites, and prompt reuse
- Native asynchronous HTTP jobs and cancellation
- Local LoRA adapters and GPU telemetry
The actual application
Compose, replay, and keep your own takes.
Create brings lyrics, musical direction, scores, and generation settings together. Library preserves songs in browser storage; API and System expose jobs, models, GPU activity, and logs.
