The first fully functional, Apple Silicon port of Tencent's SongGeneration.
100% Local. No Cloud. No NVIDIA GPU Required.
This is a specialized fork of SongGeneration Studio, engineered to run on Macs with any M series Apple Silicon chips.
The original model requires NVIDIA GPUs and CUDA dependencies to run. We have patched the entire dependency chain to translate CUDA instructions to Metal (MPS), optimized memory usage to prevent crashes, and created a universal installer that works on any Apple Silicon Mac.
- Metal (MPS) Support: Runs completely on your Mac's GPU via MPS on pytorch.
- Smart Memory Management: Dynamically detects your RAM (16GB vs 64GB) to prevent crashes.
- Base & Large Model Support: Choose between speed (>=6 mins for ~2 min song) or maximum quality (>=12 mins for ~2 min song).
- Universal Compatibility: Tested on M1 Max 64gb, Mac OS 15.7.2 Sequoia. Compatible with all M series chips.
| Component | Bare Minimum (Base Model) | Recommended |
|---|---|---|
| Chip | M1 / M2 / M3 / M4 / M5 (Pro) | M1 / M2 / M3 / M4 (Max/Ultra) |
| RAM | 24 GB | 32 GB+ (64GB Preferred) |
| OS | macOS 14 (Sonoma) | macOS 15+ (Sequoia) |
| Storage | 15 GB Free | 70 GB Free (due to Swap usage) |
| Generation Time | ~4-6 minutes | ~10-25 minutes |
❌ 16GB RAM Users: The "Large" model is not supported and will crash your system. You can attempt the "Base" model, but expect heavy slowness/system freeze.
**
⚠️ Due to the model architecture(and inefficiency of apple mps on pytorch), the total app ram usage(including swap) hits around 70gb during inference(even for the fastest base model), so expect heavy swap usage for devices with <32gb memory. Large model goes around 80gb usage...
This app is designed to be installed via Pinokio, the AI browser.
- Download Pinokio: pinokio.computer
- Copy this Repository URL: https://github.com/Rdx-ai-art/SongGen-Mac.git
- Paste into Pinokio: Open Pinokio, go to Discover > Download from URL > paste the URL into the respective field , and select Download.
- Click Install:
- The installer will automatically detect your Mac chip.
- It will fetch the initial model weights (~15GB). (the main song models, base or large, are to be downloaded later from the UI)
- It will patch all Python libraries (
librosa,torchcodec,flash-attn) for Mac compatibility.
That's it! No terminal commands needed. ✅
When you launch the app, check the "Model" dropdown:
- SongGeneration-Base(fastest): Best for 32GB Macs or quick drafts (~6 mins for ~2 mins of song).
- SongGeneration-Large: Best for 48GB+ Macs and final production (~12 mins for ~2:30 mins of song). Higher memory usage.
- Lyrics: Write your own or use an LLM to generate lyrics with structure (Verse, Chorus, Bridge).
- Style: Select "Female" or "Male" voice, and pick a Genre.
- Generate: Click Create and wait.
- Note: The progress bar may pause at 10% or 90% while the model loads/unloads. This is normal.
You can upload an MP3/WAV to clone its vibe (15 secs max, will be cut after upload).
⚠️ Important: The Reference Audio acts as the "Boss." If you upload a ** low pitch piano instrumental**, the AI might ignore your "Female Voice" setting and give you a low Male voice to match the piano's pitch.- 💡 The Fix: If this happens, add specific tags to the start of the Custom Style box, like :
"Female Vocals, High Pitch, Woman Singing"
- Heavy Swap Usage (Disk Space Warning):
- The "Large" model is massive. During generation, you may see 30GB - 40GB of Swap usage, even on 64GB Macs(totals around 80gb+ ram+swap during inference for me).
- Ensure you have at least 60GB of free disk space before running the Large model, or the OS may freeze.
- System Sluggishness:
- While generating, the AI uses nearly 100% of your system resources. It is normal for other apps to lag.
- Pre-Chorus is bugged:
- Pre-chorus block produces gibberish and buggy vocals, especially in the first added block... so maybe avoid it, if its being buggy for your lyrics.
- Original Model: Tencent AI Lab (SongGeneration)
- Studio Interface and original code: BazedFrog
- Apple Silicon Port: [Rdx-ai-art]
- Pinokio: pinokio.computer
This launcher is provided for personal, educational, and research purposes. See the original SongGeneration repository for model licensing details.