Yeufonic for MacFor Macs with Apple silicon.
Yeufonic-0.0.63.dmg
The download comes from its release on GitHub,
built there by GitHub's own Mac runners, which keep a signed record of the build:
gh attestation verify Yeufonic-0.0.63.dmg --repo yeufonic/yeufonic checks the file is the one they made. What is new in each version: the releases on GitHub.
What you need
- A Mac with Apple silicon (M1 or later) and macOS 14 Sonoma or later. Intel Macs are not supported.
- Memory: 32 GB or more is recommended, since the GPU and the system share it. We have tested with 24 GB, where a cover or a long instrumental used nearly all of it. 16 GB is the least setup accepts, and we have not tried it.
- About 30 GB of free disk space, and about 16 GB more if you choose to draw take pictures on the Mac (less with the yue2.cpp engine), 4 GB for the style LoRAs and 3 GB for recorded voices.
- An internet connection for the setup, which downloads about 18 GB of models (about 11 GB with yue2.cpp).
How fast it is
Measured on a Mac mini with an M4 (10-core GPU) and 24 GB of memory, and on one with an M4 Pro (20-core GPU) and 48 GB, beside a Windows PC with an NVIDIA RTX 4070 Ti SUPER (16 GB), all with the ComfyUI engine and the same songs. Making a song is limited by memory bandwidth, which the Pro has more than twice as much of as the M4, and the NVIDIA card about two and a half times as much as the Pro.
| NVIDIA RTX 4070 Ti SUPER | M4 Pro, 48 GB | M4, 24 GB | |
| A song, for each minute of music | about 45 seconds | about 2 minutes | about 5 minutes |
| An instrumental, for each minute of music | about 70 seconds | 2 to 3 minutes | 4 to 10 minutes |
| A cover of a 90-second recording, including reading its tune and words | under 2 minutes | 4 to 5 minutes | 7 to 9 minutes |
| Stems | under a minute | under a minute | under a minute |
| A take picture drawn on the computer | under a minute | 1 to 2 minutes | 4 to 5 minutes |
| A song's words written on the computer | about 10 seconds | about 10 minutes; seconds with yue2.cpp | about 13 minutes |
| Training a LoRA (both halves, five songs) | not timed | about 1½ hours | not tried |
| Training a recorded voice | not timed | about 11 minutes | not tried |
With the yue2.cpp engine the M4 Pro is about as fast as with ComfyUI, and the NVIDIA card faster still: a song at about 25 seconds for each minute of music.
What to know
- Lyric drafts written on the Mac are slow with ComfyUI, which runs Gemma on the CPU there. With the yue2.cpp engine (below) they take seconds; or choose an external LLM in Settings. A take's picture is described from the take itself by default, so it no longer waits on Gemma.
- It is not signed by Apple, so macOS asks before opening it the first time, and again after each update.
- The GPU meter in the header shows nothing on a Mac.
Installing
- Open the disk image and drag Yeufonic into Applications.
- Open Yeufonic from Applications. The first time, macOS refuses to open it. Choose Done, then go to System Settings → Privacy & Security, scroll down, choose Open Anyway, and open Yeufonic again.
- A Terminal window opens and sets everything up: Python, the engine and about 18 GB of models. It shows its progress, and if the connection breaks, opening Yeufonic again carries on where it stopped. It asks once whether to draw take pictures on the Mac (about 16 GB more), whether to fetch the 39 style LoRAs, and whether to set up recorded voices; the answers are kept, and No is the default for each.
- When it has finished, Yeufonic opens in your browser, and its icon appears in the menu bar at the top of the screen. That icon opens the page again, shows the logs, and quits.
Choosing the engine
Yeufonic sets up the ComfyUI engine, which does everything, LoRA training included. The smaller yue2.cpp engine is the other choice: songs at about the same speed, lyric drafts in seconds, take pictures in less memory, and no LoRA training. To switch, quit Yeufonic and run in Terminal:
bash /Applications/Yeufonic.app/Contents/Resources/setup.sh --engine yue2cpp
Its YuE2 model is the near-lossless Q8 (3.8 GB); add --cpp-model bf16 for the
full-quality one (7.2 GB, twice the memory). --engine comfyui switches back, and nothing either
engine downloaded is downloaded again. The same script takes --pictures, --loras
and --voices to add those later.
Updating
Quit Yeufonic from its menu-bar icon, download the newer build, and drag it into Applications over the old one. Opening it runs the setup again, which adds what the new version needs and keeps your models and everything you made.
Removing it
Quit it from the menu-bar icon, then delete Yeufonic from Applications, and these two folders (in Finder, Go → Go to Folder):
~/Library/Application Support/Yeufonic(the models and anything you made)~/Library/Logs/Yeufonic