Media MusicGen

Do NOT use this skill for:

• MusicGen / AudioCraft — EXCLUDED, CC-BY-NC weights despite MIT code
• Suno / Udio — commercial paid APIs, not open-source
• Lyria / MusicLM — Google proprietary, not open-weights
• Research-only models (Jukebox, MuseNet, etc.)

Step 1 — Pick a model

Full license audit (including excluded models and Stable Audio's $1M ARR cap) in references/LICENSES.md.

Step 2 — Install

# Riffusion (MIT, diffusers-based)
pip install diffusers transformers torch accelerate soundfile

# YuE (Apache 2.0) — HuggingFace transformers based
pip install transformers torch soundfile
# YuE models: m-a-p/YuE-s1-7B-anneal-en-cot (English, chain-of-thought)
#             m-a-p/YuE-s1-7B-anneal-zh-cot (Mandarin)

# Stable Audio Open (Stability Community License)
pip install stable-audio-tools soundfile
# (HuggingFace: stabilityai/stable-audio-open-1.0)

Or let the script help:

uv run ${CLAUDE_SKILL_DIR}/scripts/musicgen.py install riffusion
uv run ${CLAUDE_SKILL_DIR}/scripts/musicgen.py install yue
uv run ${CLAUDE_SKILL_DIR}/scripts/musicgen.py install stable-audio-open

Check what's available:

uv run ${CLAUDE_SKILL_DIR}/scripts/musicgen.py check

Step 3 — Generate

# Riffusion — fast, MIT, spectrogram diffusion
uv run ${CLAUDE_SKILL_DIR}/scripts/musicgen.py generate \
    --model riffusion \
    --prompt "dreamy lo-fi hip-hop beat with mellow piano" \
    --duration 10 \
    --out beat.wav

# YuE — full song with vocals
uv run ${CLAUDE_SKILL_DIR}/scripts/musicgen.py generate \
    --model yue \
    --prompt "emotional indie rock, male vocalist, 80 bpm, acoustic guitar, simple drums" \
    --duration 90 \
    --out song.wav

# Stable Audio Open — 44.1 kHz stereo, instrumental
uv run ${CLAUDE_SKILL_DIR}/scripts/musicgen.py generate \
    --model stable-audio-open \
    --prompt "ambient cinematic strings, slow rise, film score" \
    --duration 47 \
    --out cue.wav

Per-model limits:

• Riffusion native window is 5-second spectrograms. Longer durations (>5s) are produced by repeated inference + crossfade; the script handles this automatically.
• YuE accepts "chain-of-thought" prompts with sections (Verse 1 / Chorus / Bridge). Prompt engineering matters more than other models.
• Stable Audio Open caps individual generations at ~47 seconds.

Step 4 — Continue an existing audio idea

Feed an existing clip as seed; the model continues in style.

uv run ${CLAUDE_SKILL_DIR}/scripts/musicgen.py continue \
    --in sketch.wav \
    --model stable-audio-open \
    --prompt "add energy, build into chorus" \
    --duration 20 \
    --out extended.wav

Only Stable Audio Open and Riffusion support continuation reliably; YuE's continuation API is less stable.

Step 5 — Split output into stems (vocals / drums / bass / other)

Delegates to the media-demucs skill.

uv run ${CLAUDE_SKILL_DIR}/scripts/musicgen.py stems \
    --in song.wav \
    --out-dir stems/
# -> stems/vocals.wav, stems/drums.wav, stems/bass.wav, stems/other.wav

Step 6 — Post-process / mux

• Loudness normalization for streaming: hand off to media-ffmpeg-normalize (Spotify -14 LUFS, YouTube -14 LUFS, broadcast -23 LUFS).
• Export MP3 / Opus / FLAC: ffmpeg -i song.wav -c:a libmp3lame -b:a 320k song.mp3
• Use as video background: ffmpeg -i video.mp4 -i music.wav -map 0:v -map 1:a -c:v copy -shortest out.mp4
• Loop seamlessly: see ffmpeg-audio-filter skill (afade + concat demuxer).

Gotchas

• License boundary with Stable Audio Open: the Stability AI Community License permits commercial use only up to $1M in cumulative revenue. Above that you need a commercial agreement with Stability. Riffusion and YuE have no such cap. Flag this early to the user. See references/LICENSES.md.
• MusicGen / AudioCraft / Suno / Udio are EXCLUDED from this skill. MusicGen's code is MIT but weights are CC-BY-NC. Do not recommend or install them here.
• Prompt engineering dominates output quality. Bad results are almost always prompt problems. Try: genre + instrumentation + tempo + mood + production style (e.g. "80 bpm lo-fi hip-hop beat with mellow rhodes piano and vinyl crackle, chill sunset vibe").
• Riffusion generates via spectrograms. Under the hood it's Stable Diffusion v1.5 fine-tuned to produce mel-spectrograms, then Griffin-Lim or WaveNet-style decoder. This is why 5 s is the native window.
• YuE produces vocals but expects structured prompts. Its best results are with section labels: "[Verse 1]\n...lyrics...\n[Chorus]\n...lyrics...". See the YuE HuggingFace model card.
• Stable Audio Open output is 44.1 kHz stereo FP32. Save directly; no resampling needed for most mastering workflows.
• GPU is practically required for YuE (7B params). Riffusion runs on CPU slowly (~2 min for 5 s); Stable Audio Open needs GPU for any reasonable speed.
• Durations are approximate. All three models generate fixed-length latents; actual output length can be ±10% of the requested duration.
• Seeds: pass --seed N for reproducibility. Default is a random seed each run.
• Model downloads are 2-15 GB each. First run caches into ~/.cache/huggingface/.
• Stereo vs mono: Riffusion is mono-only. YuE is mono. Stable Audio Open is stereo. Upmix to stereo in ffmpeg if needed: -ac 2.
• No copyright laundering: the user is responsible for confirming generated music doesn't accidentally reproduce a copyrighted melody. Highly specific artist-imitation prompts raise the risk — use genre/mood descriptors instead of artist names.
• Avoid --chain-of-thought in prompts for non-YuE models. Prompts that reference "Verse 1 / Chorus" etc. confuse Riffusion and Stable Audio Open.

Examples

Example 1: 10-second background music for a video intro

uv run ${CLAUDE_SKILL_DIR}/scripts/musicgen.py generate \
    --model riffusion \
    --prompt "upbeat corporate bright synth pad, inspirational" \
    --duration 10 \
    --out intro_music.wav

ffmpeg -i intro.mp4 -i intro_music.wav \
    -map 0:v -map 1:a -c:v copy -shortest intro_scored.mp4

Example 2: Full 90-second indie rock song with vocals via YuE

uv run ${CLAUDE_SKILL_DIR}/scripts/musicgen.py generate \
    --model yue \
    --prompt "[Verse 1] acoustic guitar picking, soft male vocals about a road trip. [Chorus] full band, big drums, uplifting harmonies, 80 bpm. [Verse 2] same as verse 1 with harmony vocal." \
    --duration 90 \
    --seed 42 \
    --out indie_song.wav

Example 3: Stable Audio Open cinematic cue + stem split for DAW mix

# Generate
uv run ${CLAUDE_SKILL_DIR}/scripts/musicgen.py generate \
    --model stable-audio-open \
    --prompt "ambient cinematic strings, slow build, film score, Johann Johannsson style" \
    --duration 45 \
    --out cue.wav

# Split stems for further work
uv run ${CLAUDE_SKILL_DIR}/scripts/musicgen.py stems --in cue.wav --out-dir cue_stems/
# -> cue_stems/{vocals,drums,bass,other}.wav (other will dominate for instrumental)

Example 4: Extend a sketch idea

uv run ${CLAUDE_SKILL_DIR}/scripts/musicgen.py continue \
    --in sketch_8bars.wav \
    --model stable-audio-open \
    --prompt "build into a climax, more energy, bigger drums" \
    --duration 15 \
    --out sketch_extended.wav

Troubleshooting

Error: ModuleNotFoundError: No module named 'diffusers'

Cause: Riffusion backend isn't installed. Solution: pip install diffusers transformers torch accelerate soundfile or scripts/musicgen.py install riffusion.

Error: CUDA out of memory on YuE

Cause: YuE 7B needs ~16 GB VRAM in fp16. Solution: load in int4 / bnb-4bit (--load-in-4bit flag; requires bitsandbytes), shorten duration, or generate on CPU (slow — 30+ min).

Output clips at 5 s for longer prompts

Cause: Riffusion's native window is 5 s; longer requires the crossfade-loop which the script normally handles automatically. Solution: confirm --duration exceeds 5 and that the script version is up to date. Try --model stable-audio-open for native long-form.

YuE produces melody but no vocals

Cause: prompt lacks section markers or vocal cues. Solution: use [Verse 1] / [Chorus] tags and mention vocals explicitly ("male vocalist", "female vocalist", "harmonies").

Generated music sounds like a specific copyrighted song

Cause: overly specific artist / song reference in the prompt. Solution: describe the style instead of the source (e.g. "melancholic 90s grunge with distorted guitars" rather than "a Nirvana song").

Stable Audio Open refuses to run locally

Cause: Stability CC access-gated on HuggingFace. Solution: accept the model license on https://huggingface.co/stabilityai/stable-audio-open-1.0 and huggingface-cli login.

Reference docs

• references/LICENSES.md — strict confirmation of commercial-safety, including Stable Audio Open's $1M ARR cap and explicit exclusions of MusicGen / Suno / Udio / Jukebox.