Overview
Train custom LoRA adapters to capture specific vocal styles, genres, instrument sounds, or production aesthetics. Requires 3-10 songs as training data.
Dataset Preparation
- Collect 3-10 songs in the target style (WAV/FLAC preferred, MP3 OK)
- Place in a directory:
~/Music/lora-datasets/<style_name>/ - Songs should be 30-300 seconds each
- Consistent style/genre across the dataset
- High audio quality (no noise, no clipping)
Training
cd "$(python3 -c "import json; print(json.load(open('$HOME/.claude/skills/claude-music/config.json'))['ace_step_dir'])")"
uv run python3 -m acestep.training.train_lora
--checkpoint-dir ./checkpoints
--model-variant turbo
--dataset-dir ~/Music/lora-datasets/my_style/
--output-dir ./lora_output/my_style
--rank 16
--learning-rate 1e-4
--steps 1000
uv run python3 -m acestep.training.train_lora
--checkpoint-dir ./checkpoints
--model-variant turbo
--dataset-dir ~/Music/lora-datasets/my_style/
--output-dir ./lora_output/my_style
--method lokr
--rank 16
--learning-rate 1e-4
--steps 500
## LoRA vs LoKr
| Aspect | LoRA | LoKr |
|--------|------|------|
| Training time | ~1 hour | ~12 min |
| Quality | Higher fidelity | Good, slightly less detailed |
| VRAM | ~10GB | ~8GB |
| Use case | Voice cloning, precise style | Genre adaptation, quick experiments |
## Hyperparameters
| Parameter | Default | Range | Notes |
|-----------|---------|-------|-------|
| Rank (r) | 16 | 4-64 | Higher = more capacity, more VRAM |
| Learning rate | 1e-4 | 1e-5 to 5e-4 | Lower for voice cloning |
| Steps | 1000 | 200-5000 | More data = more steps needed |
| Batch size | 1 | 1-4 | Limited by VRAM |
## Using Trained LoRA
After training, the LoRA is available in ACE-Step's generation pipeline. Refer to ACE-Step documentation at:
`<ace_step_dir>/docs/en/LoRA_Training_Tutorial.md` (see config.json for path)
For detailed reference: load `references/lora-training.md`