Claude Music Lora

claude-music-lora — LoRA Fine-Tuning

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

1. Collect 3-10 songs in the target style (WAV/FLAC preferred, MP3 OK)
2. Place in a directory: ~/Music/lora-datasets/<style_name>/
3. Songs should be 30-300 seconds each
4. Consistent style/genre across the dataset
5. 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'])")"

# LoRA training (standard, ~1 hour on RTX 5070 Ti)
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

# LoKr training (5x faster, ~12 min)
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

Hyperparameters

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