Yue2 Music

YuE2 Music

Turn a musical request into a reproducible song and an audible comparison. Use released model interfaces. Retain an original song and its plan before making changes.

Choose the workflow

audio → SheetSage2 [loads MERT-v2-FullSong itself] → ABC
style + lyrics → YuE2 full/melody planning        → ABC
                                                  edit/validate
style + lyrics + ABC → YuE2 semantic generation → synthesis → latents → VAE → song
style + lyrics      → YuE2 off generation      → synthesis → latents → VAE → song

Do not feed public MERT feature tensors to YuE2 as codec tokens. YuE2 exposes no audio-reference, phoneme-alignment, or local-inpainting argument.

Set up the needed models

Read models-and-setup.md. Install the YuE2 runtime from the official GitHub repository source. Use a separate environment for SheetSage2 because dependency pins differ. Download the public model snapshots and record their revisions. This skill's original instructions, helpers, and templates are licensed under Apache 2.0. Copyright (c) 2026 the YuE2 authors. Model weights and third-party dependencies retain their applicable licenses.

Use the supported baseline: one request at a time, BF16-capable NVIDIA GPU with 24 GB VRAM, default YuE2 settings. Do not silently shorten a requested song or lower inference settings to hide an OOM. Free allocations or choose suitable hardware; report changes.

Use YuE2-Vae for listening and YuE2-Vae-legacy when reproducing the supplied benchmark protocol. Keep decoded files separate. Do not infer their roles from the word “legacy.”

Run helper paths below relative to this skill folder, with the appropriate environment's Python. Select snapshots using --model, --revision, and --vae-revision, or local model directories plus --offline. Use fresh output directories.

Generate and retain the plan

Start with assets/prompt.json, an original example. Put genre, instruments, vocal character, language and intended tempo in style; put section tags and actual words in lyrics. Keep implementation notes out of lyrics.

python scripts/run_yue2.py generate --request assets/prompt.json --output outputs/pop
python scripts/run_yue2.py all-modes --request assets/prompt.json --output outputs/modes
python scripts/run_yue2.py plan --request assets/prompt.json --output outputs/plan

Inspect result.json, truncation, score.abc, request.json, and audio. Keep exact tokens and latent.npy; the helper saves native artifacts. Preserve all requested modes and failures. A successful process or playable file does not establish musical quality.

Read generation-and-covers.md for Python, CLI, exact plan continuation, CFG, sampling, and cached decoding. Load unchanged plans with SymbolicPlan.load; submit modified ABC as a new input. CLI --resume verifies a completed result; it does not continue interrupted generation.

Cover a recording

1. Transcribe in the SheetSage2 environment. Select the vocal melody or the full lead melody, including instrumental passages.
2. Inspect warnings and correct missed notes, meter or key before attributing errors to YuE2. Preserve source audio and raw transcription.
3. Export chord-free ABC. Select a retained voice explicitly when dropping a part; removing chords alone should preserve both melodic voices and their rests.
4. Render with cot="melody", target style and suitable lyrics. This supplies a symbolic melody condition; it does not preserve the source singer's identity or waveform.

# SheetSage2 environment.
python scripts/transcribe.py reference.wav --task melody-full --output outputs/transcription
python scripts/abc_tools.py strip-chords outputs/transcription/score.abc outputs/cover.abc

# YuE2 environment; the request supplies target style and lyrics.
python scripts/run_yue2.py generate --request assets/prompt.json --cot melody \
  --abc-file outputs/cover.abc --output outputs/cover-song

cot="melody" does not remove chord symbols automatically. To retain the original harmony as well, use full transcription and cot="full"; call this score-conditioned regeneration with melody and harmony.

Edit or delegate an edit

Read editing-workflows.md and abc-editing.md before changing a score.

1. Render a baseline from the full plan. Freeze its original directory.

2. Define invariants: exact pitches; pitch plus rhythm; contour only; or bounded melodic adaptation. Specify voices/passages, lyrics, instruments, tempo, meter and structure.

3. If delegation is available, give a score-editing agent raw ABC, prompt, lyrics, the requested change and the edit brief. Request a new ABC, revised style/lyrics as needed, and an edit manifest. Give a separate reviewer the before/after artifacts and constraints. Without delegation, perform these stages yourself. Keep model generation sequential per GPU.

4. Check musical events, not character strings: ties, accidentals and compressed rests matter. Run:

   python scripts/abc_tools.py inspect edits/jazz.abc
   python scripts/abc_tools.py compare outputs/plan/score.abc edits/jazz.abc --voices Vocal
   python scripts/run_yue2.py generate --request edits/jazz.json --cot full \
     --abc-file edits/jazz.abc --output outputs/jazz
   

   Add --allow-tempo-change for intentional tempo changes. Exact comparison should fail for intentional rhythm changes; audit permitted differences from its report instead of relabeling the result “melody preserved.” Keep the edited score connected through --abc-file or request abc_path; omitting both with abc: null generates a fresh plan and discards the edit.

5. Regenerate after changing style, lyrics or ABC. Old acoustic latents can be decoded again, but cannot implement a musical or lyric edit.

6. Compare full songs and short passages around the edit. Revise when the requested effect fails; retain each attempt and its actual prompt.

For lyric translation, adapt syllables, stress, vowels and breath points. Keep a syllable/phoneme-to-note sidecar. Do not invent a phonemes field or mistake the sidecar for hard acoustic alignment. Use ASR/PER and listening as separate evidence.

Deliver an audible result

Read listening-and-evaluation.md. Return playable audio, full prompt/lyrics, before/after ABC, invariant checks and requested evaluations. Keep model/decoder identity and failures visible.

python scripts/listen.py outputs/pop outputs/jazz --output outputs/comparison

This creates a local HTML player, copies audio, and includes the exact requests. It does not publish or upload. Distinguish symbolic checks, ASR, listening and quality scores. Deliver custom edit manifests and before/after comparison reports alongside the page; the player copies a fixed set of native artifacts, not arbitrary sidecars. Do not claim exact note realization, instrument removal or sample-accurate preservation from an ABC check or SongBench score alone.