ACE-Step: AI Music Generation

ACE-Step: AI Music Generation

Generate music tracks locally using ACE-Step 1.5. Runs on local your GPU.

When to use

Trigger on: "сгенерируй музыку", "сделай трек", "music generation", "ace-step", "напиши песню", "сгенерируй аудио", "background music", "soundtrack", "jingle"

Quick start

# Simple instrumental
cd ~/your-ace-step && uv run python ~/.claude/skills/ace-step/scripts/generate.py "epic cinematic orchestral" --instrumental --duration 60

# Song with auto-generated lyrics
cd ~/your-ace-step && uv run python ~/.claude/skills/ace-step/scripts/generate.py "upbeat pop song about coding" --generate-lyrics --duration 120

# Song with custom lyrics
cd ~/your-ace-step && uv run python ~/.claude/skills/ace-step/scripts/generate.py "indie folk acoustic" --lyrics "[Verse 1]\nWalking down the road..." --duration 90

# Use XL model for better quality (slower, uses CPU offload)
cd ~/your-ace-step && uv run python ~/.claude/skills/ace-step/scripts/generate.py "jazz piano trio" --model acestep-v15-xl-turbo --instrumental --duration 60

# Fast draft with no LM thinking
cd ~/your-ace-step && uv run python ~/.claude/skills/ace-step/scripts/generate.py "techno beat" --no-thinking --instrumental --duration 30

Parameters

Parameter	Default	Description
caption	required	Music style/description prompt
--lyrics	""	Song lyrics (with [Verse], [Chorus] tags)
--duration	120	Length in seconds (10-600)
--instrumental	false	No vocals
--bpm	auto	Beats per minute
--key	auto	Key/scale (e.g. "C major", "A minor")
--language	en	Vocal language: en, zh, ja, ko, ru, etc
--model	auto	DiT model (acestep-v15-turbo, acestep-v15-sft, acestep-v15-xl-turbo, acestep-v15-xl-sft)
--seed	-1	Random seed for reproducibility
--steps	auto	Inference steps (more = better quality)
--no-thinking	false	Skip LM (faster but lower quality)
--generate-lyrics	false	Auto-generate lyrics from caption
--config	""	Use existing TOML config file
--json	false	Output as JSON (for programmatic use)

Models available

Model	Quality	Speed (16GB)	When to use
acestep-v15-turbo	Good	~5-10s	Quick drafts, iteration
acestep-v15-sft	Better	~10-20s	Default, balanced
acestep-v15-xl-turbo	Great	~20-40s	Final renders (CPU offload)
acestep-v15-xl-sft	Best	~30-60s	Production quality (CPU offload)

Output

• Files saved to ~/Music/ace-step/
• Format: WAV (lossless)
• First run downloads model weights (~15-25 GB total from HuggingFace)

Lyrics format

[Verse 1]
First verse lyrics here
Second line of verse

[Chorus]
Chorus lyrics here

[Verse 2]
Second verse

[Bridge]
Bridge section

[Outro]
Final words

Caption prompt tips

Be specific about:

• Genre: "melodic death metal", "lo-fi hip hop", "orchestral film score"
• Instruments: "acoustic guitar, soft piano, ambient synths"
• Mood: "melancholic, introspective, building to triumphant"
• Vocals: "powerful female soprano", "raspy male baritone", "whispered"
• Tempo: "slow ballad", "high-energy 140 BPM"
• Reference: "in the style of Hans Zimmer film scores"

Gradio UI (alternative)

cd ~/your-ace-step && uv run acestep
# Opens http://127.0.0.1:7860

REST API (alternative)

cd ~/your-ace-step && uv run acestep-api
# REST API on http://127.0.0.1:8001

Architecture

• Location: ~/your-ace-step/
• Runtime: Python 3.12 via uv (isolated venv)
• GPU: your GPU, CPU offload (configure as needed) for XL models
• Tier: tier6a (16-20GB config)
• LM models: 0.6B, 1.7B (for lyrics gen and thinking)